ftja-tune
Update rubric.md/criteria.json from a conversational statement of a new preference or dealbreaker. Invoke when, in the course of ANY conversation about jobs/career in this project, the user states a new preference, dealbreaker, or pivot ("I'm not interested in traditional PM roles", "drop the onsite-required ones") — proactively offer to add it, don't silently apply it.
TaegyuJEONG · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add TaegyuJEONG/ftja --skill ftja-tuneAbout this skill
name: ftja-tune description: Update rubric.md/criteria.json from a conversational statement of a new preference or dealbreaker. Invoke when, in the course of ANY conversation about jobs/career in this project, the user states a new preference, dealbreaker, or pivot ("I'm not interested in traditional PM roles", "drop the onsite-required ones") — proactively offer to add it, don't silently apply it.
FTJA tune
This is the explicit-confirmation half of "self-improving rubric" (see SPEC.md's Future Vision section for why the implicit/automatic half is deliberately out of scope). Never edit rubric.md or criteria.json without the user confirming first.
When to trigger
The user says something, in passing or directly, that reads as a job-search
preference, dealbreaker, or a scope pivot (e.g. moving from one kind of role
to another). This can happen mid-conversation about something else that
touches on career/roles — you don't need an explicit /ftja-tune invocation.
What to do
-
Read the current
rubric.mdandcriteria.json. -
Restate the new signal back to the user in one line and ask: "should I add this to the judgment criteria?" Do not just add it.
-
Check for conflict against the existing rubric/criteria. If the new statement contradicts something already there (e.g. rubric says "remote preferred" and the user now says "drop the onsite-required ones" — that's not a conflict, that's a tightening; but if it contradicts something like an existing dealbreaker being loosened, or two preferences that can't both hold), surface the conflict explicitly and ask which one wins. Don't silently pick.
-
On confirmation, edit the file (the rubric for narrative criteria,
criteria.jsonfor structured filters — a "no on-site" statement likely touches both: anexclude_keywordsentry AND a rubric Fail criterion, ask if unsure which).The rubric exists in two forms that must stay the same rubric:
rubric-criteria.json(a list of single criteria, which the user also edits in the viewer's Pipeline tab) andrubric.md(what the models read). Ifrubric-criteria.jsonexists, make the change there — add an entry (kind:pass/fail/preference, alabelthat reads on its own, the rule astext) or reword one, leaving every other entry'sidas it is — then rewriterubric.mdfrom it:(Windows: `venv\Scripts\python` in place of `venv/bin/python`.) venv/bin/python -c "import json; from ftja.verdict import save_criteria; d = json.load(open('rubric-criteria.json', encoding='utf-8')); save_criteria('.', d['criteria'], d.get('notes'))"Editing
rubric.mdas text instead would leave the list stale, and the user's next save in the Pipeline tab would overwrite your edit. Only whenrubric-criteria.jsondoesn't exist yet (no run since setup) editrubric.mddirectly. -
Do not commit:
rubric.mdandcriteria.jsonare personal files that git ignores on purpose, so they never enter the repository's history. Nothing to regenerate — the review server's Pipeline tab readscriteria.json/rubric.mdlive on every request. -
Confirm the change back to the user in one line.
What NOT to do
- Don't infer preferences from behavior (which digest jobs they clicked, applied to, etc.) — there's no tracking of that in this version. Only act on things the user actually said.
- Don't batch up multiple unrelated changes into one edit — one conversational update, one change, confirmed back to the user, so each one can be undone on its own.