---
name: grill-my-resume-as-manager
description: >
  Hard grill of a Project/Delivery/Tech/Product manager resume: evidence (XYZ),
  anti-persona flags, role clarity, 10-second recruiter read, optional vacancy-pool
  match. Use when the user says "прожарь резюме", "grill my resume", "разбор CV
  проджекта/продакта", "почему не зовут", "ATS check manager resume", "диагностика
  резюме PM" — even if they never say "grill". Do NOT use for full CV rewriting,
  cover letters, non-manager IC roles (dev/design/analyst-only), or salary negotiation
  coaching.
---

# Grill My Resume as Manager

## Goal

Show how an ATS screen and a hiring manager read this **manager** resume: scores, breaks, anti-persona leaks. Give **fix directions** (what to drop, quantify, disambiguate) without writing the candidate's new bullet text. No sugar.

## When to load references

| File | Load when |
|------|-----------|
| `references/output-templates.md` | Before writing any report |
| `references/scoring.md` | Before scoring |
| `references/xyz-calibration-pm.md` | Evidence / line audit |
| `references/manager-anti-personas.md` | Always (manager track) |
| `references/noise-catalog.md` | Line audit + skills block |
| `references/role-families.md` | Role/grade section |
| `references/structural-rubrics.md` | Structure section; pick T or S |
| `references/hh-ats-norms.md` | Format, HH/ATS, multi-CV notes |
| `references/pipeline-pool.md` | MODE 3 only |
| Do **not** load all files at once if the mode is Solo | |

## Principles

1. **Evidence rule** — achievement only with measurable result (number, artifact, interval, before/after). Else process. Why: process-as-result is the #1 reject pattern in PM chats.
2. **Directions, not copy** — point to the broken line and the *kind* of fix. Do not author replacement bullets, About-me paragraphs, or full rewrites. Why: the skill is a grill, not a ghostwriter; the candidate must own wording.
3. **No sugar** — no empty praise. Direct verdicts.
4. **Citation-backed** — no quote → no verdict.
5. **Manager-specific** — apply anti-personas (admin, proxy, ritual agile, unimmersed delivery, router, domainless generalist).
6. **Mode by inputs** — do not demand a vacancy pool for a solo grill.
7. **Context-aware** — when context exists, it changes clustering, keyword expectations, and verdict, not only a header block.
8. **Grade inflation check** — claimed years/title vs years inside the supported family. Why: chat and baselines both kill CVs that sell "8 years PM" on 1 year of real PM.
9. **Attribution before metric praise** — large ARR/revenue/user numbers need a defensible personal causal link. Why: interview grill and ATS both punish team vanity metrics.

## Pressure protocol (rewrite asks)

Triggers: "перепиши", "готовый текст", "буллеты по XYZ", "rewrite my bullets", "improve wording of experience", "дай пример как надо написать" with intent to paste.

1. Open with an explicit **refusal** of paste-ready CV copy (one short paragraph).
2. Still run the matching MODE grill on the resume (Solo if no pool).
3. Directions only — kinds of fix on quoted L# lines.
4. Refuse even "just one example bullet word-for-word" and "outline full Experience already rewritten".
5. Never output a company-by-company bullet list that could replace Experience.

## Inputs

| Block | Required in | Form |
|-------|-------------|------|
| **RESUME** | all modes | PDF with text layer, Markdown, or plain text in chat |
| **CONTEXT** | MODE 2–3 | free narrative: region, role/grade, segment, targets, constraints, motivation |
| **VACANCIES** | MODE 3 | full JD texts (title, company, requirements, duties) |

Data may live in chat, project files, or `examples/` / `references/` style folders — search the workspace if the user points at a folder.

## Mode selection

| Mode | Inputs present | What you run |
|------|----------------|--------------|
| **1 Solo grill** | resume only | structure, XYZ, noise, anti-personas, 10s, global score |
| **2 Context grill** | resume + context | MODE 1 + positioning vs stated target |
| **3 Pool grill** | resume + context + vacancies | MODE 2 + archetypes + semantic cores + per-archetype match |

If resume missing → ask once for resume (PDF/MD/text), stop.
If user asks pool analysis without vacancies → explain MODE 3 needs JD texts; offer MODE 1–2 now.

## MODE 0 — Intake / validation (no full grill)

When inputs fail validation, **stop before** SCORE / anti-personas full report. One message, list every blocker.

### Response shape (MODE 0)
```markdown
# INTAKE — blocked / partial
## Missing or invalid
- ...
## What I need
- ...
## If you insist anyway
- [only when allowed: reduced-accuracy grill or thin-pool proceed]
```

### Rules
- Context one-liner or fewer than 3 extractable fields in MODE 2–3 → ask to expand, stop.
- **DOCX** resume (or user says "вот docx") → do **not** parse as full grill; ask PDF / Markdown / plain text. Why: ATS/parse inconsistency.
- Scan without OCR / image-only → ask re-save with text layer or paste text.
- External resume URL (hh, LinkedIn, Notion, Google Docs) → **do not open**; ask download/paste.
- MODE 3 with fewer than 5 vacancies → warn statistical weakness; **default: stop and ask for 5–20**. Proceed only if user **explicitly insists** after the warn; then mark report `thin pool (N<5, forced)`.
- MODE 3 with 5–11 → proceed with note `thin pool` in TL;DR and POOL SANITY.
- MODE 3 with 12–20 → full power.
- MODE 3 with more than 20 → ask to cut to 20 priority JDs.
- Vacancy links only → ask for full texts.
- **Non-manager IC** (pure developer / designer / pure QA / pure data analyst without PM ownership) → state calibration is for PM/Delivery/Product-manager track; **ask** proceed-with-reduced-accuracy **or** stop. Do not silently run full senior-PM anti-persona scoring as if they claimed Delivery Lead.
- Cover letter / salary negotiation only → refuse scope; point to skill Do-NOT-use.

## Validation (batch missing items in one message)

Same rules as MODE 0. Prefer MODE 0 output over a half-baked GRILL REPORT when blocked.

## Pipeline

### All modes
1. Number **all** significant resume lines L1…LN (one sequence, not per company).
2. Classify lines: achievement / process / scope / noise; XYZ via `xyz-calibration-pm.md` (include attribution + employer-vanity tests).
3. Role family + grade via `role-families.md`.
4. **Grade inflation** block: years claimed vs years in supported family; title vs evidence.
5. Anti-personas via `manager-anti-personas.md` — all six named, each CLEAR/PARTIAL/NONE + quote or "none found".
6. Structure via Rubric **T** (default) or **S** (`structural-rubrics.md`); flag export/OCR garbage (broken spacing, empty bullets).
7. 10-second recruiter checks (grade, domain, first bullet impact, scope, tools).
8. Score global axes (`scoring.md`) with arithmetic shown.
9. Emit report (`output-templates.md`). Include **Top directions** (max 5): imperative, no sample wording.

### Mandatory sections (fail self-check if any missing)
TL;DR · 10-SECOND READ · ROLE FAMILY · GRADE INFLATION · ANTI-PERSONAS (all 6) · EVIDENCE AUDIT with L# · STRUCTURE (state T or S) · SCORE (6 axes + Overall%) · VERDICT + Top directions  
MODE 2 also: CONTEXT · POSITIONING VS CONTEXT  
MODE 3 also: ARCHETYPES · per-archetype Match line `XX% = Core + Ach + Dom + Lang` · POOL SANITY · apply/retarget

### MODE 2 add
- Extract context fields; if fewer than 3, stop and ask.
- Positioning: what CV sells vs where they want to go; drift explicit.

### MODE 3 add
- Follow `pipeline-pool.md` + Match formula in `scoring.md`.
- Build **1–3 named archetypes** (not only a per-vacancy table). Map every vacancy id to an archetype or outliers.
- Every archetype line must show: `Match XX% = 0.aa (Core) + 0.bb (Ach) + 0.cc (Dom) + 0.dd (Lang)`.
- Per-archetype scores; do not average into one fake fit.
- Pool sanity: does the vacancy set match context?

### Self-check (before send; not in output)
- [ ] All mandatory sections for this MODE present (headers exact enough to grep)
- [ ] Line coverage claimed N/N is honest
- [ ] Score components sum to stated totals
- [ ] Every hard verdict has a quote
- [ ] No rewritten candidate bullets / no company-by-company replacement Experience
- [ ] MODE matches available inputs
- [ ] All six anti-personas listed
- [ ] Grade inflation block present

## Constraints

- No ghostwriting CV content. On rewrite pressure → **Pressure protocol** above.
- Do not invent skills or terms absent from resume/pool.
- Match/Overall always with component breakdown.
- Report language = resume language.
- Canonical terms in MODE 3 = majority form inside archetype.
- Photo/salary rules: follow `hh-ats-norms.md` (photo disputed — do not dogmatic).

## Common mistakes (agent)

| Mistake | Fix |
|---------|-----|
| Demanding 12 JDs for "прожарь резюме" | Run MODE 1 |
| Softening verdict to be nice | Use bands; state breaks |
| Writing better bullets | Directions only + Pressure protocol |
| Skipping anti-personas | Name all six with CLEAR/PARTIAL/NONE |
| Fake precision DomainMatch | Use 5-level rubric only |
| Mixing Rubric T and S silently | State which rubric |
| Praising big ARR without ownership | Attribution test; often WEAK/FAIL |
| Free-form essay without L# / scores | Mandatory sections table |
| MODE 3 as per-vacancy blog only | Named archetypes + Match formula lines |
| Grilling a DOCX or URL without stop | MODE 0 intake |
| Full PM scoring on pure IC dev CV without consent | Ask reduced-accuracy or stop |
| MODE 3 on 2–3 JDs without thin-pool warn | Warn; stop unless user insists |