ResumeProof

Build job-specific resumes from verified evidence while keeping claims truthful, readable, and easy to audit. Treat ResumeProof as the governance layer around the host agent's document-generation tools, not as a bundled resume-template engine.

Non-negotiable gate

Never generate or overwrite HTML, DOCX, or PDF output before the user explicitly approves the complete text draft. Approval of an outline, one section, or a previous version does not count.

Workflow

1. Discover the user's sources: base resume, target JD, evidence files, preferred language, template, and output location.
2. Read references/configuration.md when setting up a new workspace or resolving file locations.
3. Read references/workflow.md before tailoring or generating a resume.
4. Create the application workspace with python scripts/resume_proof.py new. Record supported claims in evidence-ledger.json.
5. Compare the resume with the JD. Prioritize role-critical capabilities, measurable outcomes, and relevant keywords without copying the JD mechanically.
6. Draft the full resume in plain text. Keep full-time roles and internships separate when mixing them could imply instability.
7. Present the complete text draft and wait for explicit approval.
8. Only after approval, record the approved-text hash with python scripts/resume_proof.py approve, then use available document tools to generate the requested files without changing the approved meaning.
9. Run factual, language, pagination, extraction, and visual QA. Use the CLI's validate and qa commands to record results.
10. Deliver with the CLI's deliver command. It must block delivery when approval is missing, approved text has changed, QA is incomplete, or a destination collision lacks explicit --force authorization.

Writing rules

• Keep every claim traceable to a supplied source or explicit user confirmation.
• Do not invent metrics, tools, ownership, dates, employers, clients, markets, or outcomes.
• Prefer specific actions and results over generic capability claims.
• State partial ownership accurately with verbs such as supported, coordinated, analyzed, or contributed.
• Use natural language appropriate to the resume's language. If a humanizer skill is available, apply it only after factual content is settled.
• Keep the final resume within the user's requested length; default to no more than two A4 pages.
• Optimize for human scanning first and ATS extraction second. Avoid dense columns, decorative charts, and text embedded in images.

Output contract

For each application, retain:

• text-review.md: the exact draft approved by the user
• editable source, usually HTML or DOCX
• final PDF
• evidence-ledger.json: claim-to-source mappings
• manifest.json based on assets/manifest-template.json
• validation notes or command output

Use scripts/resume_proof.py as the single cross-platform entry point. Do not bypass its approval and delivery checks with direct copies unless the user explicitly requests a manual recovery.

QA checklist

• All dates, names, metrics, and tools match approved evidence.
• No placeholders, comments, hidden text, or stale content remain.
• Contact details and links are correct.
• The PDF text can be extracted in a sensible reading order.
• Required keywords appear naturally; forbidden or legacy terms are absent.
• Page count meets the requested limit.
• Page breaks, spacing, bullets, and alignment are visually clean.
• Delivered files match the approved files by checksum.