icp-prompt-builder
Interactive loop that builds and tunes an AI prompt for evaluating whether a company fits a client's ICP. Run after any list-building skill (disco-like, blitz-list-builder, google-maps-list-builder, prospeo-full-export) to qualify companies before scaling. Iterates batches of 10 companies with user feedback, stops when 2 consecutive rounds have zero corrections, saves the final prompt for reuse. Always uses Claude Code Task sub-agents — never an external API key.
growthenginenowoslawski · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add growthenginenowoslawski/coldoutboundskills --skill icp-prompt-builder技能说明
name: icp-prompt-builder description: Interactive loop that builds and tunes an AI prompt for evaluating whether a company fits a client's ICP. Run after any list-building skill (disco-like, blitz-list-builder, google-maps-list-builder, prospeo-full-export) to qualify companies before scaling. Iterates batches of 10 companies with user feedback, stops when 2 consecutive rounds have zero corrections, saves the final prompt for reuse. Always uses Claude Code Task sub-agents — never an external API key.
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:
- Save the final tuned prompt to
~/cold-email-ai-skills/profiles/<business-slug>/icp-prompt.txt - 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
- 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 rou