Coding projects

From idea to a result you can trust.

AI coding works best as a loop of orientation, decision, implementation, verification, and feedback. The loop can be fast; it should not be invisible.

The project loop

01 · MapRead the project guidance, status, recent history, relevant docs, entry point, consumers, tests, and current behavior.
02 · FrameWrite the intended outcome, acceptance criteria, constraints, non-goals, authorization boundary, and likely counterexamples.
03 · ChooseReuse an existing primitive or skill when it fits. Select the narrowest useful workflow; do not create a new abstraction to avoid understanding the old one.
04 · ChangeImplement a small complete slice. Keep changes reversible and preserve unrelated worktree edits.
05 · ProveRun checks matched to the changed surface: tests, type/lint/build, migration status, browser behavior, performance measurement, or artifact rendering.
06 · ReviewInspect the final diff and report what is done, what is not, and what evidence supports the claim.
Before code

Build a compact evidence map

Current behavior · owners · invariants · data boundary · permission boundary · reuse candidates · counterexamples · checks · non-goals.

After code

Use a completion gate

Focused test · shared check where justified · real runtime or rendered artifact when relevant · final diff · no accidental secrets or unrelated changes.

Match the check to the claim

If you claim…You need evidence such as…
“The function is correct.”Meaningful unit or integration tests, including a boundary or counterexample.
“The user can do it.”Real browser/runtime interaction, with focus, viewport, loading, error, and permission state checked as relevant.
“The migration is safe.”Schema/migration checks, existing data considerations, rollback or deploy behavior, and affected application tests.
“The system is faster.”Before/after measurement with the same scenario, not only a lower query count or an intuition.
“The asset matches.”Inspection of the actual delivered file and comparison with the requested style, dimensions, transparency, or project reference.
“The research is current.”Opened authoritative sources, dates, citations, and a clear separation between fact, inference, and uncertainty.

When to split the work

Split when the task contains independent questions, separate risk boundaries, or a large amount of context. Keep the parent goal visible, give each subtask a narrow return format, and reconcile results before changing the product. Parallelism is useful only when it improves speed or confidence.

Useful “continue” instruction

When you want autonomous follow-through, say:

Continue until the requested outcome is complete and verified. Use reasonable assumptions when the path is clear. Stop and ask only when the missing choice would materially change scope, risk, or the product result. Report evidence and remaining blockers honestly.