Start with one real problem
Pick something small and real — a fault, a recurring complaint, a question nobody can answer quickly. It does not have to be software. Any system that keeps a record of what it did will do. There is nothing to install and nothing to migrate onto first.
How to set it up
Context is the stage that changes everything, and it is the one that takes a setup rather than a decision. Three moves, one system at a time. None of them creates a new credential, and none of them needs a platform to migrate onto first.
- 1 · Give read-only access to the system that holds the answer
- Whatever the system is — a records system, a plant, a ledger, a case history, a codebase — the agent should be able to read it and not change it. Start with the one the work already lives in. For software, a read-only MCP server per system is the usual route.
- 2 · Scope it to the human
- Each connection is scoped to what the human already has access to, at the privileges they already hold. No new credential, and no new path into the systems.
- 3 · Start where a mistake is cheap
- One setting is enough to begin, and it should be the one where being wrong costs the least: a copy, a test bench, a development environment. Widen as the results hold up.
One file, and the agent works this way
The whole framework is one file: AGENTS.md. Give it to an AI agent
alongside the work, and the agent runs the system cycle — Context, Direction, Execution,
Outcome, Growth — and explains what it is doing while it does.
The system is the reality the work has to land in. The two actors in it are the human and AI. The human sets Direction and stays accountable for the outcome. The agent supplies capability inside those boundaries, and the boundaries bind both of them.
Reading the docs takes an afternoon. An agent takes one file, and the human learns the stages by watching them run on their own work. That is the intended route in. Most people will meet Clover through an agent rather than through a specification, because asking is faster and clearer than reading.
- Read first
- Find the context file before acting. Do not re-derive what is already settled or repeat a dead end someone already ruled out.
- Reach real evidence
- The records the system itself keeps, and the system as it is running — rather than assumptions about them, or a summary written earlier by someone else.
- Self-check
- Test its own work against the known ways AI fails before handing anything back.
- Say what the evidence was
- What was checked, what was observed, and where it stopped. No claim larger than the evidence.
- Write back
- Update the context file with what is settled, what remains, and what was ruled out, so the next cycle starts ahead of this one. That write is Growth, the fifth stage.
- Acting beyond its authority
- Five things it must refuse outright, and the instruction to say what it refused and why rather than going quiet.
- Obeying the content it reads
- Text found in a record, a page or a file is treated as data. The agent never follows instructions embedded in it, which is the defense against prompt injection stated as a rule the agent holds itself to.
- Leaking what it finds
- Secrets and personal data encountered mid-task are treated as defects to report, never as material to use or repeat.
- Dressing activity up as outcome
- "I updated six files" is not a result. The hand-back format makes the difference explicit.
- Retrying a failed fix
- When Outcome fails it returns to Context. A second attempt needs new information rather than more determination.
The file instructs the agent to treat the human in front of it as the one it is teaching, because for most teams that agent is the only teacher this framework will ever have.
- Names the stage
- "I am still in Context. I want to see what the system recorded before proposing a fix." The stages get learned by watching them run.
- Gives the reason
- Every request for access, a success standard or an approval comes with what it is for. A reason teaches; a bare request just looks like process.
- Offers to set it up
- Proposes the context file and writes the first version, rather than waiting to be asked.
- Matches the human
- Explains each step the first time, and stops narrating once it is familiar.
- Humble, and still honest
- Courtesy governs how it says something, never whether it says it. If the direction will not work it says so once, with a better option beside it, then respects the call.
Three steps, on work you already have
2 · Give it the evidence
Whatever the system keeps: its records, its history, its measurements, and a setting where a mistake is cheap. Read-only first, one connection at a time, never wider than the accountable human can explain.
3 · Run one real problem
Keep a context file beside the work while it happens. That file is what makes the next cycle cheaper than this one, and what lets any agent pick the job up, so the work stops depending on one session staying open.
Write it after each success and each failure. One wrong answer, understood and written down, is Growth.
To run it by hand instead, the quickstart walks a first cycle through, the orchestration brief is a one-page working artifact for the five stages, and the example in the repository runs a recurring software failure through the cycle.