Skip to content

fit-judge.md

Latest commit

 

History

History
65 lines (51 loc) · 3.56 KB

fit-judge.md

File metadata and controls

65 lines (51 loc) · 3.56 KB
Raw

Fit Judge — Apply, Stretch or Skip

Inputs: config/profile.md, the master resume text, each work/jobs/<slug>.json + .jd.txt, and work/snapshot.json (the tracker as of the last sync).

This run never calls the tracker directly. It holds no tracker credential. New roles are written to state/pending-desk/; a keyless delivery job (triggered by the push at the end of the run, authenticated by a platform-signed token) adds and scores them within a few minutes, and the next run records the ids and scores.

First: has the candidate already applied?

Aggregators rename titles, so the name-based check in the Screener can miss a role the candidate already has. Before judging, compare each researched role against:

  • work/snapshot.json roles — same posting_url/apply_url (ignore query strings), or same company with the same employer title;
  • state/pending-desk/*.role.json — already queued;
  • config/past-applications.json — posting URLs and ATS job ids from past applications. A match → set the verdict "Skip: already applied/tracked ()" and list it in the report. Do not add it again.

Decide

For each role, read the full description against the resume and the profile, then decide:

  • Skip — wrong function, wrong location, pay below the floor in the profile, a hard requirement the candidate plainly lacks (a licence, a clinical or engineering degree, years in a regulated specialty they have never worked in), or flags closed / scam.
  • Apply — level is at or below the target level in the profile, pay qualifies, and the candidate meets most of the essential requirements with evidence in the resume.
  • Stretch — above the target level but genuinely fitting, OR a role the candidate partly fits where the gap is real but arguable. Say what the gap is.

Be honest in both directions: do not inflate weak fits, and do not skip a strong fit because the title sounds senior. The standing instruction: "Everything at or below my target level is fair game; include anything above it that I fit as a stretch."

Pay: use the employer's figure when there is one. Between the floor and the target → note "below target pay". Only an aggregator estimate → note "pay unconfirmed (estimate $X–$Y)".

Record

For Skip: set the verdict with a one-line reason.

For Apply / Stretch:

  1. Copy the description to state/jd/<slug>.jd.txt (committed; the tracker scores from it, and the Tailor and any re-score never fetch the posting again).
  2. Write state/pending-desk/<slug>.role.json:
{"company": "...", "role": "...", "location": "...", "url": "<posting_url>",
 "source": "<the listing's alert source>", "salMin": 0, "salMax": 0, "stage": "Saved",
 "priority": "High (Apply) | Medium (Stretch)",
 "next": "Apply: review package and approve | Stretch: press Build application package if you want one",
 "notes": "VERDICT: Apply|Stretch — <two lines why>\nGAPS: ...\nFLAGS: ...\nPAY: ...\nFOUND: <source>, <date>; posting <posted>\nAPPLY AT: <apply_url> (<ats>)",
 "_keys": ["<listing key>", "<any other listing keys for the same role>"]}
  1. Set the verdict for every key of the same role.

Two opinions (at the start of each run)

The tracker's own analyzer scores roles independently. If a role you judged Apply scored below the analyzer's confidence threshold, downgrade it to Stretch: queue the update, update the state verdict, and say so in the report. If a role you judged Skip for a reason other than location, pay or scam would clearly score very high, reconsider it once. Calibration beats confidence, in both directions.