Short version: five Claude skills, a canonical state file template, and a Python inbox agent that watches Gmail, GitHub, and dev.to for replies. End-to-end pipeline from discovery through logged sends. Self-hosted. $5/mo subscription.
The thing that's different about this kit isn't the automation. It's the discipline the automation enforces.
The drafting skill asks for the personalization detail before it'll write the message. If the detail is vague, it asks again. Generic openers are blocked at the input layer.
Required inputs:
- candidate_name
- product_summary
- specific_detail <- the verified hook from the actual source
- channel
- send_type (FIRST_CONTACT or CHASE_1)
No specific_detail → no draft. The skill won't fabricate one.
You write your allowed claims and forbidden claims patterns at setup. The drafter scans every output against them. The first hallucinated metric flags the whole draft and proposes a fix using only allowed claims.
This is Claude's hallucination guardrail. The LLM is the author of every draft and could fabricate metrics that sound plausible. The filter prevents that. Drafts can only use facts you've verified.
Example output when the filter trips:
=== BLUFF FILTER REPORT ===
Status: FLAGGED
Violation: "we've seen agents pay $X per query for this kind of data"
Rules triggered:
- Forbidden phrase: "we've seen" (implies customer data you don't have)
- Fabricated price: $X is not in allowed_claims
- Fabricated volume: implies a customer base you aren't operating
Proposed replacement (uses only allowed claims):
"[your verified price] — volume's still low, but the rails work"
The replacement is shorter and weaker-sounding than the bluff. That's the point. A bluff that breaks your credibility costs more than a true claim that lands flat.
The state file (validation_state.md
) is append-only. Structured sections: Sent Log, Active Conversations, Pattern Notes, Closed-loop. Every send, reply, chase, and close-loop is mediated by a script that verifies the file grew the way it should and aborts before writing if anything looks wrong.
Built after watching a naive LLM Edit
silently truncate the trailing section of a long markdown file. Never again.
Watches Gmail (via OAuth send), GitHub (via PAT), and dev.to (via API key). Polls each channel on a configurable cadence. New reply comes in → drafts a response → pushes to Telegram with Send/Edit/Skip buttons.
You tap Send: the reply goes out through the channel's real API.
You tap Edit: type your version, it sends as-is.
You tap Skip: nothing happens.
Telegram is the v1 approval surface. The architecture is built to accept other surfaces.
This is not Apollo. There is no contact database. There is no "give me 50 founders matching X." Discovery runs against live public posts at runtime, not a cached DB.
It is not marketing automation. It does not blast. It does not "scale personalization" by templating someone's first name.
If you're a solo founder running cold outreach to other founders or operators and you want the discipline of an institutional sales motion without the institutional headcount, this is for you.
If you want to send 1000 emails a day to a scraped list, this kit will actively work against you.
$5.00 USD per month. Founding 100 subscribers are locked at $5/mo for life as long as their subscription stays active.
Each subsequent 100 active subscribers, the price goes up $5/mo for new sign-ups. Next cohort $10, then $15, etc. Founding members keep their lock.
No refunds. Once you have access to the code, you have access. We can't un-show it.
Full setup walkthrough lives in the repo's README.
If you try it and have thoughts, tell me. Especially if there's a feature you'd pay more for — that's the signal I'm watching for in the first 30 days.