For teams shipping with AI
Turn tribal knowledge into an operating system that makes any builder a 5x builder.
We meet your teams wherever they sit on the AI adoption curve: already AI-pilled and trying to cut review time, or adopting this year and wanting to hit the ground running.
Your teams already know how to do their hardest jobs well. That tribal knowledge is real, spread across a few concentrated circles and an unsearchable Slack channel. We turn it into an operating system that plugs into your coding agent, Claude or Codex, with standing rules, cited facts and tests that live locally and load every session. Lightweight and local by design, so humans and agents both work faster and at higher quality. You own it outright.
Scoping call is free. A 30 minute conversation, and no deck.
Sound familiar?
You ship faster than you can check.
You read every line anyway.
The agent gets it eighty percent right, which is the worst possible score. Close enough to look finished. Wrong exactly where it costs you. So somebody senior reads all of it, which is the job you bought the agent to remove.
the review you were trying to deleteNone of it was ever written down.
Your team knows how this job should be done. That knowledge lives in review comments, Slack threads and the way three people happen to do it. Not one of those is a place an agent can read.
real knowledge, no homeYou explain the same thing every time.
New starter, new contractor, new branch. The same context, out loud, from scratch, degrading a little with every retelling because nobody is reading from the same page.
context, retyped foreverThe same mistake keeps coming back.
Caught in review, fixed, explained. Nothing in the repository remembers, so it returns two months later in somebody else's branch and gets caught again by the same tired person.
no memory, no progressThe work is high stakes and hand made.
A regulatory response. A migration. A design system. It matters, it comes round constantly, and how well it goes depends on who happened to pick it up that week.
quality by coincidenceAnd you cannot hire your way out.
A new senior head is another opinion about what good looks like, plus six months of absorbing the unwritten rules by osmosis. The knowledge stays tribal. There is simply more tribe.
more people, same problemThis is the month before almost every conversation we have. Everyone involved is good at their job, and the knowledge is genuinely there. The work still comes out inconsistent, because none of it is written anywhere the agent looks when the work actually happens.
The actual problem
Your agent got faster. Your review queue got longer.
Your model already knows more than it manages to apply. The highest value you bring to it is how the context and feedback orchestration works. What is missing is the structure around the model: the standards it should hold, the facts it must cite, and the checks that tell it when it has drifted. Every session starts from zero, the model fills the gaps with something plausible, and a human supplies the standard afterwards by reading the output and wincing. A prompt cannot carry that. It is typed fresh every session and gone the second the session ends.
Put the standard where the agent already looks and the guessing stops, because the answer is there before the work starts. Everything free on this site is a worked example of exactly that.
How we do it
Capture. Encode. Enforce.
Three moves, and they are the same three whether the job is a regulatory response or a design system.
Capture
get it out of the team
- Sit with the people who already do this job well, and watch them do it
- Write down what good looks like, in the words your team already uses
- Name the decisions that stay human, so the system never takes them
Encode
put it where the agent reads
- Standing rules loaded at the start of every session, with nothing to remember
- A knowledge base it has to cite from, so it stops inventing
- Skills and scripts for the steps that should never be improvised
Enforce
make it fail loudly
- Tests that fail when a standard slips, in seconds, with no model in the loop
- Checks that run in your pipeline and stop trusting anybody's memory
- The suite is the handover, and it is why the system survives us leaving
What you end up with: a standard that survives the person, an agent that stops guessing, and work that passes review the first time.
What you get
What actually lands in your repository.
Standing rules
The standards your agent reads before it does anything, written in the words your team already uses, living in the file the agent opens on the way in.
loaded every sessionA knowledge base
Your sources of truth, structured so the agent cites them and refuses to assert what it cannot source. Everything still open comes back to you as a list.
no more invented factsSkills
The procedures that should never be improvised, loaded on demand when the work calls for them and ignored when it does not.
the steps, written onceCommitted scripts
Code that produces the deliverable, so nothing is hand exported and nothing drifts between one run and the next.
never hand exportedTests
Checks that fail when a standard slips. Deterministic, seconds to run, no model in the loop. This is the part that stops the whole thing rotting.
the part that stops rotA written scope
What the system will cover, what it will refuse to do, and what stays a human decision. Agreed before anything is built, with the price attached.
specific enough to argue withWhy this is different
Most AI enablement leaves nothing behind.
Here is what usually arrives, and what arrives here.
a deck and a workshop
Slides go stale the week they are delivered, and the enthusiasm goes with them. You get a repository your team is working in on Monday, and the workshop is the build itself.
a platform you rent
Another seat count, another renewal, another vendor sitting between your team and its own work. The system is files in your repository. There is nothing to log into and nothing that stops working when an invoice lapses.
a prompt library
A folder of prompts is a folder somebody has to remember to open, and under deadline nobody does. Standing rules load themselves every session, because they live where the agent already looks before it starts.
a wiki page nobody reads
Tribal knowledge written down once, in a place with no connection to where the work happens. These files sit in the repository the work happens in, so the agent reads them without anyone deciding to go and look.
a pilot that never lands
Six weeks of promise, one impressive demo, and nothing in production. The checks are the handover. They are what tells you six months later that the standards still hold, and they are why it survives without us.
The math
What your status quo already costs.
Leaving things exactly as they are has a price, and you are already paying it. We cannot see your cost base, so here is the shape of that price. Put your own numbers in and the conclusion holds.
One senior reviewer
Re-reading agent output for mistakes a file in the repository could have caught before the work was ever produced.
Two hours a week
Deliberately conservative. Teams tell us it is considerably worse in the weeks that actually matter, which are the weeks with a deadline in them.
Ninety hours a year
Of your most expensive person's attention, spent supplying a standard that could have been a file. Put your own loaded hourly rate against it.
Now multiply
By the number of people doing it, and by the number of jobs in your team that look like this one. That total is the number a system is competing against.
And the one you cannot price
The quarter one of them leaves and takes their share of it with them, or the week all three are busy and the deadline lands anyway.
A system is one payment, and then it keeps working. The bill above lands again every week, in the same senior hours, for as long as nothing changes.
How it runs
From the first call to the day we leave.
Four steps, and your team is in the room for all of them.
Step 01. A conversation, free
What the job is, how often it happens, and what it costs when it goes wrong. This is also where we tell you if a free system already covers it.
Step 02. A written scope, and a price
What the system will cover, what it will refuse to do, and what stays a human decision. Agreed before anything is built, and specific enough to argue with.
Step 03. Built with your team
Your people run it while it is being built, so they can extend it after. A system only we know how to change is one that dies the week we leave.
Step 04. Handover, and then silence
The checks are the handover. They are what tells you six months later that the standards still hold, and they are why you never have to call us again.
What it costs
Fixed scope. One payment. Nothing to renew.
The price is agreed in writing alongside the scope, before a line is built. No retainer, no seat count, no runtime to license, no subscription. When the work is done the repository is yours, under a licence you choose, and it keeps running whether or not we ever speak again. The number depends on the job, which is what the call is for.
Two reasons we would turn you down
The job only happens once. A system earns its cost by being run again and again, so for work you will do a single time you are better off doing it by hand, and we will say so.
A free system here already covers it. Then we send you the link and you have it today for nothing.
Either way we would rather find that out with you on a free call than three weeks into an engagement you are paying for.
FAQ
The questions everyone has.
Why would we not build this ourselves?
Plenty of teams could. The reason they have not is that it is nobody's job, so it loses every week to the thing with a deadline on it. If you have someone who will genuinely own this, take our free systems and copy the structure. We would rather you shipped it than bought it.
How long does it take?
It depends on the job, and you get the estimate in the written scope before a line is built. The scope exists precisely so that nobody is guessing about this, yours included.
Which agent does this work with?
The system is files: rules, skills, a knowledge base, scripts and tests. Anything that reads a repository can use them. The free systems here are built and tested against Claude Code, and the structure does not depend on it.
Do we need to be an AI company?
No. What matters is that you have a job that repeats, that costs something when it goes wrong, and that some of your people already know how to do properly. That knowledge is the source material.
We tried this and it did not stick.
What usually got tried is a prompt library or a wiki page, and both need somebody to remember to open them under deadline. Nobody does. Rules that load themselves need no remembering, and a test that fails when a standard slips will not let it drift quietly for six months.
What do you need from us?
Time with the two or three people who already do this job well, and access to the repositories the system will live in. They are the source material, and most of the engagement is turning what they know into files. If none of them can get an hour a week, tell us on the call and we will say so.
What if we want to stop?
Then you stop, and you keep everything built so far. There is no licence to lose and nothing that phones home. That is the entire point of handing over a repository. You were never renting access to a platform.
Check us before you call us
Read one we already published.
We publish complete systems, free and MIT, with the standing rules the agent loads and the knowledge base it has to cite from all in the open. Clone one, read every standard it holds, and judge the craft before you spend a minute on a call. It is the same work we do for your job, in domains that happen to be public.
Get in touch
Your status quo bills you every week. This bills you once.
One paragraph is plenty: what the job is, how often your team does it, and what it costs when it goes wrong. Or write to [email protected] directly.
We take a small number of builds at a time, because step three means sitting with your people and that does not parallelise. If your timing is tight, put the date in the first line and we will tell you straight away whether we can meet it.