Job Application Pipeline

• Takes a job URL, fetches the posting, and scores it against a fixed résumé profile
• Returns structured JSON: fit score, stack match and gaps, live-coding risk, and apply priority
• Drafts a cover letter, bullets, and a short recruiter note only when the score clears 7
• Appends the result to a Google Sheet and emails a digest for human review
• A second workflow lists still-open drafts on weekday mornings and evenings and does not apply

┌──────────────────┐     ┌──────────────────┐     ┌──────────────────┐
│  Job URL form    │────►│  Fetch posting   │────►│  Fit score LLM   │
│  jobUrl + notes  │     │  HTML → text     │     │  structured JSON │
└──────────────────┘     └──────────────────┘     └────────┬─────────┘
                                                           │
                        score < 7                          │ score ≥ 7
                           │                               ▼
                           │                      ┌──────────────────┐
                           │                      │  Draft package   │
                           │                      │  letter + bullets│
                           │                      └────────┬─────────┘
                           ▼                               ▼
                  ┌──────────────────────────────────────────────┐
                  │  Google Sheets          status = draft_ready │
                  └──────────────────────┬───────────────────────┘
                                         │
                                         ▼
                  ┌──────────────────────────────────────────────┐
                  │  Gmail digest     you review, never applies  │
                  └──────────────────────────────────────────────┘

• Prompts only use the posting plus a locked résumé profile — no invented employers, metrics, or skills
• Parse nodes turn Continue On Fail off so bad JSON stops the run instead of writing a half row
• Eval fixtures cover a strong .NET/React match, a stack mismatch, and a live-coding flag
• The digest workflow only reads the sheet. It does not change status and it does not call the model
• Nothing in either workflow submits an application

• Local Docker n8n instead of a hosted runner

  The workflow exports as JSON, credentials stay off the repo, and the demo costs nothing to host.

• Cheap model for the fit score, stronger model for the draft

  Scoring is a structured JSON pass. The cover letter only runs when fitScore is 7 or higher.

• Google Sheets plus email, not a database

  The row is the review queue. A separate weekday workflow reads draft_ready rows and never writes them.

• OpenAI returns nested or fenced JSON — a Code node flattens output text and JSON.parse fails loud
• Fit score under 7 — the draft step is skipped and nothing is emailed to an employer
• No draft_ready rows — the digest workflow sends no email
• Laptop asleep or n8n stopped — that digest slot is missed and the next weekday run lists the same open drafts
• Posting asks for live coding — liveCodingRisk is set to likely so the review shows it

1. 1docker compose up -d and open n8n at http://localhost:5678
2. 2Import n8n/workflow.json and n8n/digest-workflow.json, then reconnect OpenAI, Sheets, and Gmail
3. 3Open the form URL and paste a job posting
4. 4Check the Applications sheet for the fit score and, at 7 or above, the draft
5. 5Confirm the weekday digest only lists draft_ready rows and that no employer is emailed