icp-prompt-builder

ICP Prompt Builder

Before you pay to pull 5,000 companies, tune a qualification prompt on 10-50 of them. This skill walks you through the iterative loop.

Why this exists

List-builder skills (DiscoLike, Blitz, Prospeo, Google Maps) return COMPANIES, but they don't know whether those companies match your ICP. If your list-builder returns 5,000 companies and 80% are wrong fits, you'll waste money enriching them for emails that go nowhere.

The fix: build an AI qualification prompt BEFORE scaling. Pull 10 companies, have the prompt score them, compare to your judgment, refine, repeat. Once the prompt agrees with you 2 rounds in a row with zero corrections, lock it in and apply it at scale.

Always uses Task sub-agents (no API key)

This skill runs entirely inside Claude Code via the Task tool. No Anthropic SDK calls, no OpenAI calls — Claude Code does the scoring itself. This is intentional:

• No extra API spend. Uses your Claude Code plan.
• No key management. Works out of the box.
• Scaleable within reason. For 20-100 evaluations, parallel Task sub-agents batch 10-20 companies per agent.

At very large scale (5,000+ companies per batch), you may want to export the tuned prompt and run it through the OpenAI / Anthropic API with parallelism for speed. But TUNING happens inside Claude Code.

The loop (8 steps)

Step 1 — Gather ICP context

Claude asks the user (or reads client-profile.yaml from /icp-onboarding):

• Website of the client selling (to scrape for context)
• Who IS a good customer? What makes them a good fit?
• Who is NOT a good customer? What disqualifies them?
• Any specific signals? (B2B only, revenue range, tech stack, hiring status, recent fundraise, etc.)
• Any HARD disqualifiers? (competitor domains, existing customer domains, certain industries/geographies)

Step 2 — Select 10 test companies

Pull 10 companies from the list-builder output:

• Mix likely-good and likely-bad fits
• Variety in industry, size, location
• Each company needs at minimum: domain, company_name, industry, headcount, description
• Richer fields (Clay-derived: Business Type, Scale Scope, Revenue) make scoring better

Step 3 — Build the initial qualification prompt

Template:

You are an ICP evaluator for {CLIENT_NAME}.

## Target ICP
{ICP description from user or client-profile.yaml}

## Qualification criteria (MUST be true)
- {criterion 1}
- {criterion 2}
- ...

## Disqualification criteria (ANY match = disqualify)
- {disqualifier 1}
- {disqualifier 2}
- ...

## Input
You will receive a company with these fields:
- domain, name, industry, headcount, description
- (optional) Business Type, Revenue, Scale Scope

## Output
For each company, return JSON:
{
  "qualified": true | false,
  "confidence": 0.0-1.0,
  "reason": "one-sentence explanation"
}

Step 4 — Run the prompt on the 10 companies

Via the Task tool. Launch one Task sub-agent that reads the prompt + 10 companies, returns 10 JSON scores.

Step 5 — Present results to the user

Format as a table:

Company                   | Qualified | Conf | Reason
--------------------------+-----------+------+----------------------------------------
acme-corp.com             | YES       | 0.92 | B2B SaaS, 200 employees, target industry
random-nonprofit.org      | NO        | 0.95 | Nonprofit, not a business customer
edge-case-company.com     | YES       | 0.55 | Could fit but revenue model unclear

Step 6 — Collect user feedback

Ask specifically:

• Which evaluations are wrong? (e.g., "acme-corp should be NO because they're a competitor")
• Which are right but for the wrong reason?
• Any patterns the prompt missed?
• Any new disqualifiers to add?

If the user has zero corrections, log this round as "approved."

Step 7 — Refine the prompt (or move on)

If the user gave corrections:

• Add/remove qualification criteria
• Tighten/loosen disqualifiers
• Add specific examples of edge cases ("companies like X are NOT a fit because Y")
• Adjust confidence thresholds if everything is coming back 0.5

Then go back to Step 4 with a NEW batch of 10 companies.

Step 8 — Stop condition + save

The loop ends when 2 consecutive rounds have zero corrections from the user. When that happens:

1. Save the final tuned prompt to ~/cold-email-ai-skills/profiles/<business-slug>/icp-prompt.txt
2. Append metadata to client-profile.yaml:

icp_qualification_prompt:
  path: profiles/<slug>/icp-prompt.txt
  tuned_at: YYYY-MM-DD
  rounds_to_convergence: 3
  final_batch_size: 10

3. Print a one-liner for the next skill:

Prompt locked. To score your 5000 companies:
  npx tsx ~/cold-email-ai-skills/skills/icp-prompt-builder/scripts/score-batch.ts \
    --prompt-file=profiles/<slug>/icp-prompt.txt \
    --companies=path/to/companies.csv \
    --out=scored.csv

Approval-loop rules (important)

• Never auto-approve. Even if the prompt looks right, require the user to explicitly say "approved" or give zero corrections for 2 consecutive rounds.
• Reset counter on any correction. One correction resets the streak to 0.
• Don't skip the batches. Running 30 companies all at once feels faster but masks errors. 10 at a time is the right batch size — small enough to eyeball.
• Show the prompt each round. After each refinement, display the current full prompt back to the user so they can see what changed.
• Always use Task tool sub-agents for the scoring inside each round. Never call external APIs.

Using the tuned prompt at scale

Once saved, the prompt is applied to the full list via scripts/score-batch.ts. Options:

Option A (free, slow) — run through Claude Code Task sub-agents in batches of 20 companies per agent. Good for <500 total.

Option B (paid, fast) — export prompt + companies to OpenAI / Anthropic API with parallelism. Good for 500-50,000.

The script supports both. Default is Option A to keep everything inside Claude Code.

Recommended flow

1. /icp-onboarding → produce client-profile.yaml
2. /disco-like OR /blitz-list-builder OR /prospeo-full-export → pull a sample of 50-100 companies
3. /icp-prompt-builder → tune qualification prompt on that sample (3-5 rounds typical)
4. Scale the list-builder to 5,000+ companies
5. Apply the tuned prompt to the full list → only keep qualified: true with confidence >= 0.6
6. /blitz-list-builder or /email-waterfall on the qualified subset
7. Upload to Smartlead

Data points the prompt can use

From most list-builder outputs:

• domain, company_name, industry, headcount, description, LinkedIn URL

Additional fields (if enrichment skills have been run):

• Business Type (B2B / B2C / B2B2C)
• Annual Revenue range
• Scale Scope (Enterprise / Mid-Market / SMB)
• SubIndustry (more specific than primary industry)
• Tech stack (Clearbit, BuiltWith data)
• Recent signals (funding, hiring, news)

Tell the AI about the fields you have access to in the prompt preamble.

Common mistakes

• Building the prompt too tight on round 1. Start broad, narrow with feedback.
• Not including negative examples. "Companies like Netflix are NOT a fit because they're B2C" is more powerful than generic "must be B2B".
• Using only "qualified: true/false" without confidence. Always ask for confidence — 0.5-0.7 borderline cases are where you learn the most.
• Scoring 50 at once "to save time." Defeats the point of the loop.
• Not saving the prompt. The point of tuning is reuse. If you don't save, you'll re-tune next time.

Scripts

• scripts/score-batch.ts — apply tuned prompt to a CSV of companies

What to do next

Apply the tuned prompt to your full list (the list-building skill you came from — Prospeo, Blitz, DiscoLike, Google Maps, or Competitor Engagers — will walk through this). Then /list-quality-scorecard to grade the filtered output.

Or wait: if the prompt didn't converge within 5 rounds (you kept making corrections), your source data may be too thin. Enrich with more fields (company description, headcount, tech stack) before retrying.

Related skills

• /icp-onboarding — run FIRST to produce client-profile.yaml
• /disco-like, /blitz-list-builder, /google-maps-list-builder, /prospeo-full-export — pull the companies this skill qualifies
• /personalization-subagent-pattern — same approval-loop pattern, applied to copy personalization