Character Skills — per-path voices and exercises for the quest-perfection loop

The quest-perfection loop always works one (character, level) slice: the walk lane (quest-walkthrough skill → quest-walker agent) plays it as a learner, and the fix lane (quest-fix skill → quest-fixer agent) repairs what the walk witnessed. Until now "as a learner of this character class" was a single sentence of instruction; the character skills make it a real, versioned contract — one skill per character path, holding the persona, the voice, the per-level exercise checkpoints, and the friction lens for that path.

The files

One skill per _data/quests/paths.yml entry, named by the path key:

.claude/skills/quest-character-<key>/SKILL.md     # developer, system-engineer,
                                                  # security-specialist, data-scientist,
                                                  # digital-artist, game-developer

Each skill has the same shape, so the fleet reads them uniformly:

section	consumer	content
Who you are playing	walk lane	the persona: background, starting knowledge, what costs this learner confidence
Path roadmap	both	generated — the path's levels/themes/tiers/XP from paths.yml + the registry
Voice	walk lane	the character's fantasy dialect for report narration, plus the two-layer voice rule
Walk lens	walk lane	character-specific friction checks beyond the generic rubric
Per-level exercises	both	per level on the path: competence checkpoints to verify + what fixes must preserve
Fix lens	fix lane	which repairs serve this learner; guardrails against flavor injection
Hard rules	both	data sources, precedence, and the never-expands-scope clause

How the loop consumes them

Agents in the loop have no Skill tool — they Read the file. The resolution is deterministic: walk-plan.json carries character.key, and the sheet lives at .claude/skills/quest-character-<character.key>/SKILL.md.

• Walk lane — quest-walkthrough step 3 now starts by putting on the character sheet: role-play the persona while chain-reading, verify the level's exercise checkpoints, judge continuity against them, and narrate the session report in the character's dialect. A missing sheet is noted in the report's method section, never a failure.
• Fix lane — quest-fix step 3 reads the sheet before the first edit: the fix lens ranks repairs by what serves this path's learner, and the per-level "preserve" lists guard against score-gaming deletions. The sheet informs edit choice only — the M-rules (smallest edit, deterministic keep/revert, content-only surface) are unchanged and always win.

The two-layer voice rule

Levels are shared across paths (level 0100 is walked by developer, digital-artist, and game-developer alike), so quest content keeps the one neutral quest-fantasy brand voice (_data/brand/voice.yml, _data/brand/sections/quest.md). The character dialect (forge-and-craft, wards-and-watch, atelier-and-canvas, …) colors only the walkthrough report's narration and the judgment of fit. Every sheet states this rule, because the failure mode it prevents is real: a fix lane that rewrites shared quests into one class's flavor degrades them for the other five paths.

Consistency machinery (why the sheets can't drift)

The facts in a sheet come from the same sources the planner resolves slices from, and a generator keeps them mechanically identical:

• scripts/quest/character_skills.py regenerates each sheet's roadmap block (between <!-- BEGIN GENERATED: character-roadmap --> markers) from _data/quests/paths.yml + scripts/quest/quest_registry.py. Everything outside the markers is hand-authored and never touched.
• --check verifies: every path has a sheet, no orphan sheets for removed paths, roadmap blocks are byte-fresh, the authored exercises cover exactly the path's levels (one ### … Level <code> heading each), and the frontmatter name: matches the directory.
• make quest-skills (regenerate) / make quest-skills-check (verify); quest_audit.py runs the same check as its non-gating skills layer — advisory because the audit's PR triggers don't watch paths.yml, so a gating check could red an unrelated content PR. Drift heals with one make target.
• Deliberately not part of make quest-data: the fix lane runs quest-data and must never write .claude/** (fix PRs stay content-only, M4).

Changing the paths

Adding, removing, or re-leveling a character in paths.yml:

1. Edit _data/quests/paths.yml (it is curated, not generated).
2. Run make quest-skills — stale roadmaps regenerate; a new path is reported as a missing sheet.
3. For a new path, author the sheet by copying an existing one's section skeleton, then write the persona/voice/lenses and one exercises heading per level; make quest-skills fills the roadmap and make quest-skills-check proves coverage.
4. A removed path's sheet is reported as an orphan — delete the directory.

What this deliberately is not

• Not one skill per quest. ~230 per-quest files would drift with every content PR; the concrete quest list already reaches every run through walk-plan.json, and quest-grain expectations live in the per-level exercise checkpoints.
• Not a workflow change. The loop's workflows already say "use the quest-walkthrough / quest-fix skill"; the sheets ride in through those skills' step 3, so nothing under .github/workflows/ moved.
• Not new authority. A sheet never expands an agent's writable surface, never overrides quest.instructions.md, the brand guides, or an M-rule, and never outranks walk-plan.json on what the slice contains.