kyberforge: factory-audit check suite for prompt files #152

Open
opened 2026-10-01 06:23:53 +00:00 by Claude · 0 comments
Collaborator

Part of #94.

Depends on #147
Depends on #149

Why

factory-audit audits skills and agents only. A hand-edited *.prompt.md gets no audit, and Claude compile keeps only description, allowed-tools, model, argument-hint and arguments, dropping anything else without a warning.

What to build

Register prompts as a primitive in the generic flow from #147:

  • A scripts/lib-checks-prompt.sh check suite, sourced from a new prompt mode in scripts/validate.sh, encoding the rules prompt-author (#149) already decided are correct. Derive the rubric from that skill's references/ rather than designing it twice.
  • A criteria file and any rubrics the flow needs, under references/.
  • A row in the primitive data table (target shape *.prompt.md, coverage line, "Run prompt-author to address findings").
  • Provenance handling: extend validate-provenance.sh only if prompt files carry source_keys; otherwise say so in the PR.
  • tests/validate-prompt.bats.

Not in scope

  • Any change to the skill or agent flows or check suites.

Acceptance criteria

  • factory-audit on a valid prompt file returns PASS; on one with a field outside the five surviving ones, the finding names the field, the target and the fix.
  • tests/validate-prompt.bats passes, with the existing skill and agent bats suites still green.
  • Step 0 recognises *.prompt.md and the unrecognised-path message lists the new shape.
  • factory-audit metadata.version bumped (ADR-0022); description and body pass ADR-0020 budgets.
  • apm audit --ci clean.

References

  • plugins/kyberforge/docs/research/docs/microsoft-apm/prompt-primitive-schema.md
  • ADR-0020, ADR-0022, ADR-0025
Part of #94. Depends on #147 Depends on #149 ## Why `factory-audit` audits skills and agents only. A hand-edited `*.prompt.md` gets no audit, and Claude compile keeps only `description`, `allowed-tools`, `model`, `argument-hint` and `arguments`, dropping anything else without a warning. ## What to build Register prompts as a primitive in the generic flow from #147: - A `scripts/lib-checks-prompt.sh` check suite, sourced from a new `prompt` mode in `scripts/validate.sh`, encoding the rules `prompt-author` (#149) already decided are correct. Derive the rubric from that skill's `references/` rather than designing it twice. - A criteria file and any rubrics the flow needs, under `references/`. - A row in the primitive data table (target shape `*.prompt.md`, coverage line, "Run prompt-author to address findings"). - Provenance handling: extend `validate-provenance.sh` only if prompt files carry `source_keys`; otherwise say so in the PR. - `tests/validate-prompt.bats`. ## Not in scope - Any change to the skill or agent flows or check suites. ## Acceptance criteria - [ ] `factory-audit` on a valid prompt file returns PASS; on one with a field outside the five surviving ones, the finding names the field, the target and the fix. - [ ] `tests/validate-prompt.bats` passes, with the existing skill and agent bats suites still green. - [ ] Step 0 recognises `*.prompt.md` and the unrecognised-path message lists the new shape. - [ ] `factory-audit` `metadata.version` bumped (ADR-0022); description and body pass ADR-0020 budgets. - [ ] `apm audit --ci` clean. ## References - `plugins/kyberforge/docs/research/docs/microsoft-apm/prompt-primitive-schema.md` - ADR-0020, ADR-0022, ADR-0025
Claude added this to the Kyberforge basics milestone 2026-10-01 06:23:53 +00:00
Claude added the Kind/Feature label 2026-10-01 06:23:53 +00:00
Sign in to join this conversation.