Profile — raw material & positioning

Everything the candidate brings and their core public positioning live here; every other skill consumes them. Backward looks — progress reviews, the retrospective — belong to the coach.

The goal, and what must be true when it is met

Turn whatever the candidate brings into the files every other skill reads — every fact traceable, positioning unified, nothing stated more strongly than what happened.

Anything not yet true is written into its file as TODO: and resurfaces when it blocks what they ask for next — never on a schedule, never as a nag.

Prerequisites

• Required: A settled folder. Nothing is written before it.
• Optional: Documents in documents/ (read on sight); voice.md samples (calibrates voice); storybank.md (confirmed stories for pitch proof points).

Loops and sequences

Setup — the workspace (a sequence)

Runs when there is no CLAUDE.md in the working folder. Exits when the folder is settled, CLAUDE.md is written, and both drop paths (documents/, jd-inbox/) exist.

1. Settle the path: If the working folder is clean, name it plainly in chat: "everything I write lands in <path> — say the word if you'd rather use a different folder." If the folder is occupied or generic, propose ~/job-search/.
2. Never hunt filesystem: Ask once if a previous search folder exists; never search unauthorized directories.
3. Write guardrails first: Copy templates/workspace-CLAUDE.md to CLAUDE.md before any candidate file is created.

Intake — the raw material (a sequence)

Runs when a new candidate arrives, a résumé is dropped, or they ask to get set up. Exits when destination table items are written or TODO:'d, and the close summary is delivered.

• Process: Check documents/ first. Ask ONE question at a time. Decline raw LinkedIn URLs as seeding sources (request PDF export instead).
• Extraction gate: Facts told directly in conversation are written to disk; facts extracted from documents pass through § Nothing extracted is written.
• Close: Deliver a concise summary (target, band, timeline, top asset, top risk, first step). Offer the practice cold-start ("say 'drill me' for your first 15-minute score").

Base résumé — improvement rounds (the loop)

Runs when they ask to improve the base résumé with no specific job in mind. ("Tailor for X" routes to apply.) Read the base resolution ladder (references/schema.md), the audit (references/eval.md), reader craft (references/patterns.md), and the Summary's shape (../apply/references/patterns.md § The Summary).

• Standard: base-resume-brief.md § FIXED (sourced from criteria.md targets, top competencies across postings, and base-resume.md § Claim rules).
• Each round: Re-read brief in full. Apply changes reshape-only to base-resume.md. Budget: 2–3 rounds, said up front.
• Self-loop: Run ../apply/scripts/check_materials.mjs for structure and the independent wording check (references/language-check.md) for phrasing (max 2 passes).
• Score & ceiling: Score against § FIXED as N/M held count over FIXED bullets plus reader tiers. Append row to ## Rounds in base-resume.md (or base-resume-history.md). Read earlier rows first.
• Exits: Draft scores M/M; budget expires; or the ceiling: two rounds in a row with the same count and tiers → stop and present tradeoff as a decision for them to make. Never relax the standard — cutting a claim's supporting evidence to fit a limit IS relaxing it: a claim-name without its number is not the claim.

Pitch — public positioning (the loop)

Runs when they ask to build or review their pitch, or another skill routes with a diagnostic (coach on silent funnel, outreach on response rate <10%).

• Standard: External and written down before drafting — pitch-brief.md § FIXED (references/schema.md).
• Budget: 2 rounds, said up front. Read earlier rows in ## Rounds in pitch.md (or pitch-history.md) first.
• Each round: Re-read brief in full. Apply minimum change to move diagnostic score (references/eval.md). Run substitution test on core statement.
• Score & write: Score against § FIXED as N/M held count plus 1–5 diagnostic. Append row to ## Rounds in pitch.md (or pitch-history.md).
• Exits: Scores M/M; budget expires; or the ceiling: two rounds in a row with the same count and diagnostic → stop and present tradeoff as a decision for them to make. Never relax the standard — cutting a claim's supporting evidence to fit a limit IS relaxing it: a claim-name without its number is not the claim.

LinkedIn (a sequence)

Runs when they ask for a LinkedIn review or audit.

1. Verify profile URL: Match URL and profile against profile.md / base-resume.md.
2. Audit top-down: Review Headline, About, Experience against search impact and positioning thesis. Produce concrete rewrites.
3. Write linkedin-audit.md: Sync headline to profile.md § Snapshot and note consistency updates.

Exits when linkedin-audit.md is written and downstream syncing obligations completed.

Consistency sweep (a sequence)

Runs when asked "do my materials tell one story?" or after modifying any positioning surface.

• Cross-read base résumé summary, LinkedIn About/Headline, and pitch. Propose base résumé summary adjustments in chat before writing.

Exits when all surface mismatches have concrete rewrites.

Direction change (a sequence)

Runs when they ask what they are looking for, change targets, or retarget. Exits when criteria.md is current and downstream staleness is flagged.

• Update targets in criteria.md (archiving old targets as history lines).
• Flag downstream staleness in pitch.md, evaluate verdicts, and criteria.md constraints. Never restart from scratch.

State

File shapes in references/schema.md. Enforced by scripts/check_files.mjs.

• Owned files: profile.md, criteria.md, base-resume.md, voice.md, pitch.md, pitch-brief.md, pitch-history.md, linkedin-audit.md.
• base-resume.md rules:
  • Trace rule: Every number, title, date, scope claim on any surface traces here. A surface may be vaguer, never stronger.
  • Ruling gate: Concessions ("I can't defend that"), declined proposals, and approved wordings are written here immediately.
  • Direction rule: When a correction is agreed, update the base first in the same turn, then re-render dependent surfaces.
• The write moment: Every pitch.md rewrite triggers four actions in the same turn: (1) update pitch.md, (2) sync profile.md § Snapshot headline, (3) propose any novel fact for base-resume.md, (4) append row to ## Rounds in pitch.md (or pitch-history.md).
• Hands back: When the ask is answered and records are written to disk.
• Session close: Run node scripts/check_files.mjs --workspace .. The wording check runs inside the base-résumé loop; no other profile surface is candidate-voiced, so the close spawns no checker-subagent. Report outcomes, never narration (clean is 1 line; fix FAILs before reply ends).

Nothing extracted is written until proposed, checked, and confirmed

Material extracted from documents or past sessions passes through the proposal table:

• Batch confirm: Document-sourced rows are confirmed in one batch with opt-out.
• Explicit yes: Conversation-sourced claims and novel inferences require their own explicit confirmation.
• Declined proposals: Record as a ruling in base-resume.md § Claim rules ("declined at intake: <value> from <source>, <date>") so it is never re-proposed.

Guardrails

• Targets and core statements are written in candidate's words. Pitch must survive the substitution test.
• Respect voice.md and avoid artificial analogies or inflated scope words.

Every reply ends with ONE contextual next step — a sentence with its why, not a menu.