Validate startup ideas with evidence. Test them with real outbound. Measure booked calls, not vibes.
A Claude skill for idea-stage founders: five gated stages that take a raw idea to a live outbound experiment — with every claim verified, every send human-approved, and every outcome recorded.
S1 Validate → evidence-scored verdict (demand · AI-platform kill-risk · GTM viability)
S2 Hypothesize → personas + falsifiable outbound bets, ICPs as live-validated Apollo filters
S3 Source → credit-disciplined lead sourcing, fit scoring, domain dedupe, site-scan verification
S4 Write → mechanism-testing variants, linted per rendered email, verified claims only
S5 Send & Learn → 3-touch sequences, per-batch approval, intent/sentiment reply classification
As a plugin (recommended — gets updates):
/plugin marketplace add vkuchkanov/gtm-machine
/plugin install gtm-machine@gtm-machine-marketplace
As a bare skill: copy plugins/gtm-machine/skills/gtm-machine/ into your ~/.claude/skills/ directory, or package it as a .skill zip and import it in Claude/Cowork settings.
Then just describe your idea to Claude: "validate my idea: …"
- Honest by construction. Scores cite evidence by id or are rejected programmatically. The rubric is calibrated to disagree with the founder when evidence demands it — the eval suite includes a self-flattery guard and a discrimination pair (two ideas sharing every keyword, deserving opposite verdicts).
- Never lies to prospects. Any claim about a recipient's business requires a scan record with a matched token and timestamp. The scanner fails closed: "couldn't look" never becomes "nothing there".
- Nothing sends itself. No scheduled auto-send exists. Every batch needs a fresh human approval bound to the exact lead set and a hash of the rendered copy; edits kill the approval.
- The whole trail is inspectable. Ideas, hypotheses, leads, sends, and replies land in spreadsheets the founder owns, with evidence links and timestamps.
Works with zero setup for stage 1 (validation uses web search). Later stages use Apollo (paid plan with credits), a JS-rendering scraper (Firecrawl), Google Drive/Sheets, and an email send path — the skill walks founders through connecting each one at the moment it's needed (references/setup.md).
Pre-first-conversion, and the skill says so to its users: the process passed its own evals (11/11 golden-set verdicts, live-validated ICPs, a verified end-to-end send→reply→classify loop) but the reply-rate bars are starting hypotheses, not measured benchmarks. This README will be updated when they are.
.claude-plugin/marketplace.json ← makes this repo an installable marketplace
plugins/gtm-machine/
.claude-plugin/plugin.json
skills/gtm-machine/
SKILL.md ← workflow, gates, honest status
references/ ← per-stage rubrics + setup walkthroughs
scripts/ ← runnable stage modules (Python)
evals/ ← seeded test prompts
MIT — see LICENSE. The moat isn't the code; it's what you learn running it.