Job Pipeline

job-pipeline

End-to-end orchestrator. One invocation takes a job from unknown to ready-to-submit package + outreach plan with no human prompting in between. The human is only needed at two points: setting policy (once, in references/decision-policy.md) and approving submission.

State: the ledger

leads/pipeline.json at the repo root is the single source of truth. Read it before doing anything; write it after every state change.

• If leads/ or leads/pipeline.json doesn't exist yet, create the directory and initialise the ledger to {} (then proceed).
• Keyed by job URL. Statuses: discovered → scored → preparing → prepared → notified → approved → submitted → interviewing → offer, terminal: rejected | passed | expired.
• Never re-fetch, re-score, or re-prepare a URL that's already in the ledger unless its status is discovered or the user explicitly asks. This is what stops every run from redoing old work.
• When a posting turns out closed/older than 5 months, set status expired with a note — keep the entry so it stays deduplicated.
• Update the updated field on every write.

Decision policy

Read references/decision-policy.md before scoring or deciding anything. It encodes Kristian's goals (innovation, compensation, speed) as score modifiers and action thresholds. The policy decides what happens to each job — not the user, not ad-hoc judgment.

Command routing

Mode: Full run

Designed to work unattended (cron, remote trigger, --print). Never ask interactive questions in this mode; make the policy decision and record it.

1 — Sweep the existing pipeline first (don't miss what you already have)

Before searching for anything new:

1. Read leads/pipeline.json.
2. For every job in status prepared or notified: re-check it's still open (fetch-job.sh <url>, check .status). If closed → status expired, flag prominently in the final report ("missed: X closed before submission").
3. For every prepared job missing answers: true: generate application-questions.md now using generate-application Step 9, set answers: true.
4. For every GO job missing contacts: true: invoke find-linkedin-contacts, save output to the package folder, set contacts: true.
5. Flag any job sitting in prepared/notified for more than 7 days — these are the leaks; speed is a stated goal.

2 — Search

Invoke the find-kristian-jobs skill, Full Search mode (its budgets and freshness rules apply). Before scoring, drop every URL already present in the ledger. Add each genuinely new posting to the ledger as discovered.

3 — Score and decide

Score each new posting with the find-kristian-jobs matrix, then apply the modifiers and thresholds from references/decision-policy.md. Record score, recommendation, and the decided action in the ledger (scored). Apply the policy actions:

• AUTO-PREPARE: run the full generate-application skill (analysis, cv-data.js, cover letter, application-questions.md) and find-linkedin-contacts, end-to-end, without asking. Ledger → prepared, with package, answers, contacts fields set.
• NOTIFY-ONLY: ledger → scored, include in report for human decision.
• PASS: ledger → passed with one-line reason. Never silently drop a job — every discovered URL gets a ledger entry.

4 — Report and notify

Write the run summary to leads/YYYY-MM-DD-pipeline-run.md:

• New jobs found / prepared / passed (with reasons)
• Pipeline health: stale items, expired items ("missed opportunities"), items awaiting approval
• For each prepared package: path, score, deadline if known, and the exact next action ("review answers, then /job-pipeline submit <url>")
• Any ⚠️ NEEDS INPUT fields from application-questions.md files

Then push the summary to Kristian (he is not at the terminal during scheduled runs):

1. Telegram if the channel is available: send a short digest (counts + top items + needed decisions) via the telegram reply tool.
2. Otherwise PushNotification with a one-line summary.
3. The openclaw system event command if running inside the OpenClaw container (detect via $OPENCLAW_* env or /home/node/.openclaw existing).

Run budget

One full run = find-kristian-jobs budgets + max 3 AUTO-PREPARE packages. If more than 3 jobs qualify, prepare the 3 highest-scored and leave the rest as scored with action AUTO-PREPARE (deferred) — the next run picks them up.

Mode: Status report

Read the ledger, print a table grouped by status (active first), flag: stale prepared items, unknown deadlines on GO jobs, ⚠️ NEEDS INPUT markers, anything expired in the last 14 days. End with the single most valuable next action.

Mode: Approve

Set the job's status to approved. Verify the package is complete (analysis, cv-data.js, cover letter, application-questions.md, contacts); generate anything missing. Re-check the posting is still open. Then offer /job-pipeline submit.

Mode: Assisted submission (human stays in the loop)

Submission is the one irreversible, outward-facing step — it is never done unattended.

1. Job must be approved (or the user is asking interactively right now — that counts as approval).
2. Open the apply page: mcp__claude-in-chrome__tabs_create_mcp → mcp__claude-in-chrome__navigate → mcp__claude-in-chrome__read_page.
3. Fill every field from application-questions.md using mcp__claude-in-chrome__form_input; upload the CV PDF with mcp__claude-in-chrome__file_upload (ask for the path if no PDF exists in the package — or build it from cv-data.js first).
4. Refuse to proceed if any ⚠️ NEEDS INPUT field is unresolved.
5. Stop before clicking the final submit button. Screenshot the filled form (mcp__claude-in-chrome__computer, screenshot action), show it, and get an explicit yes. Only then click submit.
6. Ledger → submitted with date; suggest sending the P1 outreach messages from the contacts file.

Hard rules

• Ledger first, ledger last — a run that doesn't update leads/pipeline.json is a failed run.
• Never submit, send, or publish anything external without explicit human approval in the same conversation.
• Honest packages: same authenticity rules as generate-application — no invented experience, gaps stated.
• All money/relocation/start-date answers come from the decision policy, never improvised.