---
name: resume-system
description: >
  Mandatory entrypoint for every resume or CV request in Darsh Shah's vault,
  including quick wording questions, single-bullet feedback, coaching,
  critique, JD tailoring, content selection, formatting, export, and QA.
  Invoke on every resume-related turn even when the user does not name the
  skill or asks only a short conversational follow-up.
---

# Resume System

Use this skill for all resume work in `jobs-resume-studio/`. Keep this file as
a thin router; load only the references required for the active phase.

## Activation contract

- Activate before analyzing or suggesting any resume content.
- Keep the active workflow phase across follow-up turns; a short question does
  not become standalone copyediting.
- For bullet work, load `bullet-coaching.md`; when facts, impact, causality, or
  team composition remain unclear, also load `impact-excavation.md`.
- Do not draft a final bullet until the evidence is sufficient. Ask one
  hypothesis-led recall question, wait for the answer, and then continue.

## Default: interactive tailoring

When Darsh provides a JD and asks for a resume, do not jump directly to a final
document. Use five explicit phases:

1. **Strategy checkpoint:** Interpret the business problem, company context,
   talent bar, candidate story, proof pillars, screening risks, and tone. Show
   the readout and wait for correction before selecting the final content.
2. **Evidence checkpoint:** Search the approved vault and broader library.
   Propose experience order, bullet budget, evidence shortlist, de-emphasis
   choices, and excavation gaps. Wait before assembling.
3. **Bullet checkpoint:** Coach one bullet at a time. Use hypothesis-led,
   industry-recognizable memory prompts when evidence is missing. Lock and
   persist only wording Darsh explicitly approves.
4. **Assembly checkpoint:** Build the Markdown resume from locked content.
   Show the complete story and unresolved tradeoffs before rendering.
5. **Finalization checkpoint:** After explicit approval, render DOCX/PDF, run
   layout and visual QA, update the completion log, rebuild the broad index,
   and mark the source state `rendered-final`.

Use one-shot mode only when Darsh explicitly asks to skip checkpoints. A
one-shot output remains a draft until he approves its content.

## Reference routing

- JD intake and interactive gates:
  `references/tailoring-workflow.md`
- Talent thesis, company context, story, and evidence hierarchy:
  `references/talent-strategy-and-narrative.md`
- Bullet diagnosis, rewriting, ordering, and restraint:
  `references/bullet-coaching.md`
- Missing-impact excavation, business hypotheses, and industry recognition:
  `references/impact-excavation.md`
- Approved vault, locking, state transitions, and index semantics:
  `references/persistence-and-state.md`
- Repository paths and reusable sources:
  `references/source-map.md`
- Kellogg format, rendering, page fit, visual QA, and completion logging:
  `references/kellogg-format-export-qa.md`

Read `tailoring-workflow.md`, `talent-strategy-and-narrative.md`, and
`source-map.md` for every JD-tailoring request. Add other references only when
their phase begins.

## Core guardrails

- Strategy precedes retrieval; retrieval precedes rewriting.
- Preserve Darsh's approved wording and July 22 Cursor-derived one-page format
  unless the target creates a clear advantage for changing it.
- Default to bold but defensible positioning. Never invent hard facts.
- Make expertise visible through judgment, scale, and consequence rather than
  buzzwords or implementation trivia.
- Use official titles plus truthful functional descriptors; flag material
  reframes for approval.
- Translate Epic work through developer-platform, enterprise, licensing, or
  creator-tool context when that is the underlying work.
- Keep most bullets within two rendered lines.
- Do not edit a locked bullet silently. Reopen it, explain why, and obtain a
  new approval before replacing its canonical bank entry.

## Repository tooling

Use the repo-local `.venv` for Python. Use:

```bash
.venv/bin/python skills/resume-system/scripts/resume_bank.py --help
```

Treat `resume_vault.json` as the approved achievement bank and
`resume_library_db.json` as a broad, generated search corpus. They are not
interchangeable.