agentleFS
automatewithudayautomatewithuday/GTM-Public-Traded-AI-Accounts/.claude/skills/score/SKILL.md

score

Skill0 starsChanged 4 months agoMITView on GitHub

Score enriched companies — deterministic base score plus LLM reasoning via cheap Haiku subagents (no API cost) — then qualify/disqualify. Use after /enrich, or when asked to score, rank, or qualify accounts.

16 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use audit.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • OpenAISkill

    audit

    Audit or critique a product flow, journey, workflow, funnel, onboarding path, checkout path, settings path, screen, or multi-step product experience by capturing screenshots first, then reporting UX, design, and accessibility findings inline from that evidence. Use Figma only when the user explicitly asks for a board. Use when the user asks to audit, review, critique, inspect, assess, analyze, evaluate, or give feedback on a product experience.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 4 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 14 edited versions we read keep these.

  • 4

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Don't score packets yourself in the main conversation (burns expensive context).

    line 36
  2. Distinctive:

    If a subagent fails on a packet, re-spawn just that packet.

    line 37
  3. Distinctive:

    ALWAYS batch: one packet per agent wastes the agent's fixed overhead.

    line 38
  4. Very specific:

    Rescoring the review band: After running /enrich --source deep, prepare packets with score --prepare --statuses scored to re-score …

    line 51
  5. Distinctive:

    Never re-roll to pass, never cherry-pick the highest — the median is the verdict, and commit computes it deterministically.

    line 72

“Common” is measured: kept by at least 6 in 10 of the 14 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

/score — score + qualify (v1: Haiku subagents, zero API cost)

Step 1 — prepare packets

uv run python -m pipeline score --prepare

Writes one JSON packet per company to data/scoring_queue/ (signals + evidence + base-score math + service catalog + rubric + required output schema). Prints how many packets are queued.

Deterministic pre-gate (v3.1): --prepare skips LLM scoring for companies whose verdict cannot change the outcome — no hard signal, or base score + scoring.pre_gate.max_llm_lift below the qualify threshold. Gated companies get a synthetic deterministic verdict written straight to data/scoring_results/ (labeled deterministic/pre-gate in the DB) and are listed in the pre-gated N line; a plain --commit handles them. Only spawn subagents for the packet paths --prepare prints — pre-gated packets already have their verdict.

v2 note: Packets may also carry an angles list (dated, structured outreach events from deep enrichment). The subagent verdict must return angle_ranking (fingerprint/family/message_hook) and primary_angle (fingerprint/family/why_this_angle) per the packet's output_schema — copy fingerprints exactly from the packet, use null or [] when no angles are present.

Step 2 — spawn lean scorer subagents

Take the packet paths printed by --prepare (pre-gated packets are excluded — never score those; their verdicts already exist) and spawn Agent tool subagents with subagent_type: scorer (the lean agent in .claude/agents/scorer.md — Read+Write only, haiku, no repo baggage), giving each subagent a batch of up to 5 packet file paths. Spawn batches in parallel (single message, multiple Agent calls). The scorer agent carries the scoring contract in its own system prompt, so the task prompt is just:

Score these packets. Packets: <absolute paths, one per line>

Rules:

  • subagent_type: scorer always — it pins haiku and the read-once/write-once contract. Never use opus/sonnet or a general-purpose agent for bulk scoring.
  • Don't score packets yourself in the main conversation (burns expensive context); delegate to subagents.
  • If a subagent fails on a packet, re-spawn just that packet.
  • ALWAYS batch: one packet per agent wastes the agent's fixed overhead. Fewer, fuller agents beat many small ones.

Step 3 — commit + qualify

uv run python -m pipeline score --commit

Validates every result against the schema (invalid ones are reported — re-run their subagent), writes scores to Supabase, sets profile, and transitions status: qualified (total ≥ threshold AND ≥1 hard signal AND ≥1 active angle in v2), disqualified (total < floor), else stays scored for human review. Processed files are archived to data/scoring_archive/.

v2 output: --commit now includes a kept bucket (already-qualified/contacts_found companies, never demoted) and [no_active_angle] markers for companies blocked by the angle gate. Export now includes angle columns: angle_ready, angle_family, primary_angle, message_hook.

Rescoring the review band: After running /enrich --source deep, prepare packets with score --prepare --statuses scored to re-score companies in the review band with angles.

Median-of-3 for borderline verdicts (code-enforced)

Haiku verdicts jitter ±10–15 points. --commit itself enforces the borderline rule — you cannot skip it: any single-shot verdict whose total lands within ±scoring.median_band (config, default 8) of scoring.qualify_threshold is held (reported as median-of-3 pending; the result and packet stay in place — nothing is written to the DB). To settle a held ticker:

  1. Spawn 3 scorer subagents in parallel; each replicate scores ALL pending packets (independence comes from separate agents, amortization from batching). The packets are still in data/scoring_queue/ — no re-prepare needed. Tell each replicate to write to the packet's output_path with a distinct suffix (.run1.json, .run2.json, .run3.json).
  2. Re-run score --commit. It finds the three replicate files, computes the median total itself, commits that verdict (labeled claude-code/haiku-subagent-median3), and archives all replicate files. Fewer than 3 replicates → still held; an invalid replicate is reported — re-spawn just that one.

Never re-roll to pass, never cherry-pick the highest — the median is the verdict, and commit computes it deterministically. Reply rates > volume.

Step 4 — report

Summarize: newly qualified (ticker, total, profile, top service fit, one-line reasoning), the review band, disqualified count. Suggest /people for qualified accounts. Never change thresholds yourself — propose to the user (config/settings.yaml → scoring).

Tiering + priority (v3): --commit now computes a tier for every processed company — T1 (total ≥ scoring.tiers.t1_min, default 80, AND qualified), T2 (qualified below that bar), T3 (review band), T4 (disqualified) — stored on both the scores row and the company's tier column. It also stores a priority composite (verdict total + a deterministic-base-score stacking bonus for evidence spanning ≥3 components + the strongest fresh outreach angle). /people and messages --prepare process companies in (tier asc, priority desc) order, so the strongest accounts get worked first when a per-run cap bites.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.