How we work
Short engagements. Real examples. Files you keep.
A first engagement, Audit into Foundry, looks like this. Every week ends with something you can read, run, or reject.
Week by week.
-
Week 0
A thirty-minute call.We ask what your team does by hand that it shouldn't, and what they refuse to do even when asked. If an audit does not fit, we say so on the call and you lose half an hour.
-
Week 1
Interviews and examples.Two conversations with the people who do the work, not the people who manage it. Ten real examples of each task. We run one task through an agent against those examples, blind, and put the results next to what your people produced.
-
Week 2
The map and the readout.The task map with verdicts, the constraints file, the trial outputs, and forty-five minutes with whoever decides. You choose whether to continue. Everything from the audit is yours either way.
-
Weeks 3 to 6
Foundry, one skill a week.Each week: a skill drafted from your expert's rules, run against its eval set, corrected with the expert in the room, installed in your agent product. You see every output it produced, including the wrong ones.
-
Week 7
Handover.Your team changes a rule in a skill, re-runs the evals, and ships the change while we watch. If they can't, we are not done, and the week is on us.
-
After
Care, or nothing.A monthly option: on every model release we re-run your evals and fix what broke, and send you the digest. Or you run the scripts yourself. Review and Readable are scoped separately when you want them.
Six things we hold to.
- You own it.Skills, evals, policies and offer files live in your repository, under your name, in open formats: Markdown, JSON, plain scripts. If Tacit disappears tomorrow, nothing you have stops working.
- Real examples or nothing.We never write a rule from a description of the work. Every rule comes from a case your expert judged, and the case stays in the eval set as the proof.
- The eval is the deliverable.A skill without a test is a prompt. Every skill ships with twenty cases and a score, so "it works" is a number you can re-run, not a feeling.
- Humans keep the decisions.The review policy says what the agent may decide on its own. The default is less than you expect, and it widens as the digest shows it can.
- Numbers only when true.We do not quote efficiency gains or hours saved. We show you the outputs side by side and let you count.
- Small, and in order.Audit before Foundry, Foundry before Review. Each stage is short enough to stop after, and you keep everything when you do.
What we build with.
- ClaudeCowork and Code for the agents your team runs; the API where a skill needs to run inside your own tools.
- Codex and other coding agentsWhere your team already has one. Skills are portable; we meet the agent you use.
- The agent skills formatA folder with a SKILL.md, references and scripts. Open, readable, versioned. The same format this studio's own tools are built on.
- MCPFor connecting an agent to your inbox, calendar, ticketing or booking system without giving it the keys to everything.
- Markdown and GitEverything we hand over is text in a repository you control. No dashboard you have to log into to see your own rules.
- Your agent productIf you have one, we install into it. If you don't, we recommend one for your size and say why.
What "done" means.
Done is not "the agent works". Done is: your team can change a rule without calling us, and knows within ten minutes whether the change broke anything. We test that in the last week, with your people at the keyboard and us watching. If it does not hold, we stay until it does.
That is also why there is no lock-in to sell you. The thing that keeps you is that it keeps working.