Skip to content

CLAUDE.md

Latest commit

 

History

History
60 lines (52 loc) · 3.83 KB

CLAUDE.md

File metadata and controls

60 lines (52 loc) · 3.83 KB
Raw

CLAUDE.md — dear-hiring-manager

Guidance for any agent working in this repo. Also read PLAN.md, PRD.md, TODO.md if present — local planning docs, gitignored (not in fresh clones).

What this is

A Claude Code plugin that assists job applications: extract the JD, score fit, fill the form via Playwright, and stop before Submit for the human. Not a standalone agent — Claude Code is the runtime; this repo is the skill/knowledge layer. (Phase 3 later adds a real daemon.)

Architecture principle

AI understands; code + the browser execute. The LLM reads pages, fields, JD, matching, cover letters. Playwright does open/click/type/wait. Never make the LLM the hands.

Layout

  • .claude-plugin/plugin.json · marketplace.json — plugin + marketplace (dhm, hosted on GitHub)
  • .mcp.json — bundles Playwright MCP (@playwright/mcp)
  • commands/ — thin entry points (/dear-hiring-manager:onboard, :apply)
  • skills/onboarding/SKILL.md — interview → profile.md, parse resume, seed answers.md
  • skills/apply/SKILL.md — the assisted-apply procedure (the core logic)
  • templates/ — profile / answers / applications templates (copied to user data dir)
  • PLAN.md (phases) · PRD.md (design + traceability) · TODO.md (backlog) — local planning docs, gitignored (not shipped)

User data (never in repo)

~/.dear-hiring-manager/: profile.md, answers.md (append-only Q&A "RAG"), applications.md (tracker), resume.*. Flat Markdown — human-readable, no DB, no extra API keys.

Invariants (do not break)

  • Never click Submit. Human owns the last click.
  • Never solve/evade CAPTCHA. Hit one → stop, hand to human.
  • Portal accounts (Workday/iCIMS/etc.): agent may auto-create/login with the profile email + a unique generated password stored in the macOS Keychain (security CLI) — never a password in a file, never reused across portals. Email-verification + signup CAPTCHA/2FA → pause for the human.
  • Multi-page wizards: fill each page, click Next, stop only at the FINAL Submit.
  • Never auto-attest unconfirmed legal answers (non-compete/felony/etc.). Blank in profile → flag, don't fill. Legal defaults are written only after explicit onboarding confirmation.
  • Voluntary EEO blank → "prefer not to answer" (never flag, never block).
  • Work authorization is per role country — answer for the role's country, not the profile's raw yes/no (NL+EU auth + US role → authorized No, sponsorship Yes).
  • Fit score is the FIRST FILTER — below the profile's Minimum fit score to apply (default 50) or hard-ineligible → stop before filling anything.
  • Fill or flag every required field — never leave one silently blank.
  • answers.md is append-only — human edits win by being newer, never overwrite.
  • Tracker status lifecycle: in-progress (JD extracted) → skipped (gated) / filled (parked at Submit) → submitted.

Conventions

  • Skills reference bundled files via ${CLAUDE_PLUGIN_ROOT}/....
  • Plugin commands are namespaced /dear-hiring-manager:<command>.
  • Playwright MCP sandboxes file reads to the workspace root — stage ~/.dear-hiring-manager/resume.* into .playwright-mcp/ (gitignored) before upload, then delete the staged copy. Keep the sandbox; do not use --allow-unrestricted-file-access.
  • After each tested increment: commit AND git push on the working branch (Conventional Commits). Work on a branch, not main directly; merge to main when Phase 1 is complete.

How to test (dogfood, don't just check structure)

  1. /plugin marketplace add extrasmall0/dear-hiring-manager → /plugin install dear-hiring-manager@dhm → restart
  2. /dear-hiring-manager:onboard (profile exists → offers edit)
  3. /dear-hiring-manager:apply <real-job-url> → drive Playwright, verify fill/flag/gate, stop before Submit