skill-builder

Skill Builder

Repair one failing skill gate at a time. Preserve the artifact, command, baseline, failed contract, patch, and rerun evidence for every repair.

When To Use

Use for SKILL.md repair, skill eval hardening, release proof, Tessl score improvement, reference quality fixes, and plugin-skill readiness work.

Inputs

• Canonical skill source path, failing gate or score, allowed edit scope, and required validation command.
• Baseline package, audit, eval, Tessl, or review receipt when a score or gate is being improved.
• Factory gate answers from references/repair-policy.md before create, harden, refactor, skillify, plugin, hook, or package-design work.
• Redact secrets and sensitive data by default in prompts, outputs, temporary evidence, copied artifacts, and review notes.
• For unsafe requests, return Safety Verdict: safety constraints intact; refusing unsafe request.
• Classify the request as audit, repair, validation-only, or promotion before requesting inputs. Audit requires a target or bounded collection, not a failing score or edit authority. Ask only for information needed by the selected mode. Confirm canonical ownership before edits.

Outputs

• Exact validation commands and pass, fail, or blocked outcomes.
• Changed files, rollback path, and evidence artifact locations.
• For audit, report findings, source references, coverage gaps, and proposed edits. Changed files and rollback evidence are not applicable without edits.
• For blocked release-eval cases, the failed repair-policy item in blocker_notes.
• Lane-separated readiness summary covering SDK, OSS, Tessl, runtime, registry, and publication truth only when those lanes have current evidence.
• Schema-bound outputs include schema_version and preserve the validation command that produced the result.

Workflow

1. Identify the requested mode, target, canonical owner, and authority.
2. For audit, inspect the source and relevant references before returning evidence-backed findings and coverage gaps without edits. For validation-only work, run only authorized checks and report their results. Stop after the requested deliverable is satisfied, not merely after selecting its mode.
3. For repair, read references/repair-policy.md. Load the full promotion ladder only when promotion or SDK/Tessl handoff is explicitly selected. Handoff and rollback requests authorize only their named actions, not a fresh repair.
4. Run the focused gate; record baseline score, artifact path, and first blocker.
5. Apply one Repair Map change from references/repair-policy.md, then rerun the same gate.
6. If score or blocker is flat, undo or narrow and try the next map item. After three flat loops, stop with blocker_notes.
7. On green, run final gates. If any fail, name the gate and next patch target; do not claim release readiness.