onboard - SKILL.md Agent Skill

name: onboard description: One-time researcher onboarding interview that compiles a `taste/` corpus (SOUL.md, STYLE.md, METHODOLOGY.md, DISMISSALS.md, MEMORY.md, INDEX.md) capturing the researcher's intellectual taste, beliefs, and values. Auto-injects INDEX.md into every future session via SessionStart. Triggers on first install ("no taste corpus found"), explicit `/onboard`, `/onboard --deep` for Tier 2 voice laddering, or `/onboard --redo` / `/onboard --section <name>` to refresh.

/onboard — researcher taste interview

The marquee onboarding skill. Produces a durable, agent-injectable record of who the researcher is, what they believe, how they judge work, and what dead ends they refuse to revisit.

Sister to AGENTS.md. AGENTS.md says what to do in this repo. The taste/ corpus says who this researcher is — installed per-repo today, with cross-repo portability planned (symlink ~/.rockie/taste/ → <repo>/.rockie/taste/ is the recommended manual interim).

When to run

First-run — SessionStart hook flags absence of .rockie/taste/INDEX.md; agent suggests /onboard.
Explicit invocation — user types /onboard.
Refresh — /onboard --redo rewrites everything; /onboard --section <soul|style|methodology|dismissals> refreshes one file.
Deep mode — /onboard --deep runs Tier 2 (20–30 min, voice-first laddering). Otherwise Tier 1 (5–7 questions, ~5 min).

What the skill does

Loads prompts/interviewer-tier1.md (or interviewer-tier2.md if --deep) into the active agent's system context.
The agent runs the interview turn-by-turn following the prompt's phase-gated state machine. The user answers via text or /voice tap.
As the interview progresses, the agent writes intermediate state to .rockie/taste/.draft.json after every turn (so the session can be resumed if interrupted).
When coverage reaches the end-condition, the agent emits a summary round, asks the user to confirm or correct, then runs runtime/compile.py to produce the six markdown files in .rockie/taste/.
The agent opens INDEX.md in the user's editor for final review, then commits the corpus.

Hard rules

One question per turn. The interviewer prompt enforces this with a question-mark self-check; the agent must respect it.
Never press a dodged question. Mark SOFT in the scratchpad and move on. Pressing produces confabulation, not truth.
Never silently rewrite an existing taste file. /onboard --redo is explicit; the agent cannot decide to overwrite SOUL.md unprompted. (MEMORY.md and DISMISSALS.md are append-only and may be written to by Stop-hook flow, not this skill.)
Never invent answers. If the user is short on a topic, the generated file marks the topic as [SOFT — needs deepening] rather than embellishing.
Voice opt-in only. Tier 1 is text-default; Tier 2 prompts the user to /voice tap for each answer if they want, but never requires it.

Outputs

.rockie/taste/
├── INDEX.md           # auto-injected every session via SessionStart
├── SOUL.md            # identity, worldview, hot takes
├── STYLE.md           # written + spoken voice
├── METHODOLOGY.md     # rigor standards, success criteria
├── DISMISSALS.md      # append-only "never again" log
├── MEMORY.md          # append-only session-experience log
└── .draft.json        # interview transcript (gitignored)

The corpus is gitignored by default (personal — like .env). Users can opt to commit it by removing the entry from .rockie/taste/.gitignore. (For shared-team taste, see future roadmap entry on team-level taste compositing.)

Agent invocation protocol

When this skill fires, the agent should:

Resolve which mode: Tier 1 (default), Tier 2 (--deep), refresh (--redo or --section).
Read prompts/interviewer-tier1.md (or tier2) — this is the active system prompt for the duration of the interview.
Begin Turn 1 by greeting the user briefly and asking the first question. Do NOT explain the schema; the prompt is structured so the user discovers what they're contributing turn by turn.
After each user answer, regenerate the full scratchpad (do not append to it) following the MIRROR structure in the prompt.

Persist .draft.json after every answer. Schema:

{ "started_at": "...", "tier": 1, "topics": {
    "identity": {"q": "...", "a": "...", "soft": false, "ladder": []},
    ...
  }, "coverage": {"identity": 2, "soul": 1, ...} }

End-condition: all topics ≥ 1 AND (all topics ≥ 2 OR seven questions asked). At end-condition, summarize captured signal per file, ask user to confirm or correct, then compile.
Run python3 .claude/skills/onboard/runtime/compile.py .rockie/taste/.draft.json — this writes the six markdown files from templates and the draft JSON.
Open .rockie/taste/INDEX.md for the user to review and edit.

Failure modes

User goes off-topic for 3+ turns — agent bridges back per the prompt's anti-drift rule.
User signals exhaustion ("can we wrap up?") — agent honors; emits compile from current coverage; marks low-coverage files [SOFT — needs deepening].
Voice transcript is empty / garbled — agent asks user to type this answer instead.
Compile script errors — agent reports the error, retains .draft.json, and suggests re-running /onboard (the existing draft is loaded from disk and the interview continues from where it stopped).
Partial draft (interview interrupted before 3 topics covered) — compile.py exits 3 with a "refusing to compile" message. Agent re-runs /onboard to continue the interview rather than shipping a placeholder corpus.

Re-runs

/onboard --redo archives the existing corpus to .rockie/taste/.archive/<timestamp>/ and starts fresh. Identity shifts deserve audit trail.

/onboard --section <name> runs only the questions for that section, re-compiles only that file, archives the prior version.