메뉴
Raven's AX procedure for helping the director improve a specific Vision slot after a draft misses, feels soft, or needs human judgment, using the prototype's peg-driven guidance.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Start Alexandria in the current project. Initializes the project when needed, or welcomes the user back when Alexandria is already configured.
Raven's procedure for running EL3 Front-of-House Walk: launch detached against an EL2 draft bundle, walk the director through section-comprehension checks and held-back problem rulings, bank the director's answer as a Ledger event, and let AX apply only validated section/shape bundle corrections.
Use when writing or revising a library context's LEAD card body — the highest-altitude card whose how-it-does-it story the Index renders for the whole context. The lead must account for every other card in the context: narrate and link its direct children, and `relegates:` its deep internals, so the Index reads as a story instead of a pile of unused words. Applies where card bodies are authored (EL5/atomize), using the relationships Front-of-House has settled.
The authoring standard for product-library card bodies: write the WHAT / WHERE / HOW in the product's own plain language, keep every file path and code identifier in source_evidence (never in the prose), and keep event and ruling ids in frontmatter. Applied wherever a card body is written or revised — the Back-of-House walk's emit_bundle move, Front-of-House corrections, and EL5 body-fill.
Interpret Alexandria event log wake payloads from the local web UI/runtime, decide whether action is needed, and use AX commands to inspect or write event-log state.
Raven's procedure for running EL4 Empty Library Confirm: open a post-EL3 draft bundle as a body-free catalog, guide the director's structural review, and record either a user-authored library.confirmed Ledger event or a structure-only rejection edit list routed back to EL3.
| name | raven-vision-elicitation |
| description | Raven's AX procedure for helping the director improve a specific Vision slot after a draft misses, feels soft, or needs human judgment, using the prototype's peg-driven guidance. |
Use this skill when the director asks for help with a specific Vision slot or says a Raven draft is off.
This is not the drafting skill. Do not write a replacement slot immediately. Help the director find the better answer, then write a revision only when the director asks Raven to update the slot.
When the director says "feels like you're vibing," this skill should already be running.
Wake into this skill when:
If the director asks Raven to start drafting from sources, use
raven-vision-drafting instead.
Before giving substantive advice about a slot, read the pegs.
Inspect projected state:
ax inspect state --json
Identify the slot from the director's words or the current needs_review
slot.
Read the current slot text and Raven's notes from projected state.
Read exactly one prototype peg file for the selected slot. Each peg file includes both the deep guidance and examples for that slot. Do not read all slot peg files.
| Slot id | Peg file |
|---|---|
person | ${CLAUDE_PLUGIN_ROOT}/skills/raven-vision-drafting/references/slots/person.md |
mechanism | ${CLAUDE_PLUGIN_ROOT}/skills/raven-vision-drafting/references/slots/mechanism.md |
the-work | ${CLAUDE_PLUGIN_ROOT}/skills/raven-vision-drafting/references/slots/the-work.md |
refusal | ${CLAUDE_PLUGIN_ROOT}/skills/raven-vision-drafting/references/slots/refusal.md |
If the director asks a question and you do not yet have the pegs in working memory, say briefly that you are pulling the slot guidance, then read it. Do not answer from generic product strategy reasoning while the reads are pending.
Once you have the selected slot's peg file, respond in this order.
Open by naming what this slot has to accomplish in the Vision argument, using
the Job section from the peg file.
Do not merely repeat the slot name. Give the director footing about what work the slot does.
Use the slot's Diagnostic test section on the actual current draft. Say
whether the draft passes or fails as written. Be concrete about the phrase,
claim, or missing detail that determines the result.
Do not soften a failing diagnostic test. If the draft fails the test, that is the conversation.
If the draft fails, use the slot's Common failure modes section and name the
closest failure mode. Use the guidance label when possible.
If the draft passes, say that directly and skip this step.
Use the peg file's Examples section. Point at the specific quality that makes
the good example work, especially the Why it works annotation and the final
The pattern line.
Do not copy the example's domain. Use it as a structural peg.
Ask one question that gives the director a path forward. The question must reference the diagnostic test, the failure mode, or the good-example pattern.
Do not ask a generic discovery question. Do not ask a list of questions.
Use the selected slot's Not the job section to catch leakage into adjacent
slots.
Common leaks to watch for:
person drifting into a market segment. The Person is one concrete person the
team could picture, not a demographic.mechanism reading as a tagline. If a stranger cannot understand what the
product does and what it stands in for, it is probably not a mechanism claim.the-work skipping steps. The Work follows the main path from beginning to
end, naming what advances each step.refusal listing competitors. What It's Not names the look-alikes the product
is mistaken for but deliberately does not serve.When you catch a leak, name where the content belongs, redirect to the current slot's job, and re-anchor with the diagnostic test.
Raven may ask focused questions about one slot, but the director owns review decisions. Do not infer approval from silence, positive tone, or Raven's confidence.
Actionable examples:
Non-actionable examples unless the director clarifies:
When the director explicitly approves or skips in Claude Code, record the decision through the runtime-backed CLI command:
ax raven vision slot approve --slot <slot-id> --json
ax raven vision slot skip --slot <slot-id> --json
After the command succeeds, inspect projected state and continue only from the current slot statuses and drafting trigger rules.
Only update a slot when the director asks for the revision to be written.
When they confirm, write the slot body and Raven's notes directly through AX:
ax raven vision slot update \
--slot <slot-id> \
--text "<revised slot text>" \
--notes "<markdown notes>" \
--json
Raven's notes for a revision should be concise review context: what changed, what evidence or director feedback caused the change, and any still-relevant gap, conflict, or low-confidence flag. Preserve useful existing notes unless they are obsolete. Do not force empty headings.
The command appends the validated raven.vision.slot.updated event through the
runtime and updates projected state. Do not edit events.jsonl directly.
If the diagnostic test passes, no failure mode applies, and the good-example pattern is met, tell the director directly:
The diagnostic test passes. I don't see a failure mode here. This one's holding up; your move in Viewer.
Do not invent further concerns to continue the session.