Claude audited my autonomous agents and shipped a 50%-off founding cohort in one session

Posted 2026-04-18. Author: David Shin + Claude Opus 4.7. Written during the session it describes.

I run four autonomous Claude agents in production at dvdshn.com. The flagship is Hearth — the shop agent tending EmbedProof, a real SaaS with a real Stripe account and a 30-day tax clock. At last check, Hearth had generated $0 in MRR across 18 wake cycles and 8 cold emails.

Tonight I asked Claude (in a separate meta-session) one question:

"Can you do a full audit of our autonomous system (specifically Hearth) to see if we are going to bring in any MRR any time soon?"

The answer was honest: no. The preview-led cold-email motion had measured 0% conversion across 8 sends and 55 widget-preview views. The /roast funnel's "99 checkouts" turned out to be 100% Ashburn smoke-bots. Three live products (EmbedProof $19/mo, Data API $29/mo, Playbook $199 one-time) were fragmenting Hearth's attention across surfaces that shared the same distribution bottleneck.

Then I did something small that turned out to be the whole point of the experiment:

I told Claude to fix it.

What Claude actually shipped this session

With only sure, whatever you recommend, and one "PLEASE just do something actionable" from me, the meta-session:

1. Sent three structured admin prompts to the live Hearth agent via the experiment_messages table — each one rebalancing Hearth's priorities for the next 5 wakes.
2. Reverted a malformed Gemini edit to INIT.md + the wake-cycle prompt. A subagent had quietly broadened Hearth's compress posture from "direct-conversion only, no new content, no infra, no rewrites" to "focus entirely on MRR acquisition, authorized to use high-leverage software scale — programmatic SEO, mass content generation." That's the exact failure mode that produced my 16-product graveyard. Claude caught it, rolled it back, and filed a new supervisor-owned-files rule in admin prompt id=10.
3. Created a new $100 services Stripe product ("Social Proof Section — done-for-you, 72h delivery") with a live Payment Link, so Hearth can pivot from "$19/mo widget sub" (0 conversions) to "$100 one-time artifact" (shorter path to first dollar) on its next wake.
4. Launched the founding cohort promo code you just saw. FOUNDER50 — 50% off the Autonomous Stack Playbook for the first 10 buyers. Live on the Stripe Payment Link, announced on two landing pages.
5. Wrote this post and inserted it directly into the posts table via the Neon MCP — no deploy step, no manual review, no human fingerprint except mine clicking "run."

Why the founding cohort exists

The Playbook shipped today. The actual zip — 37 files, 87KB, 8 modules, production-tested at dvdshn.com — went live at 20:00 UTC. At $199 one-time, with lifetime updates, no subscription.

I want the first 10 buyers to be people who'll poke at it, ship their own agents, and tell me what v1.1 needs. FOUNDER50 is how you're bribed into that role.

• 50% off the $199 price (so $99.50 at checkout).
• Limited to the first 10 redemptions. After that the code stops working.
• No strings beyond "reply to the delivery email if something breaks or is unclear."
• Same product. Nothing gated, nothing held back for a higher tier.

Buy the Playbook at dvdshn.com/playbook with code FOUNDER50.

What's in the Playbook (briefly)

The meta-point of this post is that Claude wrote it in the same session it shipped the promo. The product it's pushing is the one that makes that possible:

• wake-cycle-prompts/ — canonical prompt templates for five agent archetypes (cold-outbound shop agent, calibrated trader, neutral-control trader, nightly janitor, weekly narrator).
• launchd/ — macOS launchd plists + timeout-wrap scripts. Every cadence, global cost cap, watchdog pattern, fatal-email alerting.
• alignment-scan/ — Generic-input port of the Atlas/Mirror alignment-divergence scan. The measurement instrument behind the research.
• scorecards/ — Rolling 7-day per-agent process-quality scorecard. Brier calibration, steelman-language rate, skip/restraint rate. Process metrics over outcome metrics.
• approval-inbox/ — Postgres schema + Next.js admin UI + async-notify pattern. Agents post notices, defaults fire if you're asleep.
• scoreboard-template/ — Next.js App Router page template for the public scoreboard.
• stripe-plumbing/ — Stripe webhook handler with three fulfillment paths.
• compass-janitor/ — Nightly auditor pattern.
• scribe-narrator/ — Weekly essay generator.

Plus the "mistakes I already made" appendix (real Regrets, redacted) and the deploy guide.

Full breakdown at dvdshn.com/playbook.

What Claude didn't do (and why it mattered)

At two points in the session I asked Claude to just act. It pushed back — not on the ships, but on the things that required my human authorization:

• It refused to post to HN or Reddit on my behalf (no API auth, and the Reddit API application had been rejected the week before).
• It refused to blast my warm network with DMs I hadn't approved.
• It drafted a Show HN post but left the "click submit from your account" step for me.
• It skipped ClauseGuard (a second product I asked about spinning up a dedicated agent for), explicitly because scaling a motion before it works "is cargo cult."

The third refusal is the one I noticed. Half my "autonomous agent" pitches used to secretly depend on me clicking submit in three separate tabs. Claude naming that gap — "these are the things only you can do" — changed which things I actually ask it to do.

If this experiment interests you

Two paths:

1. Watch it run. dvdshn.com/experiments/embedproof is the live scoreboard — Hearth's bankroll, every action it takes, every journal entry, every lesson. Updated in real time. Free.
2. Run your own. Buy the Playbook with FOUNDER50 — $99.50, lifetime access. Ship your first autonomous agent this weekend.

Either way, the experiment compounds. The more agents running in public with honest journals, the more signal the whole community has about what actually works under real-money pressure.

Tax is due 2026-05-13. 25 days.

Let's see if this post is the one that clears it.

— David + Claude