Jonathon's AI Wiki

Career-Ops — Multi-Agent Job Search on Claude Code

6 min read

Source: ai-research/career-ops-repo-2026-06-02.md (full README, fetched 2026-06-02) / raw/gh-star-santifer-career-ops.md (gh-stars puller stub). Repo metadata verified live via the GitHub API on 2026-06-02.

Repo: santifer/career-ops (career-ops.org) | Stars: ~48,300 (10K forks, 205 watchers, 126 open issues) | Created: 2026-04-04 | License: MIT (+ trademark policy) | Stack: Node.js + Go + Playwright

Career-Ops turns any agentic coding CLI — Claude Code (primary), plus OpenCode and Gemini CLI — into a job-search command center. It is a useful worked example of the agentic-CLI-as-command-center pattern: a single slash command (/career-ops) fans out into ~12 modes that scan job portals, score offers against a structured rubric, generate tailored PDFs, and track a pipeline — with batch parallelism via claude -p workers and Playwright driving real browser navigation. The maintainer reports using it to evaluate 740+ offers, generate 100+ tailored CVs, and land a “Head of Applied AI” role. ^[inferred — the case-study outcome is the author’s own claim, not independently verified]

Key Takeaways

  • One slash command, ~12 modes. /career-ops dispatches: full auto-pipeline (paste a JD → evaluate + PDF + tracker), scan (portal sweep), pdf (ATS-optimized CV), batch (parallel multi-offer eval), tracker, apply (form-fill), pipeline, contacto (LinkedIn outreach), deep (company research), training (evaluate a course/cert), project (evaluate a portfolio piece).
  • A–F rubric scoring, 10 weighted dimensions. Offers are scored by reasoning about your CV vs the JD (not keyword matching), with a 6-block evaluation: role summary, CV match, level strategy, comp research, personalization, interview prep (STAR+Reflection). The system explicitly recommends not applying below 4.0/5 — it is positioned as a filter, not a spray-and-pray tool.
  • Playwright portal scanner. 45+ companies pre-configured (Anthropic, OpenAI, ElevenLabs, Retool, n8n, …) + 19 search queries across Ashby, Greenhouse, Lever, Wellfound, Workable, RemoteFront. --verify launches Playwright after the API pass to drop expired postings before they reach the pipeline.
  • Batch parallelism via claude -p workers. /career-ops batch evaluates 10+ offers in parallel using headless claude -p worker processes — a concrete production use of Claude Code’s headless mode + sub-agent fan-out for throughput.
  • Self-customizing by Claude. Modes, archetypes, scoring weights, and negotiation scripts are plain markdown the agent reads and edits — “just ask Claude to change them.” Onboarding is framed as nurturing a new recruiter: feed it your CV, career story, proof points, and preferences; quality compounds with context.
  • Human-in-the-loop by design. The system evaluates and recommends but never submits an application — the user always has the final call. Pipeline integrity is enforced with automated merge, dedup, status normalization, and health checks (single source of truth in a .tsv tracker).
  • Multi-CLI + multi-model. Native Claude Code, OpenCode, and Gemini CLI support (Codex “soon”); the same modes/*.md evaluation logic drives all of them. Gemini works free (OAuth or gemini-2.5-flash API), so the architecture is genuinely model-agnostic.
  • Verified, real, organic. MIT-licensed, 2-month history (created 2026-04-04), ~48K stars with ~10K forks / 205 watchers / 126 open issues — a healthy fork-to-star ratio that rules out bot inflation. 8-language README, Discord community, trademark policy.

How it works

You paste a job URL or description
        │
        ▼
  Archetype Detection   →  classifies LLMOps / Agentic / PM / SA / FDE / Transformation
        │
        ▼
  A–F Evaluation (reads cv.md)  →  match, gaps, comp research, STAR stories
        │
   ┌────┼────┐
   ▼    ▼    ▼
 Report  PDF  Tracker
  .md    .pdf   .tsv

ATS PDFs are keyword-injected and design-templated (Space Grotesk + DM Sans). An Interview Story Bank accumulates STAR+Reflection stories across evaluations into 5–10 master stories; negotiation scripts cover salary frameworks, geographic-discount pushback, and competing-offer leverage.

Why it matters (the reusable pattern)

The job-search domain is the demo; the transferable architecture is what earns this a wiki entry. Career-Ops is a clean reference implementation of paste-input → archetype-route → rubric-score → tailored-artifact → tracked-pipeline, built entirely on an agentic CLI with:

  • Markdown-as-config the agent edits (the same modes/*.md it executes) — the skill-as-editable-markdown pattern taken to a full application.
  • Batch claude -p worker fan-out for high-volume evaluation — a productivity-scale use of headless Claude Code.
  • Playwright for real-world data acquisition behind the rubric.
  • Human-in-the-loop on the irreversible action (never auto-submits) — the same discipline the Mythos Preview system card recommends for autonomous agents.

Swap “job offers” for leads, RFPs, vendor bids, or content briefs and the same scaffold applies — which is why it sits alongside the wiki’s other Claude-Code skill-collection repos rather than in a careers silo.

Try It

  1. git clone https://github.com/santifer/career-ops && cd career-ops && npm install && npx playwright install chromium, then npm run doctor to validate prerequisites.
  2. Copy config/profile.example.yml → profile.yml, add a cv.md, then open claude in the directory and ask it to adapt the modes/archetypes to your field — the system is built to be reconfigured by Claude itself.
  3. Steal the batch claude -p worker pattern for any high-volume “evaluate N items against a rubric” task in your own stack — it’s the most reusable piece for the WEO context (lead scoring, brief triage).
  4. Steal the --verify Playwright pass idea: trust a fast API feed first, then launch a browser only to prune stale/expired entries before they pollute a pipeline.

Open Questions

  • Independent verification of the case-study outcome (740+ offers, 100+ CVs, the role landed) — these are the maintainer’s self-reported figures; no third-party validation.
  • Skill/prompt-injection surface. A Playwright agent reading untrusted job-posting HTML and editing its own config markdown is exactly the threat surface security tooling targets — the repo’s handling of malicious-listing content is not assessed here.
  • Star velocity. ~48K stars in 2 months is unusually fast for a niche productivity tool; worth watching whether engagement (issues/PRs/forks) keeps pace or plateaus.

Graph View

Backlinks