fivos-lead-scoring

Fivos Lead Scoring

Qualification + tiering for accounts and contacts. Broad-then-narrow: source wide, filter down with a prompt. Scoring runs in-session through Claude itself - no external model needed today.

How scoring runs (today)

• Claude reads the input CSV row-by-row (using the Read tool on a sample, then processing batches).
• For each row, Claude applies the validated qualification prompt and outputs a structured decision: {tier, score, reasoning}.
• Claude writes the scored CSV back to disk using Write/Edit.
• No external model wiring required. First-class, not a fallback.

If you have a dedicated local qualification model (via an MCP or API), this skill's Step 2 and Step 5 can be swapped to call it. The rest of the flow - sample, audit, two gates - stays identical.

Hard rules

• Two approval gates per the TAM Mapping and Contact Sourcing Plan - 2026:
  • Gate 1: engineer approves / revises the qualification prompt after seeing 10-row sample results.
  • Gate 2: re-estimate full-run cost with the validated prompt, manager approval before scaling.
• Every scored row carries reasoning. Store tier, score, reasoning triples for audit - not bare tiers.
• Re-scoring uses this same skill. Never write custom scoring scripts outside this skill. If criteria change mid-campaign, re-invoke.

Inputs

• CSV to score - account CSV (from fivos-account-enrichment), contact CSV, or both.
• Qualification prompt - engineer-provided. If missing, the skill suggests one from the template below.
• ICP matrix - from fivos-icp-segmentation. Gives signal weights and tier thresholds.
• Tier thresholds (optional) - override the default A/B/C cutoffs if the client has specific definitions.

Outputs

1. Scored CSV with columns appended: tier, score, reasoning, scored_at, scorer_version.
2. Tier distribution report (counts of A/B/C/DQ, top disqualification reasons).
3. Audit trail: the prompt used, the gate approvals received.

Prompt template

If the engineer didn't provide a prompt, start here and fill in the bracketed sections from the ICP:

You are qualifying <client>'s outbound list. The ICP is:

- Industry: <industry>
- Size: <headcount_range>
- Signals that raise fit: <list from ICP signal library>
- Known disqualifiers: <list>

For each row you receive, decide tier A/B/C:
- A = strong ICP fit + at least one high-weight signal firing.
- B = ICP fit, no strong signal, or signal firing on a weaker-fit account.
- C = tangential fit. Keep only if volume is needed.
- DQ = disqualify (wrong geo, wrong industry, sanctioned, bankrupt, etc.).

Return JSON per row: {"tier": "A"|"B"|"C"|"DQ", "score": 0-100, "reasoning": "<2-3 sentences>"}.

Always show the engineer the fully-rendered prompt before running.

Workflow

Step 1 - Prompt assembly

Render the template with the ICP doc's values. Show the rendered prompt to the engineer for a quick read.

Step 2 - 10-row sample (in-session, Claude-driven)

Read the first 10 rows of the input CSV. For each row, apply the prompt and produce the structured output. Present results inline:

Row 1: Acme Corp (domain: acme.com)
  Tier: A
  Score: 87
  Reasoning: Strong industry fit; Series B within 45 days (high-weight signal);
  headcount in target band; no disqualifiers.

Row 2: Globex (domain: globex.io)
  Tier: DQ
  Score: 10
  Reasoning: Wrong industry (consumer retail); also HQ in sanctioned region.

...

Step 3 - Gate 1 (prompt validation)

Ask the engineer:

10-row sample complete.

Distribution: A=X, B=X, C=X, DQ=X

Audit the reasoning. Do any look wrong?

Options:
  [approve] - proceed to Gate 2
  [revise]  - edit prompt and re-run sample
  [abort]   - stop, do not score

Iterate on the prompt until the engineer signs off. Common revisions: tightening a disqualifier, re-weighting a signal, adding a geo exclusion, adding an industry keyword.

Step 4 - Gate 2 (full-run approval)

Once the prompt is validated, present the full-run plan:

Full qualification plan - <client> / <campaign>

Rows to score: N
Scorer: Claude (in-session)
Estimated time: ~M minutes (typical: 2-4 seconds/row in batched reads)

Manager approval required to proceed.

Step 5 - Full run

Read the CSV in batches (50-200 rows per batch depending on row width), apply the prompt per row, append the scoring columns, write the result to <input>_scored.csv.

scorer_version: "claude-<model>-<date>"  # e.g. claude-opus-4-7-2026-04-22

Step 6 - Distribution + anomaly report

Qualification complete - <client> / <campaign>

Total scored: N
Distribution:
  - A: X (X%)
  - B: Y (Y%)
  - C: Z (Z%)
  - DQ: W (W%)

Top disqualification reasons:
  1. <reason> (X rows)
  2. <reason> (Y rows)
  3. ...

Flags:
  - Unexpectedly high DQ rate (>25%) - review prompt tightness.
  - Tier-A concentration in one industry - possibly over-indexed.

Step 7 - Handoff

Hand the scored CSV to fivos-copywriter. Tier drives copy depth: A = deep personalisation, B = signal-triggered, C = pattern-only.

Cutover to a dedicated qualification model

If you wire a local or API-hosted qualification model:

1. Replace Step 2 + Step 5 with calls to that model.
2. Update scorer_version in the output CSV to reflect the model.
3. Everything else (sample → Gate 1 → Gate 2 → full run → report) stays the same.

No structural change to this skill is needed beyond swapping the two step bodies.

Re-scoring

If prompt criteria change mid-campaign (new disqualifier, re-weighting), re-run this skill on the existing CSV. Always bump scorer_version and scored_at so downstream analytics can see the change.

References

• ~/.agents/skills/gtm-meta-skill/writing-outreach.md - scoring rubric examples, prompt patterns.
• ~/.agents/skills/gtm-meta-skill/prompts.json - qualification prompt templates.
• TAM Mapping and Contact Sourcing Plan - 2026 (Google Doc).