Agent instructions tricks - Definition of Done

A trick that I learned doing some agentic coding is to have agent instructions (agents.md etc) contain a section - preferably by the end - as a checklist of what needs to be met to consider a task done. It doesn’t replace hooks and other hard validation, but helps a lot in keeping harness tight.

DOD steps seem to work best when made declarative, and as a checklist. Example:

1. All unit tests and integration tests pass.
2. All architectural decisions are documented in /docs/.
3. Any new decisions have immutable record in /docs/adr/.
4. Everything is commited, commit history is clean and linear.
5. Ticket status and TODO tracking is up-to-date with changes.
6. Branch can be merged into main without conflicts, no rebase necessary.

Shorter is better, since this is a sequence model will likely run step by step each time it’s “done” with the work - you can have separate “handoff” and “task completed” DOD and trigger second one explicitly by commanding model to wrap up the task.

Fun fact: you can also use AI to put harness on other devs and check their merge requests against provided definition of done, LLM can run same checks against someone elses code.