paqad-ai
Version:
Spec-driven development framework — AI agents that think before they type
22 lines (15 loc) • 1.65 kB
Markdown
verified: 2026-10-31
1. Run the candidate task three times without any skill, each in a fresh context (a fresh subagent where the host has them).
2. Save each transcript as `<workspace>/baseline/run-<n>.md` for n from 1 to 3, where the workspace is `.paqad/tmp/authoring/skills/<name>`.
3. Judge each run against the task and write `<workspace>/baseline/verdicts.json` as `[{run, passed, note}]`.
4. Run `paqad-ai skill justify <name>`.
A skill is justified when at least one run failed. Three passes out of three stop the pipeline: no skill is needed.
Four cases have a better home than a skill. Pass the matching flag and the pipeline stops without a skill folder:
| Case | Flag | Better home |
| ---------------------------------- | --------------------------------------------------------- | ----------------------------------- |
| a one-off fix | `--exclusion one-off` | fix it in this change |
| already written down | `--exclusion documented --documented-in <path>` | point the agent at that file |
| a project convention | `--exclusion convention` | the always-loaded instructions file |
| a rule that has to hold every time | `--exclusion enforceable --enforce-with <hook or script>` | a hook or a rules-as-scripts check |
For an installed skill the entry is `paqad-ai skill improve <name>` instead: it snapshots the folder first and the snapshot is the baseline.