Datatrolleri · internal platform · briefing for 2026-08-10
Ventures Garden takes a business idea from “someone had a hunch” to “strangers are giving us their email address” in about half a day of one person’s time — then does it again, across a team, so that what is learned compounds.
10 stages5 human gates 3 turns per person, maximum27 documents 0 completed ventures
The bet
A venture never starts with code. It starts with evidence a market exists and ends with strangers handing over an email address, before a line of product is written. Most ventures are then killed — that is the design working, not a failure.
Success is not a launched venture. Success is cost per signup falling week over week, and a ledger that makes venture #40 start further along than venture #1.
The accumulated ledger of kills. Any single venture is disposable; knowing which categories, positions and creative approaches convert for us is not. Most portfolio tools delete their failures. This one is built so failure is the asset.
How work moves
Work is claimed under an expiring lease, never assigned. You are one claimant among several, and anything you hold returns to the board when you stop holding it.
* means nobody is holding it.Three turns at once, maximum. That cap is not a rate limit — it is what stops one person claiming six things and blocking five of them. Reviewer attention, not money or compute, is what caps how many ventures launch in a day.
Track 01 — app pre-sell · the only track today
Each stage has a prompt in tracks/01-app-presell/prompts/, kept in two
forms: the original verbatim source, and the hardened version we actually run.
Amber marks a stage that cannot advance without a person.
Replace guessing with evidence — where is money already moving.
Enter the hot category through a door nobody is standing in.
The one document every later stage reads from.
Make the product look real enough to sell before it exists.
The instrument that measures demand. Everything before this was preparation.
Teach Meta’s AI what a good customer looks like.
Many shots on goal, cheaply. 42 formats, 11 banned by construction.
A URL a stranger can open. Our own box, behind Caddy.
Put the offer in front of strangers. Left in draft — a person publishes.
Make the system self-improving.
Six gated stages against five platform gates: gate 5 covers both stage 8 and stage 9, because both spend money.
Not advisory
Ventures Garden is fast because the machine does the work. It is good because a human stops it at five points. A track may add a gate; nothing may remove one. A waiver changes who clears a gate, never whether it exists.
The model returns five positions. You pick one, and the four you rejected are recorded with the reason — they travel into every later briefing.
Read it end to end. Push back until it describes something you would want to exist.
Anything hallucinated, off-brand or fake-looking gets regenerated.
Models will happily write “loved by 12,000 future plant parents” on day zero. Delete it. Invented social proof does not ship, ever. Normally needs a second person.
Ads are drafted, never auto-published. This is the gate that stands between a draft and real money. Normally needs a second person.
When there genuinely is no second person, the producer clears their own gate under a recorded, expiring waiver — after a cooling-off period timed from when the work was put down, with notes, marked in the ledger forever.
Two rules that never bend
No invented counts, testimonials, ratings, press logos or urgency. The instruction goes in every generation prompt, every time, because models re-add it on regeneration. A linter greps for the patterns; a person still reads the page as a stranger would.
Stages 8 and 9 leave campaigns in draft. A person publishes. This is enforced by a hook on the Meta mutations that can start live spend — it asks, every time, rather than relying on a model having read the rule.
What you operate
Every capability the track needs, and what provides it. Health is measured by
./scripts/doctor.sh and recorded — the platform knows its own state
rather than assuming it.
| Provides | Tool | Stages | State |
|---|---|---|---|
| App revenue data | AppMagic | 0 | a person in a browser |
| Image & video generation | Higgsfield (hf CLI) | 3, 6, 9 | healthy |
| Campaign management | Meta Ads MCP | 5, 8, 9 | healthy |
| Network observation | Chrome DevTools MCP | 5 | healthy |
| Static hosting | Caddy on app-01 | 7 | not observed |
| DNS record write | Simply.com API | 7 | not observed |
| Lead capture | venturesgarden-leadd | 4 | live |
| Lead storage | Postgres on data-01 | 4, 7 | live |
| Skill | What it does | State |
|---|---|---|
| /gate-honesty | Runs the fabrication linter, reads the page as a stranger, writes findings into VENTURE.md, then opens the gate for a second person | shipped |
| /venture-domain | Registers a domain, points a venture hostname at the box, diagnoses a deploy that failed at DNS | shipped |
| /gate-review | The four gates that have no skill — position, spec, assets, pre-publish | to write |
| /standup | Steward digest: spend, what is stalling, blocked on whom | to write |
| /kill-review | Friday pass over every live venture, default kill | to write |
A skill never writes what the ledger
records — it calls venturesgarden. Gates, reviews, claims and decisions
are rows; written into markdown they become a second version that disagrees.
Read this part twice
The platform’s own voice rule is to state its gaps out loud. Every number here was measured on 2026-08-08, not estimated.
Success is defined as cost per signup falling week over week. Signups reach the
ledger; spend does not, because the Meta performance sync is not built. The ledger
renders those columns as visibly unmeasured rather than as zero — a column
nothing feeds shows —, never 0.
There is no command that records a kill. The venture.decision table
exists and is empty; none of the 30 CLI subcommands or 24 service routes writes it.
The three killed ventures were killed with raw SQL. This sits exactly where the
method claims its value — the accumulated ledger of kills — and it is the largest
hole in the platform today.
Every venture targets the US, but the ad account’s day boundary is Europe/Stockholm and Meta refuses to change it after creation. A replacement cannot be created either — the portfolio’s ad-account limit is 1. Accepted deliberately: cost per signup is a ratio and is unaffected, so this is a reporting offset. Any surface showing a date must say which clock it is on.
Monday
docs/first-venture.md — a
throwaway venture, start to kill, about an hour. It costs nothing and touches no
advertising. Do this before the first real one, so that the first time you meet a
lease expiring is not on a venture that matters.docs/running-a-venture.md — the
whole flow from having no account to killing or scaling.tracks/01-app-presell/playbook.md — the ten stages in detail.docs/honesty.md../scripts/doctor.sh — is this machine set up? Fails loudly.venturesgarden board — what needs a person, and the command that takes it.venturesgarden guide list — every document: kind, slug, path, summary.venturesgarden guide due — manual work the platform cannot do itself.venturesgarden tools — what provides each capability, and whether it works.impeccable, which the design linter shells out to, is a personal
skill in ~/.claude/skills/ and is not committed. The design surfaces
depend on it, so a second operator gets its output on git pull without
the tool that produced it. You do not need it to run a venture.