farzamhejazik/claude-linkedin-assistant

1. This repo is a Claude Code command pack for job search automation, not a conventional app. The architecture is centered on slash commands in .claude/commands/ that Claude Code loads and executes inside the desktop client.

2. The main orchestration boundary is the /jobs command family. jobs.md is the entry point, while the subcommands in .claude/commands/jobs/ split the workflow into setup, discovery, tracking, outreach, daily orchestration, and Indeed-specific setup.

3. The system has a very small state model built around local files. The primary sources of truth are:

- resumes/ for user identity, target roles, skills, and pitch material

- job_tracker.csv for all job records and status

- optional resumes/search_profile.md for preference filtering

- outreach/*.md for per-company outreach logs

4. The job tracker CSV is the central operational datastore. It holds job metadata, status transitions, referral flags, deadlines, and URLs. The README explicitly says status updates are edited directly in the spreadsheet/CSV rather than through a separate app layer.

5. The resume folder is an input analysis layer. Claude reads all resume variants to infer the user’s name, target roles, skills, and outreach pitch. This drives both platform verification and job matching, so the resume content affects multiple downstream flows.

6. The search profile is an optional policy layer that narrows discovery. It adds user preferences such as remote-only, salary floor, or deal-breakers. This is read fresh by /jobs find, so it acts as a lightweight config file rather than a persistent settings system.

7. The repo depends on Claude in Chrome + the Claude Code desktop app as the execution runtime. Claude Code handles the commands, while the Chrome extension is the browser automation bridge. That browser bridge is essential because LinkedIn and Indeed interactions happen through real browser sessions, not APIs.

8. The external service boundary is primarily LinkedIn, with Indeed as optional discovery support and web search as an additional source for finding roles. LinkedIn automation is intentionally limited to safe, repetitive actions: search, profile inspection, connection requests, and first-message drafts/sends.

9. The daily flow is the top-level pipeline. /jobs daily coordinates setup checks, dashboard review, job discovery, manual add capture, outreach for companies needing referrals, and a final git commit. This makes the repo’s operational shape “check → find → add → outreach → commit.”

10. There is a strong human-in-the-loop boundary. Claude is allowed to do repetitive work and update the tracker, but replies, application submission, file uploads, and other judgment-heavy or fragile actions stay manual. The architecture is deliberately conservative to reduce risk with LinkedIn limits and account safety.

11. Privacy is handled by local-only artifacts and gitignore boundaries. Resumes and outreach logs are excluded from version control by default, while the committed resumes/README.md and the tracker template document expected local structure.

12. In diagram terms, this repo is best modeled as: Claude Code commands at the center, local filesystem state on one side, and browser-driven external services on the other side, with job_tracker.csv as the main shared data store and /jobs daily as the orchestration entrypoint.