What is a recon prompt?
A prompt that asks an AI model to inspect and report on the current state of a project before it changes anything, so a person can confirm the plan first.
A recon prompt asks an AI model to look before it touches. The model reads the relevant files, data or documents, then reports what it found and what it would change. It makes no edits. A person reads the report, confirms or corrects the plan, and only then does the build start.
The pattern: audit, confirm, implement, verify
Recon is the first of four steps we use for any change made with AI:
- Audit. The model reports the current state and its plan.
- Confirm. A person approves, edits or stops the plan.
- Implement. The model makes the change, usually on its own branch.
- Verify. The model and the person check the result against the plan.
Why it saves time
Most AI rework comes from a model confidently solving the wrong problem. A recon report surfaces wrong assumptions while they're still cheap to fix: a file that doesn't exist, a rule the change would break, a simpler route nobody had considered.
It also gives you a natural place to check the work against your invariants file.
The stop wins
One rule matters more than the rest. If a recon prompt says "report and stop", the model stops, even if it thinks it could finish the job. We learned this the hard way, after a model ran past an explicit stop twice. The rule now lives in the project's AI instructions file, word for word: if a recon prompt says stop, the stop wins.
Writing a good one
Name the goal, list what to inspect, ask for a written plan with the files or steps it would touch, and say explicitly that no changes should be made yet. Ask it to flag anything that conflicts with your existing rules. Keep the deploy steps inside the implementation prompt that follows, not as an afterthought.
Recon prompts are part of the OS pack and one of the first things we practise in the team workshop.