V1 in five steps, on one page
Six teams are done or closed. Six jobs are left, and every one of them is buildable now except one that waits on your five answers. Then the loop runs once for real.
Done and checked: T1 tickets, T2 live reader, T3 coach (it wrote a real brief, graded B), T5 Umgesetzt with the live check, T10 switch (built, the key is not live yet) and T11 openers (closed: 7.7% are not recognised, and that is accepted). The T6 judge design is written too.
Nothing is blocked by an agent. What is left is one five-point answer from you, then building.
One chat with 6,000 messages costs more than the $0.30 day on its own, so the reader stops there every minute and never gets to the rest. It is a cost wall, not a logic error. The fix is noted on #37.
The road to V1, in order
Five steps. Each one needs the one before it, so this is a real sequence, not a wish list.
You answer the 5 judge pointsHow the judge decides: #38. This is the only step that waits on a person.
you · #38T6 builds the judgeThe design is done; the build starts the moment the five points are in.
agents · T6You give the yes for A33.writer and A34.liveBoth cost cents each. They unlock wording that a person approves and a live check.
you · cents eachT8 runs the fake full cycleThe whole loop once on fake data: idea, wording, approval, Umgesetzt, verdict. #35.
agents · T8You and Max run the first real cycleSame loop, real chats, real booking. When it lands, V1 is proven. #36.
you + Max · #36Who feeds whom, and where each one stands
Read left to right. An arrow means "this team's output is what the next one needs." Green is done, blue is building now, grey waits for its input.
What is left, one line each
Each team has its own files and branch, so two teams never edit the same thing. Six teams are finished and are not listed here.
T6 Judge
#38after your 5 pointsThe design is written: it decides on how many people reply to the changed message, with bookings shown beside it. The build is one team, and it starts the day the five points are answered.
Hands on: the verdict at the end of every cycle.
T4 Lists
#39#40#42building nowThe said-yes list leaves out known clients, lists never time out, and the daily "must be zero" rules go in. Nothing here needs a person.
Hands on: lists people can trust without re-reading them.
T7 Close the built ones
#17#24#26checks onlyBuilds nothing. Runs the missing checks on work that is already finished and closes the tickets that pass, so the open list finally means something.
Hands on: a GitHub list that matches reality.
T8 Fake full cycle
#35after T6 + your yesThe whole loop once on fake data: idea, wording, approval, Umgesetzt, verdict. Nothing real is touched, so a mistake costs nothing.
Hands on: proof the loop works before it runs on real chats.
T9 Control room
#43after your 7 answersEvery agent and its cost live on the dashboard, with a pause switch per part. Spec is written; it waits on the seven answers.
Hands on: you can see and stop anything.
T12 Backup
#13not urgentThe nightly backup goes live as soon as there is a second Supabase project to write into. It is the only job here that can wait without hurting anything.
Hands on: a second copy of everything.
Your list and Max's list
Your list is what unlocks the agents; Max's list is essentially done.
You
- The 5 judge points (#38)unlocks step 2, the judge build
- Yes to A33.writer and A34.live (cents each)unlocks the fake cycle, T8
- #8 the Umgesetzt screencloses what T5 built
- #21 point B: may the chat promise a transformation?closes the rules file
- #24: the reader questions under 95%closes the reader
- #43: the 7 control-room questionsunlocks T9
- #13: a 2nd free Supabase project (no card)unlocks the backup, T12
Max
Send the opener textsdone 20-11, and T11 is closedName the takeover tagdone: it isai_off, andkiausmeans friendsThe 2 KI Flow fixesdone 20-11Switch test with you, 3 passesdone 3 times · #31- Keep the switch test going: does the AI's answer fit what was said while it was off?alongside, not blocking
- Run the first real cycle with youthe last step, #36
Every ticket, in one of three piles
22 open, 22 closed on GitHub. Most of the open ones are built and wait on a yes or a check, not on code.
Built: only needs checking or closing
Agents build now
Waits on you or Max
Checking for logic errors without trusting the AI
Your question. Each check rests on code, the CRM or a person, not on a model's opinion.
Count it twiceSeparate code counts every number again. Last run: 1,068 numbers, 0 differences.
existsCompare with the CRMA sample of our numbers checked against the CRM itself.
T7 finishes it"Must be zero" rulese.g. nobody on the said-yes list has an appointment. If the count is ever above 0, the logic is broken.
going in with T4Planted answersFake data where we know the right result. The judge ran 1,000 times on it.
existsYour answer keyThe reader is scored against the 50 chats you graded by hand.
existsRead 10 rows yourselfOnce a week per list. One read already showed only 1 in 6 on the said-yes list were real misses.
new · you or MaxDid they book?The final test: people on the lists booking, according to the CRM's own appointments.
new · you + Max