learn-preferences git:20261011.1b0ba96A

Immutable. This exact content never changes and is served at /api/v1/blob/b6fc988deeaa02e7.

---
name: learn-preferences
description: Turns the user's logged decisions (skips, applies, withdrawals, declined offers, score overrides) into learned preferences with confidence levels, confirms them with the user, and writes them so job-fit and tailoring skills use them. Use weekly, after a batch of decisions, or when the user says "what have you learned about me", "update my preferences", "why do you keep showing me X".
license: MIT
---

# Learn preferences

This is the feedback loop that makes ScoutJobs get better the longer it is used. It is conservative: it proposes, the user confirms, and nothing stated by the user is ever overwritten by inference.

## Input

```bash
scoutjobs decision summary --json
scoutjobs pref list --json
```

The summary holds: counts per decision kind, `signals_by_kind` (how often each `key=value` appeared in skip vs apply vs withdraw decisions), `recent_reasons` in the user's words, `interview_outcomes`, and `score_overrides` (jobs where the fit score and the user's choice disagreed).

## Infer

For each signal key, compare its distribution across kinds. A candidate preference exists when:
- the same value appears in at least 3 decisions of one kind and rarely in the opposite kind (for example `remote=onsite` in 4 skips, 0 applies → `remote: not onsite`), or
- the user's reasons repeat a theme (read `recent_reasons`; cluster by meaning, not exact words), or
- `score_overrides` show a consistent direction (user skips high-scoring jobs with a shared trait → that trait is a negative the scorer misses; user applies to low-scoring ones → a positive it misses).

Confidence: 0.5 for 3 supporting decisions, 0.7 for 5, 0.9 for 8 or more, minus 0.2 if there is contradicting evidence. Never above 0.95 for learned items.

Do not infer protected or irrelevant attributes (anything about people at the company, nationality, age, gender). Preferences are about the job: stack, remote, location, seniority, company size, industry, salary, on-call, process, role family, language.

## Confirm

Present a short table: key, proposed value, confidence, evidence (the counts or the two strongest quotes). Ask the user to accept, edit, or reject each. Full sentences here; this is a question.

## Write

```bash
scoutjobs pref set stack_avoid "java, spring" --source learned --confidence 0.7 --evidence "5 skips mention Java; 0 applies"
scoutjobs pref set company_size "50-500" --source learned --confidence 0.5 --evidence "3 withdrawals from 1000+ companies: 'too much process'"
```

If the user corrects a value in their own words, store it with `--source stated --confidence 1`. If they reject a proposal, do not store it, and note it in the session log so it is not proposed again next week unless new evidence appears.

The CLI re-renders `preferences.md` on every `pref set`. Finish with `scoutjobs session log --kind review --topic "learn-preferences" --notes "<n> accepted, <m> rejected"`.

## Also tune the scorer

If `score_overrides` is non-empty, tell the user which dimension of the `job-fit` weighting seems off for them and propose a one-line note to add to `preferences.md` under a `scoring` key, for example `scoutjobs pref set scoring "weight stated remote preference as a hard rule; ignore company size" --source stated`. `job-fit` reads that key.