twenty-partner-application-triage

Rank APPLICATION-stage partners by the value they would add — geographies and languages we don't yet cover, plus proof of real Twenty work — and hand back a short chase-list: the high-value applicants worth a personal nudge.

The door stays open for everyone. This skill does not reject or filter anyone out of the pipeline. Booking a call is the motivation filter; this just makes sure the good applicants who didn't book bubble up so they don't rot. The point is a few high-confidence partners, not maximum coverage.

Read-only. It never mutates a record.

Credentials

Needs ~/.twenty/credentials.env (same file the other partner skills use):

TWENTY_PARTNERS_API_URL=https://partners.twenty.com
TWENTY_PARTNERS_API_KEY=<your key>

The key lives in packages/twenty-apps/internal/twenty-partners/.env.prod (gitignored) — copy it to ~/.twenty/credentials.env on first setup.

Phase 0 — Run the ranker

python3 "$(dirname "$0")/rank.py"          # or: python3 rank.py from the skill dir

rank.py is the deterministic core. It pulls every partner, computes each applicant's net-new geo / language / scope / skills vs the VALIDATED baseline, detects a "real Twenty work" proof signal in the notes, scores, and prints ranked JSON. It does not call an LLM — the judgment lives in you (Phase 1).

Each ranked entry: name, score, tier (A/B/C), new_geo, new_lang, new_scope, new_skills, proof{workspace_url|customers|migration}, team, contact_name, email, linkedin, website, notes.

Scoring (in rank.py, tune there if it drifts): geo +3 each, language +3 each, scope +1, skills +1 capped at 3 (so generic dev shops that spray skill lists can't dominate), proof +6. Any proof signal ⇒ at least tier A.

Booking state: if the JSON has booking_state_wired: false, the callBookedAt field isn't on the model yet, so the ranker scores all applications. Say so in the output. Once callBookedAt exists, the ranker auto-narrows to the un-booked (the true chase set) — no skill change needed.

If the run prints a missing-credentials error, stop and tell the user exactly which key to add and where.

Phase 1 — Judgment pass (this is the point)

The score surfaces; you decide. Read the notes of the top ~15 plus anything tier-B/C with a non-trivial note, and adjust:

• Rescue the motivated-but-unobvious. Someone whose checkboxes are thin but whose notes show real intent, a live Twenty instance, named customers, or a thoughtful pitch is a chase even at a low score. This is the whole reason a human/LLM reads the notes — the applicants who do things we don't see in the form are exactly who we don't want to lose.
• Sanity-check volume inflation. A high score driven by 6 net-new languages from one solo, or a long skills list, may be aspirational. Confirm it against the notes before ranking it top. Real net-new geography with proof beats a long list every time.
• Demote noise. Empty notes, agency spam, or "Tally submission: " with nothing else is tier C regardless of score. Don't chase them.
• Note proof quality. workspace_url + customers together (a live workspace with named clients) is the strongest signal — stronger than the raw score. Call it out.

Don't invent facts. If a note is ambiguous, say so rather than upgrading on a guess.

Phase 2 — Output the chase-list

A tight digest, grouped by tier, A first. Lead with the count and the booking-state caveat.

# Partner application triage — N applications ranked (booking state: not wired / un-booked only)

## Chase now (Tier A — fills a gap AND/OR proven)
- **<name>** (<team>) — +<geo>/<lang>; proof: <workspace+customers/…>
  why: <one line, drawn from notes>
  reach: <email> · <linkedin>

## Worth a look (Tier B)
- <name> — <one line>; reach: <email>

## Skip for now (Tier C) — <count>, not listed (empty/spam/no gap)

Rules:

• Tier A is the actual worklist. Keep it short — if it's 18 long, the proof-backed gap-fillers go first and the volume-only ones go last.
• Always give a reach handle (email from contact_name/email, else linkedin, else website). If none, say "no contact on record" — that itself is a data-quality flag.
• Be honest. If only a handful are genuinely worth chasing, say so; don't pad the A-tier.

What this is not

• Not a gate. It never moves anyone to REJECTED or out of the funnel.
• Not a writer. It never edits a record. Surfacing only.
• Not the production cron. This is the dev surface for the ranking. Once the chase-list is trustworthy, the deterministic core (rank.py) is what gets ported to a daily logic-function cron in the partner app that writes ranking + a tier onto each un-booked record so the workspace view sorts itself. The LLM judgment pass stays here, for the runs where you want a human in the loop. Build skill → trust it → port the cheap part. Don't build both.

Self-check

python3 rank.py --selftest asserts the scoring orders a gap-filler-with-proof above a skill-sprayer above an empty record, and that skill volume stays capped. Run it after any edit to the weights or signal regexes.