Updating stuff

.gitignore

@@ -18,7 +18,7 @@ __pycache__
 
 !assets/*.png
 !.pre-commit-config.yaml
-!data/prompt.txt
+!data/skills/**
 !data/vocab_base.txt
 !requirements.txt
 !util/tl_inspector/TLInspector.js

README.md

@@ -176,8 +176,8 @@ Format: Japanese name, English name in parentheses, then gender.
 
 > **Note:** A very large vocab file can increase API costs and potentially reduce quality. Focus on the most important characters and terms.
 
-### data/prompt.txt
-This is the system prompt sent to the AI. A default `data/prompt.txt` is included and works well for most games. You generally don't need to edit it unless you want to customize the translation style.
+### data/skills/system.md
+This is the system prompt skill sent to the AI on every translation call. A default `data/skills/system.md` is included and works well for most games. You generally don't need to edit it unless you want to customize the translation style. Per-game voice rules go in `<game>/skills/quirks.md` instead.
 
 ---
 
@@ -230,7 +230,7 @@ How it works (all engine modules are supported automatically):
 
 1. **Pass 1 (collect)** — files are processed normally, but instead of calling the API each
    request is queued to `log/batch_requests.json`. Requests are byte-identical to live ones:
-   the static `prompt.txt` block is cached with a 1h TTL, matched vocab and translation
+   the static `data/skills/system.md` block is cached with a 1h TTL, matched vocab and translation
    history ride along per request, and structured output enforces the exact line count.
    Speaker/variable names still translate live during this pass (they get embedded into the
    dialogue payloads, so both passes must resolve them identically) — they're a tiny share

TODO.md

@@ -1,19 +1,13 @@
 # TODO
 
-## Custom game prompt
-
-- Keep the default prompt (`prompt.txt`) as the base for every game.
-- Add an optional **per-game custom prompt** in the UI that merges with the default (custom rules layered on top of shared instructions, not a full replacement unless explicitly chosen).
-- Surface this in settings / project setup so each game folder can carry its own tone, terminology notes, or content warnings without forking the global prompt.
-
 ## Local translation cache (game retranslate)
 
 - Persist every successful translation locally so a full retranslate from scratch can reuse prior results instead of paying again.
-- **Cache key:** hash the translation **payload** (source Japanese text sent to the model) plus target language — same approach as the existing global cache in `util/translation.py` (`get_cache_key`, `log/translation_cache.json`).
+- **Cache key:** hash the translation **payload** (source Japanese text sent to the model) plus target language - same approach as the existing global cache in `util.translation` (`get_cache_key`, `log/translation_cache.json`).
   Payload keys naturally dedupe identical lines across files and survive re-extract / re-import as long as the source string is unchanged.
 - **Scope:** consider a per-game cache file under the game project (e.g. `<game_root>/translation_cache.json` or alongside `wolf_json/`) in addition to (or merged with) the tool-wide `log/translation_cache.json`, so caches travel with the game folder and are easy to back up.
 - **Write path:** on every successful translate (live and batch consume), store `{cache_key: translation}` before or as results land in `translated/`.
 - **Read path:** before calling the API, check cache; on hit, skip the request and apply the cached translation (respecting file-specific re-expansion where the payload is Japanese-only).
 - **Retranslate workflow:** when the user re-imports fresh JSON and runs Translate again, cache hits should fill `translated/` automatically for unchanged source lines; only new or edited source text should incur API cost.
-- **UI (later):** cache stats (hits / misses / estimated savings), export/import, clear cache for one game, optional “prefer cache” toggle.
+- **UI (later):** cache stats (hits / misses / estimated savings), export/import, clear cache for one game, optional "prefer cache" toggle.
 - **Invalidation:** optional metadata (model id, prompt version) in the key or entry if we need to avoid reusing stale translations after prompt or model changes.

data/skills/project_setup.md

@@ -0,0 +1,192 @@
+# DazedMTLTool — Project Setup
+
+You are analysing a Japanese game project to produce configuration artifacts for DazedMTLTool.
+Work in the game repository. Scan files; do not invent content you did not see.
+
+---
+
+## Task
+
+In **one pass**, discover everything needed for translation setup and emit the labeled output blocks below.
+Do **not** translate the game. Do **not** rewrite formatting codes, wrap widths, or tool pipeline settings.
+
+Default: emit **all four** blocks.
+Regenerate mode: if the user asks for only one block (`glossary`, `speakers`, `translation_quirks`, or `game_skill`), emit **only** that block using the same schema and exclusivity rules.
+
+---
+
+## Ownership (no duplication)
+
+| Block | Owns | Must NOT include |
+|-------|------|------------------|
+| `glossary` | Named characters (gender, role, **per-character** speech register) + worldbuilding terms | Global dialect / person rules; honorific policy; speaker-format flags; formatting |
+| `speakers` | Tool flag ENABLE/SKIP + short evidence | Character bios; quirks; full glossary |
+| `translation_quirks` | Cross-cutting voice rules (battle-log person, global dialect, item-description style, **unusual** honorific habits) | Per-character register; "always keep -san" (tool base prompt already does); codes/wrap/line counts; speaker flags |
+| `game_skill` | Game identity, which files matter, pointer to `skills/quirks.md`, what the tool owns | Verbatim quirks list; verbatim glossary; restating formatting/honorific base policy |
+
+Hard rules:
+1. Per-character voice → `glossary` only.
+2. Category-wide / cross-cutting voice → `translation_quirks` only.
+3. Default honorifics policy is owned by the tool base prompt - only note **unusual** honorific habits in quirks.
+4. `game_skill` **references** `skills/quirks.md` only; do not reprint those rules or cite legacy names like `translation_quirks.txt`.
+5. `speakers` is config, not lore.
+
+---
+
+## Shared scan strategy
+
+Map / event files can be huge. Do **not** read them sequentially end-to-end.
+
+1. Read small DB files in full first (richest, always small).
+2. For large event/map files: **search/grep**; sample early story maps; stop when patterns stabilize.
+3. One scan feeds all four blocks - do not rescan from scratch per block.
+
+<!-- engine:rpgmaker -->
+
+### RPG Maker file strategy
+
+**DB (read in full):**
+- `Actors.json` (mandatory for major characters, actor IDs, `\\N[n]` mappings)
+- `Classes.json`, `Troops.json`, `Skills.json`, `Items.json`, `Armors.json`, `Weapons.json`, `States.json`, `System.json`
+
+**Events (grep / sample):**
+- `CommonEvents.json`, `Troops.json`, `Map001.json`–`Map010.json` (early maps first)
+- Prefer code `401` dialogue + nearby `101` speaker/name params; `405` scrolling text when present
+- Speaker markup evidence: `【Name】`, `[Name]`, `Name：`, `\\n<Name>`, `\\k<Name>`, colour-wrapped name lines
+
+--- attach your game data files / open the game repo before continuing ---
+
+### Speakers analysis (for `speakers` block)
+
+Code `101` opens the text window. Code `401` is a dialogue line. Multiple `401`s form one message box.
+
+**Always-on formats (NO FLAG needed):**
+- `101` param[4] name
+- `\\n<Name>` or `\\k<Name>` (angle brackets)
+- `【Name】` alone or with dialogue
+- `[Name]` alone or with dialogue
+- `\\c[N]Name\\c[0]` colour-wrapped name on its own `401` line
+- `Name：` full-width colon name line
+
+**Critical:** `\\N[X]` / `\\n[X]` (square brackets + number) are actor variable codes, NOT speaker formats. Do not count them as always-on speaker hits.
+
+**Flags (only when some speakers lack an always-on format):**
+- `INLINE401SPEAKERS` — name immediately before `「` on a `401` line (e.g. `エレナ「今日は…`)
+- `FIRSTLINESPEAKERS` — first `401` is a short plain name (< 40 chars, has Japanese); next line starts with `「` `"`` `（` `*` `[`; often empty face on `101`
+- `FACENAME101` — last resort only when no always-on and neither flag above; face filename in `101` param[0], empty param[4]. If ENABLE, list every unique face filename from `101` param[0].
+
+### Glossary rules (for `glossary` block)
+
+- Separator: plain hyphen-minus `-` only (never em/en dash).
+- Descriptions entirely in English; refer to other characters by English name.
+- Commit to one spelling - never `Sylfia / Sylphia`.
+- Characters: gender, role, speech register, personality, player-chosen name (Actors.json ID 1).
+- Real named actors get full `# Game Characters` entries, not only `\\N[n]` placeholders.
+- Worldbuilding: factions, lore locations, unique systems/titles - exclude skill/item/weapon/armour names and generic RPG words.
+
+### Quirks rules (for `translation_quirks` block)
+
+Find translation-only quirks, for example:
+- Battle log / system messages consistently 3rd person (or other fixed person)
+- Global dialect (old-timer speech, archaic narration)
+- Recurring item/skill description style
+- Unusual honorific habits (who uses what with whom)
+
+Exclude: formatting codes, wrap, line counts, speaker flags, character name lists.
+
+Output as short imperative bullets suitable to paste into `skills/quirks.md`.
+
+### Game skill rules (for `game_skill` block)
+
+Produce a durable IDE skill for this title saved at `skills/translation.md`:
+- Short game identity (tone/genre/content notes)
+- Which files matter for dialogue/DB in this repo
+- A **Voice rules** section that points only at ``skills/quirks.md`` (authoritative)
+- What not to change (codes, formatting - tool-owned via base prompt)
+- Do **not** reprint quirks or the glossary
+
+**Path names (required):**
+- Voice quirks file: ``skills/quirks.md`` (never ``translation_quirks.txt`` or any other legacy name)
+- This IDE skill file: ``skills/translation.md``
+- Optional user overlays (if mentioned): other ``skills/*.md`` files except ``quirks.md`` / ``translation.md``
+
+<!-- /engine:rpgmaker -->
+
+<!-- engine:wolf -->
+
+### Wolf file strategy (stub)
+
+When analysing a WolfDawn / WOLF project later: use `wolf_json/` extracts (`names.json`, DB sheets, map/common scenes). Prefer foundation DB + narrative sheets; grep large map dumps. Same four-block ownership rules apply.
+
+<!-- /engine:wolf -->
+
+---
+
+## Output format
+
+Emit **only** labeled fenced code blocks (plus a one-line heading before each block for navigation).
+No essay outside the blocks.
+
+### Block `glossary`
+
+Label the fence language as `glossary`. Inside:
+
+```
+# Game Characters
+Japanese (English) - description
+
+# Worldbuilding Terms
+Japanese (English) - description
+```
+
+### Block `speakers`
+
+Label the fence language as `speakers`. Inside:
+
+```
+Patterns detected:
+- ...
+
+Flag decisions:
+INLINE401SPEAKERS : ENABLE|SKIP — <one-line reason>
+FIRSTLINESPEAKERS : ENABLE|SKIP — <one-line reason>
+FACENAME101       : ENABLE|SKIP — <one-line reason>
+
+Examples:
+- ...
+
+Face filenames (only if FACENAME101 ENABLE):
+- ...
+```
+
+### Block `translation_quirks`
+
+Label the fence language as `translation_quirks`. Inside: imperative bullet list only (no preamble).
+
+### Block `game_skill`
+
+Label the fence language as `game_skill`. Inside: full markdown for `skills/translation.md`.
+
+Required skeleton (fill in the blanks; keep the Voice rules path exactly):
+
+```markdown
+# <Game title> - Translation skill
+
+## Identity
+<1-3 sentences: tone, genre, content notes>
+
+## Files that matter
+- <dialogue / events paths for this repo>
+- <database paths>
+
+## Voice rules
+Follow `skills/quirks.md` as the authoritative cross-cutting voice guide for this title.
+Do not invent alternate quirk filenames.
+
+## Tool boundaries
+- DazedMTLTool owns formatting codes, wrap, line counts, and honorifics defaults via its base system skill.
+- Per-character register lives in the glossary (`vocab.txt`), not here.
+- Do not reprint quirks or glossary entries in this file.
+```
+
+If known speakers were prepended in a `<known_speakers>` block above this skill, prefer those names in the glossary, then cross-check Actors.json for other major named actors.

data/skills/system.md

@@ -30,6 +30,7 @@ You will be translating erotic and sexual content. You will receive lines of dia
 - **Speech register:** If the entry describes how a character speaks (flustered, blunt, formal, childlike, crude, etc.), mirror that register in their English dialogue. A character described as speaking in a "cute, flustered register" should sound different from one described as "cold and terse".
 - **Role & context:** Use the role/personality notes to inform tone — a villain's lines should feel threatening, a comic-relief NPC's lines should feel goofy, etc.
 - Japanese omits pronouns constantly. Infer the correct subject and pronoun from context, translation history, and the character list.
+- **Person / perspective consistency (1st / 2nd / 3rd):** Keep the grammatical person consistent with the source and with surrounding lines. Before writing English, verify what perspective the Japanese line actually uses - do not assume first-person dialogue when the line is third-person narration, second-person address (お前 / 君 / あなた), or impersonal UI text. Match "I / you / he / she / they" (and possessives) to that verified perspective; do not flip person mid-scene unless the Japanese does.
 - **Preserve third-person self-reference.** Some characters refer to themselves by name instead of using "I" (e.g. ワタシ used as a name, or a character saying their own name). When a character is clearly speaking about themselves in the third person as a stylistic trait, maintain that in English (e.g. "Feris doesn't know" rather than "I don't know").
 - Third-person pronouns (彼, 彼女, あいつ, こいつ, そいつ, コイツ) should match the known gender of the person being referenced.
 - Translate **コイツ** as "this bastard" (male) or "this bitch" (female) depending on the referenced character's gender.
@@ -38,7 +39,9 @@ You will be translating erotic and sexual content. You will receive lines of dia
 
 ## Honorifics and Names
 
-- Preserve Japanese honorifics in the translation: -san, -kun, -chan, -senpai, -sensei, -sama, -dono, etc.
+- **Always preserve and enforce Japanese honorifics** in the English translation: -san, -kun, -chan, -senpai, -sensei, -sama, -dono, etc.
+- Do **not** drop honorifics to make the English sound more "natural" unless the source Japanese also lacks them.
+- If `# Game Characters` (or `<game>/skills/quirks.md`) specifies how a character addresses someone, follow that address pattern.
 - The `=` or `＝` character in a Japanese name marks a foreign or nickname component. Wrap that part in parentheses.
   - Example: `バンカー＝ベット` → `Bunker (Bet)`
 - Always translate speaker tags to English: `[クロネ]:` → `[Kurone]:`

gui/main.py

@@ -595,6 +595,7 @@ from gui.config_tab import ConfigTab
 from gui.translation_tab import TranslationTab
 from gui.workflow_tab import WorkflowTab
 from gui.wolf_workflow_tab import WolfWorkflowTab
+from gui.skills_tab import SkillsTab
 from gui.batch_tab import BatchTab
 
 class DazedMTLGUI(QMainWindow):
@@ -838,9 +839,16 @@ class DazedMTLGUI(QMainWindow):
         sidebar_layout.addWidget(btn_batches)
         self.nav_buttons.append(btn_batches)
 
-        # Configuration button (fourth)
+        # Skills button (fourth)
+        btn_skills = self.create_nav_button("📚", "Skills")
+        btn_skills.setToolTip("Skills — edit system prompt, Project Setup, and translation contexts")
+        btn_skills.clicked.connect(lambda: self.switch_page(3))
+        sidebar_layout.addWidget(btn_skills)
+        self.nav_buttons.append(btn_skills)
+
+        # Configuration button (fifth)
         btn_config = self.create_nav_button("⚙️", "Configuration")
-        btn_config.clicked.connect(lambda: self.switch_page(3))
+        btn_config.clicked.connect(lambda: self.switch_page(4))
         sidebar_layout.addWidget(btn_config)
         self.nav_buttons.append(btn_config)
         
@@ -906,7 +914,11 @@ class DazedMTLGUI(QMainWindow):
         self.batch_tab = BatchTab(self)
         self.content_stack.addWidget(self.batch_tab)
 
-        # Configuration Tab (index 3)
+        # Skills Tab (index 3)
+        self.skills_tab = SkillsTab(self)
+        self.content_stack.addWidget(self.skills_tab)
+
+        # Configuration Tab (index 4)
         self.config_tab = ConfigTab()
         self.config_tab.config_changed.connect(self.on_config_changed)
         self.content_stack.addWidget(self.config_tab)

gui/skills_tab.py

@@ -0,0 +1,247 @@
+"""Skills Tab - edit tool-level skills and translation contexts under data/."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+from PyQt5.QtCore import Qt
+from PyQt5.QtGui import QFont
+from PyQt5.QtWidgets import (
+    QHBoxLayout,
+    QLabel,
+    QMessageBox,
+    QPushButton,
+    QTabWidget,
+    QTextEdit,
+    QVBoxLayout,
+    QWidget,
+)
+
+from util.paths import PROMPT_PATH, SKILLS_DIR, TRANSLATION_CONTEXTS_PATH
+from util.skills.contexts import reload_contexts
+
+
+class _PlainPasteTextEdit(QTextEdit):
+    """QTextEdit that always pastes as plain text."""
+
+    def insertFromMimeData(self, source):  # noqa: N802
+        self.insertPlainText(source.text())
+
+
+_EDITOR_STYLE = (
+    "QTextEdit{background-color:#1e1e1e;color:#d4d4d4;"
+    "border:1px solid #3c3c3c;border-radius:4px;padding:10px;"
+    "selection-background-color:#264f78;}"
+)
+
+_HINT_STYLE = "color:#9d9d9d;font-size:13px;"
+_PATH_STYLE = "color:#569cd6;font-size:12px;font-family:Consolas,monospace;"
+
+
+class SkillsTab(QWidget):
+    """Edit ``data/skills/*.md`` and ``data/translation_contexts.json``."""
+
+    def __init__(self, parent=None):
+        super().__init__(parent)
+        self._pages: dict[str, dict] = {}
+        self._build_ui()
+        self.reload_all()
+
+    def _build_ui(self):
+        root = QVBoxLayout(self)
+        root.setContentsMargins(20, 16, 20, 16)
+        root.setSpacing(10)
+
+        title = QLabel("Skills")
+        title.setStyleSheet("color:#ffffff;font-size:18px;font-weight:bold;")
+        root.addWidget(title)
+
+        intro = QLabel(
+            "Tool-level AI instructions live here. "
+            "Per-game quirks, the game IDE skill, and optional custom skills are edited "
+            "in Workflow Step 2 (they travel with the game folder)."
+        )
+        intro.setWordWrap(True)
+        intro.setStyleSheet(_HINT_STYLE)
+        root.addWidget(intro)
+
+        self.tabs = QTabWidget()
+        self.tabs.setStyleSheet(
+            "QTabWidget::pane{border:1px solid #3c3c3c;border-radius:4px;top:-1px;}"
+            "QTabBar::tab{background:#2d2d30;color:#cccccc;padding:8px 14px;"
+            "border:1px solid #3c3c3c;border-bottom:none;margin-right:2px;}"
+            "QTabBar::tab:selected{background:#1e1e1e;color:#ffffff;"
+            "border-top:2px solid #007acc;}"
+        )
+        root.addWidget(self.tabs, 1)
+
+        self._add_file_page(
+            key="system",
+            tab_title="System",
+            path=PROMPT_PATH,
+            hint=(
+                "Runtime system prompt for every translation API call. "
+                "Honorifics, formatting, and universal quality rules live here. "
+                "Do not put game-specific voice quirks here - use <game>/skills/quirks.md."
+            ),
+            is_json=False,
+        )
+        self._add_file_page(
+            key="project_setup",
+            tab_title="Project Setup",
+            path=SKILLS_DIR / "project_setup.md",
+            hint=(
+                "Clipboard skill copied from Workflow. "
+                "Instructs the IDE to emit glossary, speakers, quirks, and game_skill blocks."
+            ),
+            is_json=False,
+        )
+        self._add_file_page(
+            key="contexts",
+            tab_title="Contexts",
+            path=TRANSLATION_CONTEXTS_PATH,
+            hint=(
+                "Per-call history templates used by RPG Maker / WolfDawn "
+                "(names.npc, database.item, events.label_108, …). "
+                "Edit carefully - keys must match what the engine modules look up."
+            ),
+            is_json=True,
+        )
+
+        btn_row = QHBoxLayout()
+        reload_btn = QPushButton("↺  Reload All")
+        reload_btn.setFixedWidth(140)
+        reload_btn.clicked.connect(self.reload_all)
+        btn_row.addWidget(reload_btn)
+        btn_row.addStretch()
+        open_btn = QPushButton("Open skills folder")
+        open_btn.setFixedWidth(160)
+        open_btn.clicked.connect(self._open_skills_folder)
+        btn_row.addWidget(open_btn)
+        root.addLayout(btn_row)
+
+    def _add_file_page(
+        self,
+        *,
+        key: str,
+        tab_title: str,
+        path: Path,
+        hint: str,
+        is_json: bool,
+    ):
+        page = QWidget()
+        layout = QVBoxLayout(page)
+        layout.setContentsMargins(12, 12, 12, 12)
+        layout.setSpacing(8)
+
+        path_lbl = QLabel(str(path))
+        path_lbl.setStyleSheet(_PATH_STYLE)
+        path_lbl.setTextInteractionFlags(Qt.TextSelectableByMouse)
+        layout.addWidget(path_lbl)
+
+        hint_lbl = QLabel(hint)
+        hint_lbl.setWordWrap(True)
+        hint_lbl.setStyleSheet(_HINT_STYLE)
+        layout.addWidget(hint_lbl)
+
+        editor = _PlainPasteTextEdit()
+        editor.setFont(QFont("Consolas", 10))
+        editor.setStyleSheet(_EDITOR_STYLE)
+        layout.addWidget(editor, 1)
+
+        row = QHBoxLayout()
+        save_btn = QPushButton("💾  Save")
+        save_btn.setFixedWidth(120)
+        save_btn.clicked.connect(lambda _=False, k=key: self._save_page(k))
+        row.addWidget(save_btn)
+
+        reload_btn = QPushButton("↺  Reload")
+        reload_btn.setFixedWidth(120)
+        reload_btn.clicked.connect(lambda _=False, k=key: self._reload_page(k))
+        row.addWidget(reload_btn)
+
+        status = QLabel("")
+        status.setStyleSheet("color:#6a9955;font-size:12px;")
+        row.addWidget(status)
+        row.addStretch()
+        layout.addLayout(row)
+
+        self.tabs.addTab(page, tab_title)
+        self._pages[key] = {
+            "path": path,
+            "editor": editor,
+            "status": status,
+            "is_json": is_json,
+        }
+
+    def reload_all(self):
+        for key in self._pages:
+            self._reload_page(key)
+
+    def _reload_page(self, key: str):
+        page = self._pages[key]
+        path: Path = page["path"]
+        editor: QTextEdit = page["editor"]
+        status: QLabel = page["status"]
+        try:
+            if not path.is_file():
+                editor.setPlainText("")
+                status.setText(f"Missing: {path.name}")
+                status.setStyleSheet("color:#f0ad4e;font-size:12px;")
+                return
+            editor.setPlainText(path.read_text(encoding="utf-8"))
+            status.setText("Loaded")
+            status.setStyleSheet("color:#6a9955;font-size:12px;")
+        except Exception as exc:
+            status.setText(f"Load failed: {exc}")
+            status.setStyleSheet("color:#f44747;font-size:12px;")
+
+    def _save_page(self, key: str):
+        page = self._pages[key]
+        path: Path = page["path"]
+        editor: QTextEdit = page["editor"]
+        status: QLabel = page["status"]
+        text = editor.toPlainText()
+        if page["is_json"]:
+            try:
+                parsed = json.loads(text)
+                text = json.dumps(parsed, ensure_ascii=False, indent=2) + "\n"
+                editor.setPlainText(text.rstrip("\n") + "\n")
+            except json.JSONDecodeError as exc:
+                QMessageBox.warning(
+                    self,
+                    "Invalid JSON",
+                    f"Cannot save {path.name}:\n{exc}",
+                )
+                status.setText("Invalid JSON")
+                status.setStyleSheet("color:#f44747;font-size:12px;")
+                return
+        try:
+            path.parent.mkdir(parents=True, exist_ok=True)
+            path.write_text(text if text.endswith("\n") else text + "\n", encoding="utf-8")
+            if key == "contexts":
+                reload_contexts()
+            status.setText(f"Saved {path.name}")
+            status.setStyleSheet("color:#6a9955;font-size:12px;")
+        except Exception as exc:
+            QMessageBox.warning(self, "Save failed", str(exc))
+            status.setText(f"Save failed: {exc}")
+            status.setStyleSheet("color:#f44747;font-size:12px;")
+
+    def _open_skills_folder(self):
+        folder = SKILLS_DIR
+        folder.mkdir(parents=True, exist_ok=True)
+        import os
+        import subprocess
+        import sys
+
+        try:
+            if sys.platform.startswith("win"):
+                os.startfile(str(folder))  # type: ignore[attr-defined]
+            elif sys.platform == "darwin":
+                subprocess.run(["open", str(folder)], check=False)
+            else:
+                subprocess.run(["xdg-open", str(folder)], check=False)
+        except Exception as exc:
+            QMessageBox.information(self, "Skills folder", f"{folder}\n\n{exc}")

gui/translation_tab.py

@@ -3018,6 +3018,20 @@ class TranslationTab(QWidget):
                 except Exception:
                     pass
 
+                # Per-game quirks overlay: Workflow Step 0 folder → DAZED_GAME_ROOT
+                try:
+                    game_root = ""
+                    if self.settings:
+                        game_root = str(
+                            self.settings.value("last_game_folder", "") or ""
+                        ).strip()
+                    if game_root and Path(game_root).is_dir():
+                        os.environ["DAZED_GAME_ROOT"] = game_root
+                    else:
+                        os.environ.pop("DAZED_GAME_ROOT", None)
+                except Exception:
+                    pass
+
                 # Try to create a hard link at legacy location so modules that
                 # still write to log/translationHistory.txt end up in this file.
                 # This will be created when the run_log_path file is first written to

modules/rpgmakermvmz.py

@@ -15,14 +15,15 @@ from retry import retry
 from tqdm import tqdm
 from util.translation import TranslationConfig, translateAI as sharedtranslateAI, getPricingConfig, calculateCost, getPricingConfig, calculateCost, get_var_translation, set_var_translations_batch, convert_corner_brackets
 from util.speakers import SPEAKER_BRACKET_INNER, strip_speaker_prefix
+from util.skills import ctx, load_system_prompt
+from util.paths import VOCAB_PATH
 
 # Globals
 MODEL = os.getenv("model")
 TIMEOUT = int(os.getenv("timeout"))
 LANGUAGE = os.getenv("language").capitalize()
-from util.paths import PROMPT_PATH, VOCAB_PATH
 
-PROMPT = PROMPT_PATH.read_text(encoding="utf-8")
+PROMPT = load_system_prompt()
 VOCAB = VOCAB_PATH.read_text(encoding="utf-8")
 LOCK = threading.Lock()
 THREAD_CTX = threading.local()
@@ -240,7 +241,7 @@ def _pat355655_captured_text(match):
 
 
 def handleMVMZ(filename, estimate):
-    global ESTIMATE, TOKENS, FILENAME, MISMATCH
+    global ESTIMATE, TOKENS, FILENAME, MISMATCH, PROMPT
     ESTIMATE = estimate
     FILENAME = filename
     MISMATCH = []  # Reset per-file; prevents cross-file contamination in CLI mode
@@ -250,6 +251,10 @@ def handleMVMZ(filename, estimate):
     except Exception:
         pass
 
+    # Reload base prompt + optional per-game quirks (DAZED_GAME_ROOT).
+    PROMPT = load_system_prompt()
+    TRANSLATION_CONFIG.prompt = PROMPT
+
     # Translate
     start = time.time()
     translatedData = openFiles(filename)
@@ -979,7 +984,7 @@ def parseMap(data, filename):
     if "Map" in filename:
         response = translateAI(
             data["displayName"],
-            "Reply with only the " + LANGUAGE + " translation of the RPG location name",
+            ctx("names.location"),
             False,
         )
         totalTokens[0] += response[1][0]
@@ -1097,7 +1102,7 @@ def _normalize_sg_desc(text: str) -> str:
             # Translate
             response = translateAI(
                 modifiedJAString,
-                "Reply with only the " + LANGUAGE + " translation.",
+                ctx("database.generic"),
                 False,
             )
             translatedText = response[0]
@@ -1132,7 +1137,7 @@ def translateNoteOmitSpace(event, regex):
         # Translate
         response = translateAI(
             jaString,
-            "Reply with the " + LANGUAGE + " translation of the location name.",
+            ctx("names.location_short"),
             False,
         )
         # Defend against unexpected response shapes
@@ -1181,7 +1186,7 @@ def translateLBNames(events):
         names_to_translate = [item[1] for item in lb_events]
         response = translateAI(
             names_to_translate,
-            "Reply with only the " + LANGUAGE + " translation of the name.",
+            ctx("names.generic"),
             True,
         )
         translated_names = response[0] if isinstance(response[0], list) else [response[0]]
@@ -1670,21 +1675,21 @@ def searchNames(data, pbar, context, filename):
 
     # Set the context of what we are translating
     if "Actors" in context:
-        newContext = "Reply with only the " + LANGUAGE + " translation of the NPC name"
+        newContext = ctx("names.npc_only")
     if "Armors" in context:
-        newContext = "Reply with only the " + LANGUAGE + " translation of the RPG equipment name"
+        newContext = ctx("database.armor")
     if "Classes" in context:
-        newContext = "Reply with only the " + LANGUAGE + " translation of the RPG class name"
+        newContext = ctx("database.class")
     if "MapInfos" in context:
-        newContext = "Reply with only the " + LANGUAGE + " translation of the location name"
+        newContext = ctx("names.location")
     if "Enemies" in context:
-        newContext = "Reply with only the " + LANGUAGE + " translation of the enemy NPC name"
+        newContext = ctx("names.enemy")
     if "Weapons" in context:
-        newContext = "Reply with only the " + LANGUAGE + " translation of the RPG weapon name"
+        newContext = ctx("database.weapon")
     if "Items" in context:
-        newContext = "Reply with only the " + LANGUAGE + " translation of the RPG item name"
+        newContext = ctx("database.item")
     if "Skills" in context:
-        newContext = "Reply with only the " + LANGUAGE + " translation of the RPG skill name"
+        newContext = ctx("database.skill")
 
     # Names
     with open("log/translations.txt", "a", encoding="utf-8") as file:
@@ -1753,7 +1758,7 @@ def searchNames(data, pbar, context, filename):
     # --- Batch translate all notes ---
     translatedNotesBatch = []
     if notesBatch:
-        response = translateAI(notesBatch, f"Reply with only the {LANGUAGE} translation of the note text.")
+        response = translateAI(notesBatch, ctx("database.note"))
         translatedNotesBatch = response[0]
         totalTokens[0] += response[1][0]
         totalTokens[1] += response[1][1]
@@ -2002,7 +2007,7 @@ def searchNames(data, pbar, context, filename):
                 if descriptionList:
                     response = translateAI(
                         descriptionList,
-                        f"Reply with only the {LANGUAGE} translation of the text.",
+                        ctx("events.generic_text"),
                         True,
                     )
                     translatedDescriptionBatch = response[0]
@@ -2954,7 +2959,7 @@ def searchCodes(page, pbar, jobList, filename):
                         question = codeList[i]["parameters"][3]["messageText"]
                         response = translateAI(
                             matchList,
-                            f"Previous text for context: {question}\n",
+                            ctx("events.plugin_with_context", context=question),
                             True,
                         )
                         totalTokens[0] += response[1][0]
@@ -3716,7 +3721,7 @@ def searchCodes(page, pbar, jobList, filename):
                         # Translate
                         response = translateAI(
                             matchList[0],
-                            "Reply with the " + LANGUAGE + " translation of the NPC name.",
+                            ctx("names.npc"),
                             False,
                         )
                         translatedText = response[0]
@@ -3812,7 +3817,7 @@ def searchCodes(page, pbar, jobList, filename):
                     if len(matchList) > 0:
                         # Translate
                         text = matchList[0]
-                        response = translateAI(text, "Reply with the " + LANGUAGE + " Translation")
+                        response = translateAI(text, ctx("events.generic"))
                         translatedText = response[0]
                         totalTokens[0] += response[1][0]
                         totalTokens[1] += response[1][1]
@@ -3851,7 +3856,7 @@ def searchCodes(page, pbar, jobList, filename):
                     if len(matchList) > 0:
                         # Translate
                         text = matchList[0]
-                        response = translateAI(text, "Reply with the " + LANGUAGE + " Translation")
+                        response = translateAI(text, ctx("events.generic"))
                         translatedText = response[0]
                         totalTokens[0] += response[1][0]
                         totalTokens[1] += response[1][1]
@@ -3894,7 +3899,7 @@ def searchCodes(page, pbar, jobList, filename):
                         question = translatedText
                         response = translateAI(
                             choiceList,
-                            f"Previous text for context: {question}\n",
+                            ctx("events.plugin_with_context", context=question),
                             True,
                         )
                         totalTokens[0] += response[1][0]
@@ -3949,11 +3954,11 @@ def searchCodes(page, pbar, jobList, filename):
                     if len(textHistory) > 0:
                         response = translateAI(
                             choiceList,
-                            f"Reply with the English translation of the dialogue choice.\n\nPrevious text for context: {str(textHistory)}\n",
+                            ctx("events.choice_with_context", context=str(textHistory)),
                             True,
                         )
                     else:
-                        response = translateAI(choiceList, "Reply with the English translation of the dialogue choice.")
+                        response = translateAI(choiceList, ctx("events.choice"))
                     
                     translatedTextList = response[0]
                     totalTokens[0] += response[1][0]
@@ -4143,7 +4148,7 @@ def searchCodes(page, pbar, jobList, filename):
 
         # 122
         if len(list122) > 0:
-            response = translateAI(list122, "Keep your translation as brief as possible")
+            response = translateAI(list122, ctx("events.code122_brief"))
             list122TL = response[0]
             totalTokens[0] += response[1][0]
             totalTokens[1] += response[1][1]
@@ -4168,7 +4173,7 @@ def searchCodes(page, pbar, jobList, filename):
 
         # 108
         if len(list108) > 0:
-            response = translateAI(list108, "This text is a label. Use title capitalization and keep it brief.")
+            response = translateAI(list108, ctx("events.label_108"))
             list108TL = response[0]
             totalTokens[0] += response[1][0]
             totalTokens[1] += response[1][1]
@@ -4224,7 +4229,7 @@ def searchCodes(page, pbar, jobList, filename):
         # 324
         if len(list324) > 0:
             # Generic short-text translation for parameter index 1
-            response = translateAI(list324, "Reply with only the " + LANGUAGE + " translation of the text.")
+            response = translateAI(list324, ctx("events.generic_text"))
             list324TL = response[0]
             totalTokens[0] += response[1][0]
             totalTokens[1] += response[1][1]
@@ -4236,7 +4241,7 @@ def searchCodes(page, pbar, jobList, filename):
         # 325
         if len(list325) > 0:
             # Use same short-text speaker-style translation as other name fields
-            response = translateAI(list325, "Reply with the " + LANGUAGE + " translation of the NPC name.")
+            response = translateAI(list325, ctx("names.npc"))
             list325TL = response[0]
             totalTokens[0] += response[1][0]
             totalTokens[1] += response[1][1]
@@ -4388,7 +4393,7 @@ def searchSS(state, pbar):
     # --- Batch translate all notes ---
     translatedNotesBatch = []
     if notesBatch:
-        response = translateAI(notesBatch, f"Reply with only the {LANGUAGE} translation of the note text.")
+        response = translateAI(notesBatch, ctx("database.note"))
         translatedNotesBatch = response[0]
         totalTokens[0] += response[1][0]
         totalTokens[1] += response[1][1]
@@ -4444,7 +4449,7 @@ def searchSS(state, pbar):
 
 def searchSystem(data, pbar):
     totalTokens = [0, 0]
-    context = "Reply with only the " + LANGUAGE + ' translation of the UI textbox."'
+    context = ctx("database.ui_term")
 
     # Title - batch as a single-item list
     title_src = _system_scalar_source(data, "gameTitle")
@@ -4452,7 +4457,7 @@ def searchSystem(data, pbar):
     if not (IGNORETLTEXT and not re.search(LANGREGEX, title_src)):
         response = translateAI(
             [title_src],
-            " Reply with the " + LANGUAGE + " translation of the game title name",
+            ctx("database.game_title"),
             False,
         )
         totalTokens[0] += response[1][0]
@@ -4507,7 +4512,7 @@ def searchSystem(data, pbar):
     if armor_values:
         response = translateAI(
             armor_values,
-            "Reply with only the " + LANGUAGE + " translation of the armor type",
+            ctx("database.type_armor"),
             False,
         )
         totalTokens[0] += response[1][0]
@@ -4534,7 +4539,7 @@ def searchSystem(data, pbar):
     if skill_values:
         response = translateAI(
             skill_values,
-            "Reply with only the " + LANGUAGE + " translation",
+            ctx("database.generic"),
             False,
         )
         totalTokens[0] += response[1][0]
@@ -4561,7 +4566,7 @@ def searchSystem(data, pbar):
     if equip_values:
         response = translateAI(
             equip_values,
-            "Reply with only the " + LANGUAGE + " translation of the equipment type. No disclaimers.",
+            ctx("database.type_equipment"),
             False,
         )
         totalTokens[0] += response[1][0]
@@ -4591,7 +4596,7 @@ def searchSystem(data, pbar):
     if element_values:
         response = translateAI(
             element_values,
-            "Reply with only the " + LANGUAGE + " translation of the element type",
+            ctx("database.element"),
             False,
         )
         totalTokens[0] += response[1][0]
@@ -4621,7 +4626,7 @@ def searchSystem(data, pbar):
     if weapon_values:
         response = translateAI(
             weapon_values,
-            "Reply with only the " + LANGUAGE + " translation of the weapon type",
+            ctx("database.type_weapon"),
             False,
         )
         totalTokens[0] += response[1][0]
@@ -4650,7 +4655,7 @@ def searchSystem(data, pbar):
         if var_values:
             response = translateAI(
                 var_values,
-                'Reply with only the ' + LANGUAGE + ' translation of the title',
+                ctx("database.actor_title"),
                 True,
             )
             totalTokens[0] += response[1][0]
@@ -4680,7 +4685,7 @@ def searchSystem(data, pbar):
         if switch_values:
             response = translateAI(
                 switch_values,
-                'Reply with only the ' + LANGUAGE + ' translation of the switch name',
+                ctx("database.switch"),
                 True,
             )
             totalTokens[0] += response[1][0]
@@ -4712,9 +4717,7 @@ def searchSystem(data, pbar):
         if msg_values:
             response = translateAI(
                 msg_values,
-                "Reply with only the "
-                + LANGUAGE
-                + ' translation of the battle text.\nTranslate "常時ダッシュ" as "Always Dash"\nTranslate "次の%1まで" as Next %1.',
+                ctx("database.battle_messages"),
                 False,
             )
             totalTokens[0] += response[1][0]
@@ -4805,7 +4808,7 @@ def getSpeaker(speaker: str):
         pass
     response = translateAI(
         speaker,
-        "Reply with the " + LANGUAGE + " translation of the NPC name.",
+        ctx("names.npc"),
         False,
     )
     try:
@@ -4822,7 +4825,7 @@ def getSpeaker(speaker: str):
             pass
         response = translateAI(
             speaker,
-            "Reply with the " + LANGUAGE + " translation of the NPC name.",
+            ctx("names.npc"),
             False,
         )
         try:
@@ -5033,7 +5036,7 @@ def finalizeSpeakerParse():
                 pass
             resp = translateAI(
                 to_translate,
-                "Reply with the " + LANGUAGE + " translation of the NPC name.",
+                ctx("names.npc"),
                 True,
             )
             try:

modules/wolfdawn.py

@@ -58,7 +58,8 @@ import traceback
 
 from colorama import Fore
 from tqdm import tqdm
-from util.paths import PROMPT_PATH, VOCAB_PATH
+from util.paths import VOCAB_PATH
+from util.skills import ctx, load_system_prompt
 from util.translation import (
     TranslationConfig,
     translateAI as sharedtranslateAI,
@@ -78,7 +79,7 @@ MODEL = os.getenv("model")
 TIMEOUT = int(os.getenv("timeout"))
 LANGUAGE = os.getenv("language").capitalize()
 
-PROMPT = PROMPT_PATH.read_text(encoding="utf-8")
+PROMPT = load_system_prompt()
 VOCAB = VOCAB_PATH.read_text(encoding="utf-8")
 LOCK = threading.Lock()
 MAXHISTORY = 10
@@ -143,7 +144,7 @@ def _batch_phase() -> str:
 
 def handleWolfDawn(filename, estimate):
     """Entry point used by the CLI/GUI dispatchers. Returns a summary string or 'Fail'."""
-    global ESTIMATE, TOKENS, FILENAME, SPEAKER_CONFIG, VOCAB
+    global ESTIMATE, TOKENS, FILENAME, SPEAKER_CONFIG, VOCAB, PROMPT
     ESTIMATE = estimate
     FILENAME = filename
     # Re-read workflow-configured settings so edits made this session take effect
@@ -153,6 +154,9 @@ def handleWolfDawn(filename, estimate):
     # that an earlier Phase 0 (names) harvested into vocab.txt.
     VOCAB = VOCAB_PATH.read_text(encoding="utf-8")
     TRANSLATION_CONFIG.vocab = VOCAB
+    # Reload base prompt + optional per-game quirks (DAZED_GAME_ROOT).
+    PROMPT = load_system_prompt()
+    TRANSLATION_CONFIG.prompt = PROMPT
 
     start = time.time()
     translatedData = openFiles(filename)
@@ -361,7 +365,7 @@ def getSpeaker(speaker: str):
 
     response = translateAI(
         speaker,
-        "Reply with the " + LANGUAGE + " translation of the NPC name.",
+        ctx("names.npc"),
     )
     translated = _normalize_speaker_name(
         response[0] if isinstance(response[0], str) else str(response[0])
@@ -371,7 +375,7 @@ def getSpeaker(speaker: str):
     if re.search(r"([a-zA-Z？?])", translated) is None:
         response = translateAI(
             speaker,
-            "Reply with the " + LANGUAGE + " translation of the NPC name.",
+            ctx("names.npc"),
         )
         translated = _normalize_speaker_name(
             response[0] if isinstance(response[0], str) else str(response[0])

util/paths.py

@@ -8,11 +8,21 @@ PROJECT_ROOT = Path(__file__).resolve().parent.parent
 DATA_DIR = PROJECT_ROOT / "data"
 VOCAB_PATH = DATA_DIR / "vocab.txt"
 VOCAB_BASE_PATH = DATA_DIR / "vocab_base.txt"
-PROMPT_PATH = DATA_DIR / "prompt.txt"
+SKILLS_DIR = DATA_DIR / "skills"
+# Runtime translation system skill (formerly data/prompt.txt).
+PROMPT_PATH = SKILLS_DIR / "system.md"
+LEGACY_PROMPT_PATH = DATA_DIR / "prompt.txt"
 LAST_UPDATE_SHA_PATH = DATA_DIR / "last_update_sha.txt"
 ENV_PATH = PROJECT_ROOT / ".env"
 ICON_PATH = PROJECT_ROOT / "assets" / "icon.png"
 ENGINE_ICONS_DIR = PROJECT_ROOT / "assets" / "engine_icons"
+TRANSLATION_CONTEXTS_PATH = DATA_DIR / "translation_contexts.json"
+# Per-game quirks skill (API overlay). Legacy flat file still migrated on load.
+GAME_QUIRKS_RELATIVE = Path("skills") / "quirks.md"
+LEGACY_QUIRKS_FILENAME = "translation_quirks.txt"
+GAME_SKILL_RELATIVE = Path("skills") / "translation.md"
+# Built-in skill filenames under <game>/skills/ (not user-custom overlays).
+GAME_SKILL_RESERVED_NAMES = frozenset({"quirks.md", "translation.md"})
 
 _ROOT_DATA_FILES = (
     "vocab.txt",
@@ -32,7 +42,19 @@ def migrate_root_data_files() -> None:
             src.rename(dst)
 
 
+def migrate_prompt_to_skills() -> None:
+    """Move legacy ``data/prompt.txt`` to ``data/skills/system.md`` when needed."""
+    SKILLS_DIR.mkdir(parents=True, exist_ok=True)
+    if PROMPT_PATH.is_file():
+        return
+    for legacy in (LEGACY_PROMPT_PATH, PROJECT_ROOT / "prompt.txt"):
+        if legacy.is_file():
+            legacy.rename(PROMPT_PATH)
+            return
+
+
 migrate_root_data_files()
+migrate_prompt_to_skills()
 
 
 def ensure_vocab_file() -> None:

util/skills/__init__.py

@@ -0,0 +1,29 @@
+"""Skill and prompt loaders (editable files under ``data/skills/``)."""
+
+from __future__ import annotations
+
+from util.skills.contexts import ctx, reload_contexts
+from util.skills.setup import load_project_setup, skills_dir
+from util.skills.system import (
+    custom_skill_path_for_game,
+    game_skill_path_for_game,
+    list_custom_skill_paths,
+    load_system_prompt,
+    migrate_game_skill_text,
+    quirks_path_for_game,
+    sanitize_custom_skill_stem,
+)
+
+__all__ = [
+    "ctx",
+    "custom_skill_path_for_game",
+    "game_skill_path_for_game",
+    "list_custom_skill_paths",
+    "load_project_setup",
+    "load_system_prompt",
+    "migrate_game_skill_text",
+    "quirks_path_for_game",
+    "reload_contexts",
+    "sanitize_custom_skill_stem",
+    "skills_dir",
+]

util/skills/contexts.py

@@ -0,0 +1,63 @@
+"""Load per-call translation history templates from data/translation_contexts.json."""
+
+from __future__ import annotations
+
+import json
+import os
+from typing import Any
+
+from util.paths import TRANSLATION_CONTEXTS_PATH
+
+_cache: dict[str, Any] | None = None
+_cache_mtime: float | None = None
+
+
+def _load() -> dict[str, Any]:
+    global _cache, _cache_mtime
+    path = TRANSLATION_CONTEXTS_PATH
+    if not path.is_file():
+        raise FileNotFoundError(f"Translation contexts file missing: {path}")
+    mtime = path.stat().st_mtime
+    if _cache is not None and _cache_mtime == mtime:
+        return _cache
+    data = json.loads(path.read_text(encoding="utf-8"))
+    if not isinstance(data, dict):
+        raise ValueError(f"translation_contexts.json must be an object: {path}")
+    _cache = data
+    _cache_mtime = mtime
+    return data
+
+
+def ctx(path: str, *, language: str | None = None, **kwargs: Any) -> str:
+    """Return a formatted history string for ``section.key``.
+
+    ``{language}`` defaults to the ``language`` env / argument (capitalized).
+    Extra kwargs fill other ``{placeholders}`` (e.g. ``context``).
+    """
+    if not path or "." not in path:
+        raise KeyError(f"Context path must be 'section.key', got {path!r}")
+    section_name, key = path.split(".", 1)
+    data = _load()
+    section = data.get(section_name)
+    if not isinstance(section, dict):
+        raise KeyError(f"Unknown translation context section: {section_name!r}")
+    template = section.get(key)
+    if not isinstance(template, str) or not template.strip():
+        raise KeyError(f"Unknown translation context key: {path!r}")
+
+    lang = language or os.getenv("language", "English")
+    lang = str(lang).capitalize()
+    values = {"language": lang, **kwargs}
+    try:
+        return template.format(**values)
+    except KeyError as exc:
+        raise KeyError(
+            f"Missing placeholder {exc.args[0]!r} for context {path!r}"
+        ) from exc
+
+
+def reload_contexts() -> None:
+    """Clear the in-memory cache so the next ctx() re-reads from disk."""
+    global _cache, _cache_mtime
+    _cache = None
+    _cache_mtime = None

util/skills/setup.py

@@ -0,0 +1,62 @@
+"""Load editable clipboard skills from data/skills/."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from util.paths import SKILLS_DIR
+
+_ENGINE_MARKERS = {
+    "rpgmaker": ("<!-- engine:rpgmaker -->", "<!-- /engine:rpgmaker -->"),
+    "wolf": ("<!-- engine:wolf -->", "<!-- /engine:wolf -->"),
+}
+
+
+def _read_skill_file(name: str) -> str:
+    path = SKILLS_DIR / name
+    if not path.is_file():
+        raise FileNotFoundError(f"Skill file missing: {path}")
+    return path.read_text(encoding="utf-8")
+
+
+def _extract_engine_section(text: str, engine: str) -> str:
+    """Keep shared body + the requested engine block; drop other engine blocks."""
+    start_tag, end_tag = _ENGINE_MARKERS.get(engine, (None, None))
+    if not start_tag:
+        return text
+
+    result_parts: list[str] = []
+    pos = 0
+    while pos < len(text):
+        next_starts = [
+            (text.find(marker[0], pos), eng, marker)
+            for eng, marker in _ENGINE_MARKERS.items()
+            if text.find(marker[0], pos) != -1
+        ]
+        if not next_starts:
+            result_parts.append(text[pos:])
+            break
+        next_starts.sort(key=lambda x: x[0])
+        idx, eng, (s_tag, e_tag) = next_starts[0]
+        result_parts.append(text[pos:idx])
+        end_idx = text.find(e_tag, idx)
+        if end_idx == -1:
+            raise ValueError(f"Unclosed engine block {s_tag} in skill file")
+        block_body = text[idx + len(s_tag) : end_idx]
+        if eng == engine:
+            result_parts.append(block_body.strip("\n") + "\n")
+        pos = end_idx + len(e_tag)
+    return "".join(result_parts).strip() + "\n"
+
+
+def load_project_setup(engine: str = "rpgmaker", *, prepend: str = "") -> str:
+    """Load the Project Setup clipboard skill for *engine*."""
+    raw = _read_skill_file("project_setup.md")
+    body = _extract_engine_section(raw, engine)
+    if prepend:
+        return prepend.rstrip() + "\n\n" + body
+    return body
+
+
+def skills_dir() -> Path:
+    return SKILLS_DIR

util/skills/system.py

@@ -0,0 +1,153 @@
+"""Load the static system prompt and optional per-game quirks overlay."""
+
+from __future__ import annotations
+
+import os
+import re
+from pathlib import Path
+
+from util.paths import (
+    GAME_QUIRKS_RELATIVE,
+    GAME_SKILL_RELATIVE,
+    GAME_SKILL_RESERVED_NAMES,
+    LEGACY_QUIRKS_FILENAME,
+    PROMPT_PATH,
+)
+
+_SKILL_NAME_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$")
+
+
+def quirks_path_for_game(game_root: str | Path | None) -> Path | None:
+    """Return ``<game_root>/skills/quirks.md`` when *game_root* is set.
+
+    Also migrates a legacy ``translation_quirks.txt`` in the game root if present.
+    """
+    if not game_root:
+        return None
+    root = Path(game_root)
+    preferred = root / GAME_QUIRKS_RELATIVE
+    legacy = root / LEGACY_QUIRKS_FILENAME
+    if not preferred.is_file() and legacy.is_file():
+        preferred.parent.mkdir(parents=True, exist_ok=True)
+        try:
+            legacy.rename(preferred)
+        except OSError:
+            # Fall back to reading the legacy path if rename fails.
+            return legacy
+    return preferred
+
+
+def game_skill_path_for_game(game_root: str | Path | None) -> Path | None:
+    """Return ``<game_root>/skills/translation.md`` when *game_root* is set."""
+    if not game_root:
+        return None
+    return Path(game_root) / GAME_SKILL_RELATIVE
+
+
+def migrate_game_skill_text(text: str) -> str:
+    """Rewrite legacy quirk path names inside a game skill markdown body."""
+    if not text:
+        return text
+    # Common stale pointers from older Project Setup output.
+    updated = text.replace("translation_quirks.txt", "skills/quirks.md")
+    updated = updated.replace("`translation_quirks.md`", "`skills/quirks.md`")
+    # Bare relative quirks.md without skills/ prefix (keep skills/quirks.md intact).
+    updated = re.sub(
+        r"(?<!skills/)(?<![\w./-])quirks\.md(?![\w.-])",
+        "skills/quirks.md",
+        updated,
+    )
+    # Avoid double skills/skills/
+    updated = updated.replace("skills/skills/quirks.md", "skills/quirks.md")
+    return updated
+
+
+def game_skills_dir(game_root: str | Path | None) -> Path | None:
+    """Return ``<game_root>/skills`` when *game_root* is set."""
+    if not game_root:
+        return None
+    return Path(game_root) / "skills"
+
+
+def sanitize_custom_skill_stem(name: str) -> str | None:
+    """Return a safe skill stem (no ``.md``) or None if invalid/reserved."""
+    raw = (name or "").strip()
+    if not raw:
+        return None
+    if raw.lower().endswith(".md"):
+        raw = raw[:-3].strip()
+    if not _SKILL_NAME_RE.match(raw):
+        return None
+    filename = f"{raw}.md"
+    if filename.lower() in {n.lower() for n in GAME_SKILL_RESERVED_NAMES}:
+        return None
+    return raw
+
+
+def custom_skill_path_for_game(
+    game_root: str | Path | None, name: str
+) -> Path | None:
+    """Return ``<game_root>/skills/<name>.md`` for a valid custom skill name."""
+    stem = sanitize_custom_skill_stem(name)
+    skills = game_skills_dir(game_root)
+    if not stem or skills is None:
+        return None
+    return skills / f"{stem}.md"
+
+
+def list_custom_skill_paths(game_root: str | Path | None) -> list[Path]:
+    """List user custom skill markdown files under ``<game>/skills/``.
+
+    Excludes built-in ``quirks.md`` and ``translation.md``.
+    """
+    skills = game_skills_dir(game_root)
+    if skills is None or not skills.is_dir():
+        return []
+    reserved = {n.lower() for n in GAME_SKILL_RESERVED_NAMES}
+    out: list[Path] = []
+    for path in sorted(skills.glob("*.md"), key=lambda p: p.name.lower()):
+        if path.name.lower() in reserved:
+            continue
+        if path.is_file():
+            out.append(path)
+    return out
+
+
+def load_system_prompt(game_root: str | Path | None = None) -> str:
+    """Load ``data/skills/system.md`` plus optional per-game overlays.
+
+    Overlay order:
+      1. ``skills/quirks.md`` (voice quirks)
+      2. Any other ``skills/*.md`` except ``translation.md`` (custom skills)
+
+    Path resolution order for the game root:
+      1. Explicit *game_root* argument
+      2. ``DAZED_GAME_ROOT`` environment variable
+    """
+    base = ""
+    if PROMPT_PATH.is_file():
+        base = PROMPT_PATH.read_text(encoding="utf-8")
+
+    root = game_root or os.getenv("DAZED_GAME_ROOT") or ""
+    root = str(root).strip()
+    if not root:
+        return base
+
+    parts = [base.rstrip()] if base.strip() else []
+
+    quirks_file = quirks_path_for_game(root)
+    if quirks_file and quirks_file.is_file():
+        quirks = quirks_file.read_text(encoding="utf-8").strip()
+        if quirks:
+            parts.append("## Game-Specific Translation Quirks\n\n" + quirks)
+
+    for path in list_custom_skill_paths(root):
+        body = path.read_text(encoding="utf-8").strip()
+        if not body:
+            continue
+        title = path.stem.replace("_", " ").replace("-", " ").strip() or path.stem
+        parts.append(f"## Custom Game Skill: {title}\n\n{body}")
+
+    if not parts:
+        return base
+    return "\n\n".join(parts) + "\n"

util/translation.py

@@ -1477,8 +1477,8 @@ class TranslationConfig:
         # Load prompt and vocab files if not provided
         if prompt is None:
             try:
-                from util.paths import PROMPT_PATH, VOCAB_PATH
-                self.prompt = PROMPT_PATH.read_text(encoding="utf-8")
+                from util.skills import load_system_prompt
+                self.prompt = load_system_prompt()
             except FileNotFoundError:
                 self.prompt = ""
         else:
@@ -1945,7 +1945,7 @@ def createContext(config, subbedText, formatType, history=None):
     cache invalidation caused by changing vocabulary matches.
 
     Cached in static_system:
-      - prompt.txt content
+      - data/skills/system.md content (plus optional game quirks overlay)
 
     Dynamic in vocab_text:
       - only vocab terms found in the current batch text
@@ -2061,7 +2061,7 @@ def buildClaudeRequest(system, user, history, formatType, model, numLines=None,
 def translateText(system, user, history, penalty, formatType, model, numLines=None, vocab_text=""):
     """Send translation request to the selected API.
 
-    system:     Static system prompt (prompt.txt). Cached by Claude.
+    system:     Static system prompt (data/skills/system.md). Cached by Claude.
     vocab_text: Per-batch vocabulary (dynamic, never cached to avoid cache busting).
     """
     # Ensure system content is not empty
@@ -2830,7 +2830,7 @@ def translateAI(text, history, config, filename=None, pbar=None, lock=None, mism
 
             continue
 
-        # Create context — static_system is the stable prompt.txt content;
+        # Create context — static_system is the stable system.md content;
         # vocab_text is the per-batch matched vocabulary (dynamic).
         static_system, vocab_text, user = createContext(config, subbedT, formatType, history)