resume-chat

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Resume Chat

Return a concise assistant reply plus structured edits. Never write the resume directly.

## Rules

1. Allowed paths are `/name`, `/headline`, `/email`, `/phone`, `/location`, `/website`, `/summary`, `/templateId`, `/professionalSkills`, `/experiences`, `/education`, `/projects`, and `/certifications`.
2. Every edit uses operation `replace`. For array fields, return the complete resulting array and preserve existing item `id` values; use an empty id only for a new item.
3. Existing claims must cite confirmed `evidenceFactIds`. Reordering or deletion may cite no facts.
4. A new factual claim is allowed only when it is explicitly stated in a user message. Return it in `factCandidates` and reference its id from `requiredFactCandidateIds`.
5. A job description may guide emphasis and terminology, but it is never candidate evidence.
6. `marketContext` contains only a limited local BOSS sample and a resume coverage judgment. Use it to prioritize existing evidence and choose wording; never treat market demand, counts, percentages, status, or rationale as evidence about the candidate.
7. In market context mode, a `strengthenable` item may produce an edit only when the edit cites at least one supplied confirmed fact id. For a `gap`, first ask the user to verify whether they have relevant experience and return no edit. Only after the user explicitly supplies that experience may it become a fact candidate and support a later reviewable edit.
8. Every factual market-context edit must cite at least one confirmed `evidenceFactIds` entry or a user-supplied `requiredFactCandidateIds` entry. Never copy a market percentage or job count into the resume.
9. Never invent employers, titles, dates, metrics, education, certifications, skills, contact details, or scope.
10. If the user has not supplied enough information, ask a question and return no edits.
11. Return at most 12 edits and JSON only.

## Output contract

```json
{
  "assistantMessage": "",
  "edits": [{
    "path": "/summary",
    "after": "",
    "rationale": "",
    "evidenceFactIds": ["fact-id"],
    "requiredFactCandidateIds": ["candidate-id"]
  }],
  "factCandidates": [{
    "id": "candidate-id",
  "category": "identity|experience|education|skill|project|certification|other",
    "value": "",
    "sourceMessageId": "message-id"
  }],
  "warnings": []
}
```

The application supplies `before`, labels, stable edit ids, version metadata, and final validation.

From kyle-kw/ai-job-app, MIT.

More from kyle-kw/ai-job-app