Whoever has just taken a decision is the worst person to challenge it: they built the reasons why it is right. When correctness matters more than speed (production, security, irreversible operations), it pays to have the decision examined by someone who does not know the path that led there.
The cycle
1. The thesis. You write in two or three lines what you take to be true, and why it matters: "the new cache layer is safe under concurrent read loads; a race condition here would corrupt data and does not show in tests". If you cannot write it in a few lines, it is not a decision, it is an impression.
2. The extraction. You isolate the smallest examinable unit: the change or the function, not the whole file; the proposal in three to five sentences with the constraints it must respect. You hand over the artefact and the contract, not the reasoning. Handing over the conclusions means getting a confirmation of the conclusions in return. A request of five hundred lines must first be broken down.
3. The doubt. A reviewer with fresh context (another person, a second agent started from zero) receives the artefact and the contract, not the thesis, and an explicitly adversarial brief: find what is wrong, assume the author is overconfident, look for undeclared assumptions, unhandled edge cases, hidden couplings, ways to violate the contract, existing conventions that get broken, ways to fail with unexpected input. Do not confirm, do not summarise: find problems, or state that none were found after a careful examination. A question like "is it fine?" produces a yes.
4. The reconciliation. Each finding is classified against the text of the artefact: actionable (fix it), already covered (say where), not relevant (explain). You do not take the reviewer's output as truth: you reread it against the artefact.
5. The stop. The cycle is bounded, not recursive. You stop when the findings become trivial, after three cycles, or when the user decides. Beyond three cycles without a result, you take the question to the user.
When it pays
When correctness weighs more than speed, when working in code you do not know, when the stakes are high (production, security logic, irreversible operations) or when the decision has so far been validated only by whoever took it.
When it is not needed
For renaming a variable or reformatting. The method has a cost, and for a one-line change it is disproportionate.
Warning signs
- Starting a reviewer for a trivial change.
- Treating the reviewer's output as authoritative without rereading the artefact.
- Doing more than three cycles without involving the user.
- Asking "is it fine?" instead of "find the problems".
- Skipping the doubt under time pressure on a high-stakes decision.
- Rerunning the reviewer on an unchanged artefact: same findings, you are only buying time.
- Doubt theatre: in two or more cycles with substantive findings, none was classified as actionable. You are validating the thesis instead of testing it.
An example
A thesis: "the migration can be rerun with no side effects". The script and the idempotency rule are extracted. The reviewer, who does not know the thesis, notices that the script inserts a log row before checking whether the migration has already been applied: rerun, it duplicates the logs and, worse, resets a counter. The defect could not be seen by whoever wrote the script, who was convinced of its idempotency.
This method derives from the open-source agent-skills family by Addy Osmani (MIT licence), rewritten and adapted.
This article is an OIAR skill
What you have just read is the content of a skill from the OIAR catalogue, written for people. Knowledge like this, but richer and deeper, written for the agents, drives the interviews, the analysis and the report inside the platform. Skills are used in OIAR: to try this one on a project of yours you need an account.
doubt-driven-developmentA five-step method for putting an important decision before an adversarial reviewer with fresh context: thesis, extracting the artefact and contract, doubt, reconciling the findings, stopping. Use it when correctness matters more than speed or the stakes are high (production, security, irreversible operations).