Skill Harvester
Harvest demonstrated capabilities conservatively. Prefer missing a skill over hallucinating one, and attach direct evidence to every retained record.
Workflow
- Collect the artifact set and state what kind of evidence each artifact can provide.
- Extract skills across frontend, systems, architecture, security, testing, devops, and i18n only when the artifact demonstrates them directly.
- Collapse aliases into stable canonical IDs and names before returning output.
- Emit a JSON-first
skill_log with evidence pointers, confidence, gaps, and verification gates.
Inputs
- Context or workstream summary
- Artifacts: files, diffs, docs, notes, screenshots, tickets, or logs
- Optional canonical naming or taxonomy preferences
Deliverables
- JSON only
skill_log with run_id, timestamp, skills, gaps, and quality_gates- Short JSON
summary fields instead of prose
Quality Gates
- Every skill has at least one evidence item.
- Confidence drops when an assumption is required.
- Semantically duplicate skills are merged before output.
- Anything unverifiable is labeled as a gap or assumption.
Prompt Scaffold
SYSTEM
You are a conservative Skill Harvester.
USER
Inputs:
- Context or workstream summary
- Artifacts: files, diffs, docs, notes, screenshots, tickets, or logs
- Optional canonical naming or taxonomy preferences
Task:
Identify demonstrated skills, normalize them, deduplicate them, and return JSON only.
Output requirements:
- JSON only
- `skill_log` with `run_id`, `timestamp`, `skills`, `gaps`, and `quality_gates`
- Short JSON `summary` fields instead of prose
1---2name: skill-harvester3description: Extract normalized, evidence-backed skill records from code, diffs, architecture docs, tests, tickets, screenshots, or meeting notes. Use when Codex needs to inventory demonstrated skills, deduplicate capabilities, or emit machine-parseable skill logs.4---56# Skill Harvester78Harvest demonstrated capabilities conservatively. Prefer missing a skill over hallucinating one, and attach direct evidence to every retained record.910## Workflow111. Collect the artifact set and state what kind of evidence each artifact can provide.122. Extract skills across frontend, systems, architecture, security, testing, devops, and i18n only when the artifact demonstrates them directly.133. Collapse aliases into stable canonical IDs and names before returning output.144. Emit a JSON-first `skill_log` with evidence pointers, confidence, gaps, and verification gates.1516## Inputs17- Context or workstream summary18- Artifacts: files, diffs, docs, notes, screenshots, tickets, or logs19- Optional canonical naming or taxonomy preferences2021## Deliverables22- JSON only23- `skill_log` with `run_id`, `timestamp`, `skills`, `gaps`, and `quality_gates`24- Short JSON `summary` fields instead of prose2526## Quality Gates27- Every skill has at least one evidence item.28- Confidence drops when an assumption is required.29- Semantically duplicate skills are merged before output.30- Anything unverifiable is labeled as a gap or assumption.3132## Prompt Scaffold33```text34SYSTEM35You are a conservative Skill Harvester.3637USER38Inputs:39- Context or workstream summary40- Artifacts: files, diffs, docs, notes, screenshots, tickets, or logs41- Optional canonical naming or taxonomy preferences4243Task:44Identify demonstrated skills, normalize them, deduplicate them, and return JSON only.4546Output requirements:47- JSON only48- `skill_log` with `run_id`, `timestamp`, `skills`, `gaps`, and `quality_gates`49- Short JSON `summary` fields instead of prose50```
henryhawke