Daily pipeline. Searches job boards using your profile keywords, scores each listing with a local LLM, and for the top matches generates a tailored resume + cover letter with Claude. Writes per-job folders and emails a digest.

Job Hunt — Daily

Daily automation: search → score → tailor outputs for high-match jobs → digest.

What it does

profile.json + resume.txt           (cached by jobhunt-setup)
        │
        ▼ extract-keywords     →  "python,kubernetes,senior,…"
        ▼ job-search           →  ~50 listings across 4 boards
        ▼ score-listings       →  per-listing {url, score 0-10, rationale}
        ▼ generate-outputs     →  for each score ≥ 7: tailored resume + cover letter
        ▼ send-digest          →  "5 new matches today" message to the user

Per-job output folder structure:

/home/me/jobhunt/2026-05-13/
├── acme-corp-senior-backend-engineer/
│   ├── listing.md         ← original posting + match score + rationale
│   ├── resume.md          ← Claude-tailored, emphasizing matching experience
│   └── cover-letter.md    ← personalized to this listing
├── …
└── DIGEST.md              ← one-line summary per match

Setup before first run

Run jobhunt-setup first. It produces the two files this workflow depends on (profile.json + resume.txt).
Credentials: - ollama — pointed at your local Ollama server. Used for cheap, private scoring. - claude-api — Anthropic API key. Used for tailored resume + cover letter.
Adzuna (optional). For broader board coverage (Indeed/LinkedIn via aggregation), sign up free at developer.adzuna.com and paste your app_id + app_key into the search-jobs task properties. Without it, the workflow still uses Remotive, We Work Remotely, and HN Who-is-hiring.
Schedule it. Use InTouch's Workflows view to attach a daily schedule (e.g. weekdays at 8 AM). Or run on-demand whenever you want a refresh.

Tunable knobs (in the YAML)

Where	What	Default	Why change
`score-listings.model`	Ollama model for scoring	`llama3.2`	Use `llama3.1` (8B) or `mistral` for better JSON adherence.
`generate-outputs.SCORE_THRESHOLD`	Min score to generate outputs	`7`	Raise to `8` for stricter; lower to `5` to see more.
`generate-outputs.MAX_OUTPUTS`	Cap on tailored outputs per run	`5`	Each tailored output is 2 Claude calls (~$0.02 each).
`generate-outputs.MODEL`	Claude model	`claude-sonnet-4-6`	`claude-haiku-4-5` is cheaper if you'd rather spread across more listings.
`search-jobs.maxPerSource`	Listings per source	`30`	Lower if scoring runs slow on your machine.
`search-jobs.country`	Adzuna country code	`us`	`gb`, `ca`, `au`, `de`, `fr`, `in`, etc.

Cost rough estimate per daily run

Step	Provider	Calls	Approximate cost
`score-listings`	Ollama (local)	1	$0 — local compute
`generate-outputs`	Anthropic (claude-sonnet-4-6)	2 × MAX_OUTPUTS (default 10)	$0.10-0.30
`search-jobs`	HTTP only	3-4	$0

Adzuna's free tier is 100 calls/day — well above a daily run's needs.

Failure modes worth knowing

Ollama returns malformed JSON for scoring → the python step strips markdown fences but if the model goes off-script entirely, the JSON parse will fail. Fix: change score-listings.model to a larger model.
No matches at threshold → script exits 0 with "No matches at score >= 7" and send-digest sends an empty-ish message. Lower the threshold or broaden keywords.
Anthropic 429 / overload → script raises and the task fails. Re-run.
User has no email channel → send-digest falls back to whatever channels the user has (in-app inbox if nothing else configured). Set up SMTP / a contact's email for a real digest email.

Manual inspection helpers

After a run:

ls /home/me/jobhunt/$(date +%Y-%m-%d)/
cat /home/me/jobhunt/$(date +%Y-%m-%d)/DIGEST.md

Listings + scores from the latest run:

cat /tmp/jobhunt-listings.json | jq '. | length'
cat /tmp/jobhunt-scored.json   | jq '.[] | select(.score >= 7)'