meta-prompt-builder
Generates structured prompts to send to each agent in an agent dispatch plan. Triggered by "write the instructions", "what do I say to the agent?", "write the prompt for me", "meta-prompt-builder". Bridges agent-composer (which agents) and prompt content (what to say). Uses Goal/Context/Constraints/Done When structure.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 33
- Forks
- 1
01 前言 Front Matter
- name
- meta-prompt-builder
- model-note
- session-inherit — Sonnet base is first-class (sonnet_floor_doctrine.md); depth-critical judged steps route to dispatch (opus agent / cross-family sidecar, consent-gated), never a substrate requirement
- description
- Generates structured prompts to send to each agent in an agent dispatch plan. Triggered by "write the instructions", "what do I say to the agent?", "write the prompt for me", "meta-prompt-builder". Bridges agent-composer (which agents) and prompt content (what to say). Uses Goal/Context/Constraints/Done When structure.
- allowed-tools
- [ "Read", "Bash", "Glob", "Grep" ]
- user-invocable
- true
02 资源文件 1
- 参考文档SKILL_detail.md1.7 KB
03 SKILL.md 正文
meta-prompt-builder — Prompt Delegation Skill
When agent-composer decides "what to orchestrate", meta-prompt-builder designs "what to say to each agent." Delegates prompt writing itself to AI — interprets Goal / Context / Constraints / Done When structure within the FH context.
Primary Users
agent-composer proficients — People who already have an orchestration plan and want to structure the specific instructions for each agent. Use this when you already know "which agents to use."
Beginners (unsure which agents to use) → start with agent-composer first. meta-prompt-builder is a post-orchestration step.
Triggers
/meta-prompt-builder- Automatically suggested immediately after agent-composer outputs an orchestration plan when user says "how do I instruct each one specifically?"
- "Write the prompts per Wave", "What should I say to each agent?", "Design the instructions"
- "I don't know what to tell the agents"
- "Tell me how to instruct this task so it's handled well"
- "Write the prompt for me"
- "Tell me how to request this from Claude"
- "How should I write the instructions?"
Step 0. Receive Task
Input: natural language task description + (optional) agent-composer orchestration plan
If no input is provided, ask for 4 fields:
What task are you designing a prompt for?
- Goal: What are you trying to achieve?
- Context: What project, files, or harness state is involved?
- Constraints: What should be avoided? Are there budget or time limits?
- Done When: When can you say "it's done"?
Natural Language Goal → Auto-suggest Recommended Chain
When Goal is provided in natural language (e.g., "I need to report to the team lead"), suggest a recommended chain using the keyword mapping below.
| Goal keyword | Recommended chain | Reason |
|---|---|---|
| "report", "team lead", "executive", "CTO", "persuade", "review" | apex-review → (sim-conductor if needed) | Decision-maker simulation then quality validation |
| "analyze", "diagnose", "something seems off" | harness-doctor → context-doctor | Structural + contextual diagnosis simultaneously |
| "simulate", "validate", "meta" | sim-conductor (D-code or Area B) | Multi-perspective validation |
| "install", "setup", "onboarding" | plugin-recommender → install-wizard | Recommend then install |
| "review", "code check" | sim-conductor D-code → harness-pr-reviewer | Code review chain |
Output format:
Recommended chain: {Goal keyword} →
Wave 1: {Agent A} — {1-line reason}
Wave 2: {Agent B} — {1-line reason}
→ Proceed with this chain? (Y: immediately orchestrate with agent-composer / N: enter 4-field directly)
Step 1. Reinterpret as FH 4-field
Structure the input task according to FH context.
| Field | FH interpretation |
|---|---|
| Goal | The harness state change this agent combination must achieve |
| Context | Harness version + latest sim result M/S tier + target file paths |
| Constraints | Simplification guard + no self-recursion + no inter-Wave file conflicts |
| Done When | M-tier 0 items OR specified agent fan-in complete OR N naming candidates produced |
Step 2. Generate Per-Agent Prompt Drafts
Receive agent-composer orchestration plan as input and generate instructions to deliver to each agent per Wave.
Format:
Wave {N}-{ID}: [{agent name}]
Goal: {1 line}
Context: {harness state in 3 lines max}
Constraints: {prohibited items in 2 lines max}
Done When: {evaluation criterion in 1 line}
───────────────────────────────
Actual instruction:
"{Natural language instruction for the agent — 2–4 sentences}"
When called standalone without an agent-composer plan: collect 4-field directly via Step 0 questions, then generate a single-agent prompt.
Step 3. Quality Self-Validation
Before outputting the generated prompt draft, check against 3 criteria.
-
Goal achievability: Is the Goal achievable within the agent's allowed-tools range?
-
Done When measurability: Is Done When a measurable event in the M/S/R framework?
-
Context alignment: Does the Context match the "target" scope defined in the agent's SKILL.md?
-
Restate check (cold reader) — the procedure for the judged Done-When row below («read one draft cold»). Hand each draft — its text only, no conversation, no repo — to a separate reader (a fresh subagent, Sonnet or Haiku, or a different model family) and ask it to restate GOAL (what is true when finished) · DONE (the observable check) · CONSTRAINTS (what not to touch), marking each
RESTATABLEorGUESS, and to list the questions it would need answered. AnyGUESSfield is[WARN: brief not restatable — <field>]; its questions are the fix list. Do not tell the reader which draft you suspect. Restating your own draft yourself does not count — the failure this answers is the author filling gaps from context the receiver never gets. Record the reader's model and keep its raw reply next to the draft. Dispatching the reader follows the usual dispatch consent; if no reader can be dispatched, mark the rowNOT RUN, neverRESTATABLE. Measured once (2026-11-14, one reader per brief, briefs written by the same author —UNCALIBRATED): 4 briefs made vague on purpose → 4/4 not restatable on all three fields; 4 briefs meant to be clear → 3/4 restatable, the fourth flagged an undefined term the author then judged real. This shows the check flags briefs made vague, not a false-alarm rate. Idea from the comprehension stage of https://github.com/google/elephant-goldfish-model ; the procedure here is FH's own.
Items that fail the check are marked [WARN] and the decision is delegated to the user.
Step 3 bias prevention: When checking Goal achievability, Read the agent's SKILL.md to directly verify (directly compare allowed-tools list). Since generator and validator are the same LLM, self-validation without external reference repeats bias.
Done When completeness check: Even if Done When is measurable in the M/S/R framework, the following 3 must also be stated to pass without WARN:
- Measurement subject (which agent's output is the basis)
- Measurement timing (absolute basis for this run vs relative to previous run)
- Version / baseline (harness-doctor version, sim-conductor Area, etc.)
Done When Ambiguity Patterns — Automatic WARN Triggers
If Done When contains any of the following patterns, automatically emit [WARN: Done When not measurable].
| Ambiguity type | Example expression | WARN reason |
|---|---|---|
| Subjective completion | "if it goes well", "if it's done properly", "if it feels right" | No binary judgment criterion |
| Speed without numbers | "quickly done", "finished soon" | No count/time criterion |
| Absence description | "if there are no issues", "if there are no errors" | Cannot count as M/S/R |
| Unconditioned completion | "when it's all done", "when finished" | What state counts as complete is undefined |
Acceptable form examples:
- "M-tier 0 items" ✅
- "harness-doctor fan-in complete + S-tier 3 or fewer" ✅
- "Scenario 1 PASS verdict received" ✅
Integration with Existing FH Assets
| Asset | Integration direction |
|---|---|
| agent-composer | Input source: orchestration plan → meta-prompt-builder elaborates Wave-by-Wave prompts |
| sim-conductor | Output validation: generated prompts can be quality-validated via D-code mode |
| fact-checker | Wave 0 gating: check if generated prompts overlap or conflict with existing assets |
| field-harvest | Context field input source: field patterns can be reflected in Constraints |
| harness-doctor | Done When cross-validation: verify alignment with harness-doctor checklist |
Meta-Simulation Validation Criteria (required after onboarding)
Detail: See
SKILL_detail.md §Meta-Simulation Validation Criteria— 3-axis functional/performance measurement (goal achievement / context alignment / constraint validity) + 3 validation scenarios + marketplace-gate onboarding gate — read after onboarding when validating the skill itself, not per-invocation.
Done When
This skill's own Done When must satisfy the 3-part completeness check it imposes on others
(measurement subject · measurement timing · version/baseline). The previous version —
All steps 0–3 completed — met 0 of 3 and matched its own Unconditioned completion WARN pattern.
Measurement subject: this skill's own output — the prompt draft set
produced in this invocation (not the dispatched
agents' results, which this skill never sees)
Measurement timing: absolute, at end of this run — no comparison to a
previous run is implied or required
Version / baseline: the Step 3 checklist and the WARN-trigger table in
THIS file; note the SKILL.md revision used if the
table has changed since the draft was written
☐ Steps 0–3 each produced their named artifact — a step with no
artifact is incomplete, not "completed" (measured: 4 steps, 4 artifacts)
☐ One prompt draft exists per agent named in the dispatch plan (measured: drafts == agents
in plan)
☐ Every draft carries all four fields (Goal / Context /
Constraints / Done When), none empty (measured: 4 fields × N drafts,
0 empty)
☐ Each draft's own Done When was run through the 3-part
completeness check and the WARN-trigger table, with the
result recorded per draft (mandatory-pass)
☐ Every Goal was checked against the target agent's declared
allowed-tools by READING that agent's SKILL.md — not from
recall (mandatory-pass)
☐ [WARN] items are surfaced to the user with the reason, and
the decision is the user's (mandatory-pass)
☐ Drafts are handed over for review, not dispatched (mandatory-pass)
☐ Step 3 item 4 ran on every draft with a separate reader;
each GUESS field is a recorded [WARN] or was fixed and re-read (mandatory-pass — NOT RUN is
recorded, never counted as pass)
☐ The drafts are usable by the target agent as written (judged — adversarial pairing:
generator and validator are the
same model, so self-validation
repeats its own bias. Have the
target agent, or a different
model, read one draft cold and
state what it would do; a
divergence from the intended
task is a FAIL)
Simplification Guard
- meta-prompt-builder does not execute agents directly. Prompt design only.
- Requests without Done When must be confirmed before proceeding.
- Generated prompts are always reviewed by the user before execution.
04 同仓库更多 Skills
todo-orchestrator
TODO CLI 프로젝트의 기능 추가·버그 수정·리팩터링·테스트 보강을 개발자와 QA 에이전트로 조율한다. todo 앱 관련 코드 작업 요청에 사용한다. 다시 실행, 재실행, 수정, 보완, 이전 결과 개선, QA만 다시 같은 후속 요청에도 사용한다. 단순 질문에는 사용하지 않는다.
todo-dev
TODO CLI(todo/cli.py)에 명령을 추가하거나 버그를 고치고 테스트를 작성하는 방법. todo 앱 코드 수정·기능 추가·테스트 작성·리팩터링 작업에 사용한다. 수정, 보완, 다시 구현 요청에도 사용한다.
todo-dev-team
TODO CLI(todo/cli.py) 개발 작업을 개발자·테스터·리뷰어 팀으로 진행하는 오케스트레이터. 명령 추가, 버그 수정, 테스트 작성, 리팩터링, 커밋·푸시 요청에 사용한다. 다시 실행, 재실행, 수정, 보완, 리뷰 반영, 테스트만 다시 같은 후속 요청에도 사용한다. 단순 질문에는 쓰지 않는다.
todo-dev
TODO CLI 개발 작업(명령·기능 추가, 버그 수정, 테스트 작성, 리뷰, 커밋·push)을 개발자·테스터·리뷰어 에이전트 팀으로 진행하는 오케스트레이터. todo CLI, todo.json, todo/cli.py 관련 구현·수정·테스트 요청에 사용한다. 재실행, 수정, 보완, 이전 결과 개선, 테스트만 다시, 리뷰만 다시 요청에도 사용한다. 단순 질문에는 쓰지 않는다.
weekly-summary-write
정리된 회의록으로 주간 회의 요약 문서를 작성하는 형식과 문체 규칙. 주간 요약 작성, 요약 문서 수정·보완, 검토 의견 반영 시 사용한다.
weekly-meeting-summary
주간 회의록을 정리하고 요약 문서를 만드는 전체 작업을 조율하는 오케스트레이터. 회의록 요약, 주간 회의 정리, 주간 보고용 요약 요청 시 사용한다. 다시 실행, 재실행, 요약 수정·보완, 이전 결과를 바탕으로 한 개선, 정리만/검토만 다시 같은 후속 요청에도 사용한다.