Skill Workshop

Build the smallest skill that owns the behavior, then prove its declared claims. Compose the repo's existing authoring, arena, review, and install tools; do not create another controller around them.

Non-negotiable design rules

• Write or update SPEC.md before SKILL.md or executable behavior.
• Treat a wall as evidence that the design is wrong. Re-derive the design; do

not hide it behind a flag, shim, special case, fallback, or parallel path.

• Keep dry-run, static, smoke, behavioral, and representative-live evidence

separate. One never stands in for another.

• Keep every attempted adapter call, failure class, and recovery action visible;

omitting a failed integration attempt is itself a failed gate.

• A final review needs at least one model family other than the author. Missing

capacity blocks completion; it does not authorize substitution.

• Record requested/effective effort because latency and reasoning depth trade

directly against each other.

• Never persist credentials. Device config is ignored and non-secret.
• Do not add hashes, fingerprints, or timestamps as self-attestation. Actual

command output and independent review own authenticity.

1. Discover the behavior

Capture the real trigger, non-trigger, user, repository, authority, and the failure or successful workflow being generalized. Describe the desired behavior without preserving accidental details of the first solution.

Redirect a narrow request to the skill named in the description instead of running the full workshop.

2. Select the fixed risk tier

Choose the highest matching tier before build. Tiers are cumulative.

| Tier | Select when | Required evidence | |---|---|---| | method | Instructions/reference only; no executable or live dependency. | Spec, real eval, two-family blind test, lint, four-agent install, fast councils. | | scripted | Bundled code, validator, schema, or generated artifact. | Method plus public-seam TDD, repo tests, deep final council, Thermos. | | integration | Auth, third-party CLI/API, provider/model identity, capacity, or device setup. | Method plus deep final council, Thermos, device config contract, smoke and representative qualification; add TDD/tests when code is bundled. |

For integration work, read references/integration-tier.md before specifying the contract.

Effort defaults to standard for method and deep for scripted/integration. Use deep for a security-sensitive or architectural method skill.

3. Specify and gate

Before promising evidence, exercise each required external CLI end to end: run its help/version command and its smallest non-mutating validate, identity, or smoke operation. A name on PATH is not readiness.

Create one source SPEC.md containing:

• outcome, triggers, non-triggers, fixed tier, and effort;
• public seams and exact observable behavior;
• required artifacts and receipts;
• behavioral cases, blind forward-test fixtures, and binary completion rubric;
• resolved decisions and deviations.

Gate that exact file with the repository's plan-review provider and a fast cross-family adversarial review. Apply valid findings to the source spec. Record rejected findings with the design boundary they would violate; do not add machinery just to satisfy a reviewer.

4. Establish the baseline

Before implementing the skill:

1. Create dirty cases for every invariant and clean guards for valid boundary

combinations.

1. Run the arena's dry mode only to prove wiring.
2. Run at least one real model as the no-skill baseline and retain its receipt.

Do not tune cases after seeing baseline answers. If semantic review finds a mislabeled case, version the suite, record the defect, and establish a new real baseline before comparison; never reuse the old score. If the baseline is perfect, report no measured lift; the suite remains a regression guard.

5. Build the smallest artifact

Use skill-creator for the canonical scaffold and writing-great-skills for the instruction surface. Keep SKILL.md lean and move only integration-specific detail into the reference.

When code is bundled, use TDD at the public seam:

1. Write one failing subprocess or external-interface test.
2. Run it and retain the red result.
3. Implement only enough behavior to pass.
4. Add boundary cases, then run the full relevant suite.

Do not manufacture scripts or tests for a code-free method/integration skill.

6. Evaluate the same cases

Add a with-skill arena variant and rerun the same cases and model used for the baseline. Preserve errors, regressions, and missing cells. A dry run is never behavioral evidence.

7. Forward-test blindly

Run two fresh isolated agents from distinct model families. Give each only:

• the skill path;
• one raw realistic task.

Do not provide expected answers, arena cases, another agent's output, or the workshop's implementation discussion. Capture commands and receipts, then remove temporary fixtures. One family may equal the author family; the two forward-test families must differ.

Install and qualify the skill through each repository target: Claude Code, Codex, Cursor, and Cline. A target that cannot discover or execute the skill is visible incomplete evidence, not permission to relabel another process.

8. Review and receipt

Run tier-required tests, static lint, install checks, final council, and Thermos. Review the semantic contents of external receipts; the helper checks their structure and existence but is not a second evaluator.

Initialize the receipt from the skill directory:

bash
python3 scripts/workshop.py init \
  --tier integration \
  --bundles-code true \
  --work-unit f84bad5c-2488-41f4-ae0b-75003a5f0b5b \
  --author-family openai \
  --output WORKSHOP.json

Each step leaves a readable receipt, not an opaque blob: copy templates/receipt.md — the real command, its verbatim output, and a one-line verdict. Fill WORKSHOP.json only from those retained receipts, then check it:

bash
python3 scripts/workshop.py check WORKSHOP.json

Render the reviewer-facing report so semantic review is one read instead of fifteen file-opens:

bash
python3 scripts/workshop.py report WORKSHOP.json --format md   -o REPORT.md
python3 scripts/workshop.py report WORKSHOP.json --format html -o REPORT.html

report inlines every receipt's contents next to the checker verdict. The checker proves the receipts exist, are non-empty, and are distinct; reading the report is what confirms they are true. A green checker never ships alone.

Exit 0 means the declared structural gates are complete. Exit 1 means valid but incomplete evidence. Exit 2 means the receipt violates the contract. Never edit an exit code, family, model identity, status, or substituted field to make the checker pass.

A complete exit is necessary but not sufficient: the checker validates receipt structure, distinctness, and existence, not the truth of what a receipt records. Only semantic review of the receipt contents plus the independent council make the completion claim real. Never treat a green checker as a standalone ship signal.

Ship only when the checker reports complete, semantic receipt review is clean, deviations are explicit, and the requested commit/PR authority exists. Ask before merge, publication, or release unless higher-level instructions already grant that exact action.

Failure handling

| Issue/Error | Fix | |---|---| | Required CLI is absent or its identity check fails | Record the gate incomplete and provide the exact install/setup action; do not replace the tool. | | Auth, model, capacity, timeout, or adapter qualification fails | Preserve the typed failure and recovery from the integration reference; retry the same declared seat only after recovery. | | Plan review times out or is dismissed | Reopen the exact unchanged artifact and obtain a recorded decision; timeout is not approval. | | Arena dry-run passes but a live cell errors or regresses | Keep the error visible, inspect the raw receipt, and fix the skill or version a mislabeled suite before establishing a new baseline. | | workshop.py check exits 1 or 2 | Read the sorted diagnostics, repair the missing evidence or invalid contract at its source, and rerun the producing command. | | Independent model family is unavailable | Leave final review blocked and report recovery; never substitute another process from the author family. |