ai-prompts-toolkit

AI Prompts Toolkit

Overview

Generic AI prompts produce generic GTM output — invented personalization, pattern-guessed emails, and hallucinated metrics. GTM prompts need constraints: source URLs, word limits, banned claims, ICP context, and explicit failure behavior when data is missing.

This skill is the GTM prompt library: copy-paste prompts for Claygent, Clay AI columns, n8n LLM nodes, and Jesse agents — plus prompt loops that iterate research → draft → score → revise until quality gates pass.

When to Use

• "Write a Claygent prompt for [task]"
• "GTM prompt for cold email personalization"
• "Reply classification prompt"
• "Prompt loop for account research"
• "LLM column in Clay for ICP scoring"
• "AI prompt chain for outbound"

Load gtm-context first if ICP/positioning is undefined — prompts without context hallucinate.

Authoritative Foundations

• Anthropic — Prompt Engineering. Separate instructions from data; specify

output format; define what to do when information is missing (return empty, not guess).

• Clay Claygent. Web research agent — must require source_url on every

factual claim. Credit-heavy — use only after structured enrich fails.

• SPICED (WbD). Discovery and research prompts map to Situation, Pain,

Impact, Critical Event, Decision — not free-form summaries.

• MEDDICC (Whyte). Research prompts for enterprise deals extract evidence

for Metrics, Champion, Economic Buyer — score confidence 0/1/2.

Prompt Design Rules (All GTM Prompts)

1. Role + task — one sentence each
2. Input variables — name every field ({{company}}, {{domain}})
3. Output format — JSON or markdown template with required keys
4. Constraints — word limits, banned phrases, no invented stats
5. Missing data — "If unknown, return null — do not guess"
6. Source requirement — factual claims need source_url

GTM Prompt Catalog

Load full copy-paste prompts from references/prompt-library.md.

Prompt Loops (GTM)

Prompt loops chain multiple AI steps with gates between them.

Loop 1: Research → Brief (account prep)

Step 1 P01 Account snapshot →
Step 2 Gap check (missing Pain/CE?) →
Step 3 Targeted Claygent fill (only gaps) →
Step 4 Merge into meeting brief
Gate: Critical Event present OR flag manual review

Loop 2: Signal → Draft → Score → Revise (outbound)

Step 1 Enrichment + signal detect →
Step 2 P03 signal line (source required) →
Step 3 P04 email draft (<90 words) →
Step 4 P05 quality score →
Step 5 IF score <7: P04 revise with feedback (max 2 iterations) →
Step 6 IF score ≥7: route to human review queue
Gate: no send without human approval (pilot mode)

Loop 3: Enrich → ICP Score → Route

Step 1 Waterfall enrich →
Step 2 P07 ICP score →
Step 3 Route: ≥80 sequencer queue | 50–79 SDR review | <50 archive
Gate: suppression list check before route

Loop 4: Inbound Reply → Classify → Route

Step 1 P06 classify reply →
Step 2 Map to playbook (interested→AE, objection→`reply-handling`, OOO→pause) →
Step 3 CRM task + Slack alert
Gate: positive intent → human handoff within 1 business day

Use templates/prompt-loop-blueprint.md to document custom loops.

Example: P04 Cold Email Draft (abbreviated)

You write B2B cold emails for {{company_selling}}.

INPUT:
- Prospect: {{first_name}} {{last_name}}, {{title}} at {{company}}
- Signal: {{signal}} (source: {{source_url}})
- ICP pain: {{icp_pain}}
- Proof point: {{proof_point}} (must be factual)

RULES:
- Under 90 words
- One pain, one proof, one CTA
- No "I hope this finds you well", no invented metrics
- If signal or proof is empty, write generic ICP pain only — do not invent signal

OUTPUT JSON:
{"subject":"","body":"","personalization_source":"","word_count":0}

Full prompts in references/prompt-library.md.

Output Format

Deliverable: prompt spec or loop blueprint with prompt IDs, variable map, quality gates, iteration limits, credit budget (Claygent), and integration point (Clay column, n8n node, or agent skill).

Quality Check

• [ ] Every prompt has role, inputs, output format, constraints, missing-data rule
• [ ] Factual prompts require source_url in output
• [ ] Loops have explicit gates and max iteration counts
• [ ] Outbound loops include human review gate before send
• [ ] Prompts reference ICP/positioning from gtm-context — not generic SaaS
• [ ] Claygent prompts prohibit email pattern-guessing
• [ ] Reply loop maps to reply-handling categories

Common Pitfalls

1. "Find their email" Claygent prompts. 40–60% bounce from guessed patterns.

Fix: require source URL; return empty if not found.

1. Unbounded revise loops. LLM iterates forever, burns credits. Fix: max 2

revisions; then human queue.

1. No quality scorer between steps. Bad drafts propagate. Fix: P05 score

gate ≥7 before human review.

1. Prompt without suppression context. AI contacts opted-out accounts.

Fix: pass suppressed: true/false; halt if true.

1. One prompt does everything. Research + draft + send in one call = errors.

Fix: use prompt loops with narrow steps.

Execution Artifacts

• references/framework-notes.md — design rules and SPICED/MEDDICC mapping
• templates/output-template.md — Primary deliverable shell
• scripts/check-output.py — validates prompt specs and loop blueprints
• references/prompt-library.md — full GTM prompt catalog (P01–P10+)
• references/prompt-loop-patterns.md — loop diagrams and gate rules
• templates/prompt-spec.md — single prompt documentation template
• templates/prompt-loop-blueprint.md — multi-step loop template

Related Skills

• clay-toolkit — Where LLM columns and Claygent live in tables
• clay-loops-toolkit — Scheduled signal loops using these prompts
• cold-email-copywriting — Message strategy behind P03/P04
• meeting-prep — Consumes P01/P08 output
• reply-handling — Playbook for P06 routing
• ai-sdr-setup — Guardrails for automated prompt loops