agentleFS
kevin-burnskevin-burns/claude-skills/spec-ready/SKILL.md

spec-ready

Skill11 starsChanged 3 months agoMITView on GitHub

Check that a plan is ready to spec before anyone writes one, then write the PRD (why) and the spec (how). A script reads the person's own turns from the session transcript and has an independent verifier confirm each required item (goal, who it is for, non-goals, anti-goal, success check, verification strategy, boundaries touched); anything missing comes back as one round of questions, least-certain first. Use after an interview or brainstorm and before an implementation plan: "is this ready to spec", "write the PRD", "turn this into a spec", "are we done planning". Not an interviewer from scratch (frontier-rounds, brainstorming) and not the implementation plan (writing-plans).

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 2 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    No invented facts: Every decision line in the PRD cites its turn as [turn:<id>].

    line 23
  2. Distinctive:

    Ask: When items are open, ask them all in one round, least-certain first, in the order the gate printed, each with your recommendation:

    line 49
  3. Very specific:

    Run gate.py override only when the person says so: when they tell you a verifier got an item wrong, run gate.py override --out $D --item …

    line 59
  4. Very specific:

    Boundaries: If the boundaries item names a boundary (not "none"), check the implementation decisions against the matching file of the …

    line 66
  5. Very specific:

    Never write prd.md or spec.md while gate.py exits non-zero, except after exit 2 when the person explicitly agrees to continue unverified.

    line 80
  6. Distinctive:

    Every verifier must quote the span of the turn it relied on.

    line 99

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

spec-ready

Planning with an agent usually ends when the agent decides it understands, and agents over-claim. This skill answers one question before a spec is written: has the person actually decided the things a spec depends on? It answers from the person's own turns in the session transcript, which the harness wrote and you did not, checked by a verifier the script calls. You propose; the script decides.

The rules above all others

  1. Evidence comes from the person, not the agent. An item is settled only when it cites one of the person's turns and a verifier the script calls agrees that turn states it, quoting the exact words it relied on. Your summaries, ledgers and playbacks are never evidence.
  2. Never pretend to be verified. With no readable transcript the gate runs in attested mode (exit 2) and says so; the PRD header then reads Readiness: attested.
  3. No invented facts. Every decision line in the PRD cites its turn as [turn:<id>]. A figure with no source is refused. Guesses go under Assumptions, labelled.
  4. No personal data in the skill. Machine settings come from environment variables the user sets.

Steps

Scripts live in this skill's scripts/ directory; run them with uv run or python3, always from the repository root (the gate finds the session transcript and .spec-ready/items.json from there), and give the Bash call a timeout of at least 180000 ms: verifier calls run in parallel but each may take up to a minute. Keep proposals and the gate's files in a scratch directory $D (mktemp -d) until you write.

  1. Propose rows. gate.py turns lists the person's turns with their ids. For each required item, find the turn that states it and write $D/proposals.json: a list of {"item", "turn", "status": "settled" | "guessed", "statement"}. statement is your one-line paraphrase, used as the recommendation if the item is asked. Only risks may be guessed.
  2. Gate. gate.py check --proposals $D/proposals.json --out $D. It finds the transcript itself, calls the verifiers it chooses, writes $D/readiness.json and prints the open questions. Exit 0 ready, 1 items open, 2 no transcript, 4 no verifier could run.
  3. Exit 4, or "no verifier available". On exit 4 the gate printed "verifier unavailable" and why (a broken login, for example): tell the person, since they may want to fix it. Otherwise, for each item dispatch a fresh subagent and paste the output of gate.py brief verbatim as its whole prompt (gate.py brief --item <key> --turn <id>); collect its JSON replies into $D/answers.json keyed by item, and re-run step 2 with --answers $D/answers.json --answers-from subagent. You never answer a brief yourself while a subagent tool exists; only without one, use --answers-from session-model, and the PRD will say the check was self-assessed.
  4. Ask. When items are open, ask them all in one round, least-certain first, in the order the gate printed, each with your recommendation:

    ❓ **Q1** — **<item label>**: <the gate's question>
    
    ➡️ <your recommendation, and the reason in a clause>

    The person's answers are new turns. Update proposals.json to cite them and go to step 2. Run gate.py override only when the person says so: when they tell you a verifier got an item wrong, run gate.py override --out $D --item <key> --turn <id> --quote "<their words>", where the turn is the one in which they said so and the quote is copied from it. The gate checks the quote is in that turn and records the override as a calibration label.

  5. Exit 2 (attested). Tell the person no transcript could be read, so their decisions cannot be checked against their own words, and nothing is settled. Continue only if the person explicitly agrees to continue unverified; then run gate.py accept-attested --out $D --reason "<their words>".
  6. Boundaries. If the boundaries item names a boundary (not "none"), check the implementation decisions against the matching file of the software-design-rules skill when it is installed; otherwise ask the questions in references/boundary-questions.md. Record which rules were applied or knowingly broken in the spec.
  7. Write. Find where this repository keeps specs (docs/specs/, docs/superpowers/specs/, docs/prd/, an existing pattern). If the person has asked for the PRD and spec and there is one such place, or none (then use docs/specs/<date>-<slug>/), write there and say where. Ask only when the location is ambiguous, for example two conventions or a path the person named that does not exist. Copy readiness.json there, then write prd.md from references/prd-template.md (its Readiness: and Verifier: header lines copy mode and verification from readiness.json) and spec.md from references/spec-template.md. Run check_sources.py <dir>/prd.md <dir>/readiness.json --spec <dir>/spec.md and fix every finding. A question nobody has answered goes under Open questions in the PRD, never under Assumptions; a user story or out-of-scope line nobody stated starts Proposed:.

Never write prd.md or spec.md while gate.py exits non-zero, except after exit 2 when the person explicitly agrees to continue unverified.

After writing

The next step is an implementation plan (superpowers:writing-plans). Offer adr-author for a contested decision worth a record. Create tickets only when the person asks, through what they use: the use-linearis skill (linearis), beads (bd create), or gh issue create.

Verifiers

The script picks them, strongest first, and records which ran:

  1. headless claude -p on Haiku, with no tools and no saved session;
  2. a local model through Ollama, alongside 1, when SPEC_READY_DECIDER=ollama:<model> and OLLAMA_HOST are set; a disagreement between them becomes a question;
  3. a fresh subagent given the script-built brief (step 3);
  4. you, as a last resort, labelled session-model.

Every verifier must quote the span of the turn it relied on; a yes whose words are not in the turn counts as unsure.

Optional guard

An opt-in PreToolUse hook, hooks/guard.py, refuses writes to prd.md or spec.md while the readiness.json beside them has not passed. The skill never installs it; README.md shows how.

Credit

The PRD and spec sections adapt Matt Pocock's to-spec, and the question format his grilling (github.com/mattpocock/skills, MIT). It sits between superpowers' brainstorming and writing-plans.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.