Feedback loops

Coding agents stop working once their task appears to be ‘complete’. Only when the coding agents receive confirmation that your test suite, yout linter, etc. have run without errors is the feedback loop complete and the task finished.

Python helps you with this by providing excellent error messages that can be fed straight back to the coding agent; the more guidance – or even suggested solutions – they contain, the better.

See also

Once the check exists, you should specify how strictly it controls the stop:

In a single prompt

Ask the coding agent to perform the check and iterate within the same message. This currently works for every task.

Across a session

In Claude Code, you can also set the check as a /goal condition. A separate evaluator checks it again after every move, and Claude continues working until it is fulfilled. This ensures that an unattended run is completed correctly even without your intervention.

As a deterministic criterion

A stop hook runs your test as a script and prevents the step from ending before it has passed.

Have the coding agent provide evidence rather than simply claiming success; this could be the test output or a screenshot of the result. Checking the evidence is quicker than running the verification itself again, and also works for sessions you haven’t observed.

Through a second opinion

A verification sub-agent or a dynamic workflow that checks its own results allows a new model to attempt to refute the result, so that the agent performing the work is not the one evaluating it.

Example

## Documentation review process

1. Write your documentation in accordance with the guidelines in `DOCSTRINGS.md`
2. Check the documentation against the checklist:
   - Ensure consistency in terminology
   - Make sure the examples follow the standard format
   - Ensure that all required sections are present
3. If any issues are identified:
   - Note down each issue with a specific reference to the relevant section
   - Revise the documentation
   - Review the checklist again
4. Only proceed once all requirements have been met
5. Finalise the document and save it

Review and Revise

As with test-driven development, you should first review the results of your skill. This is the only way to ensure that your skill solves real problems rather than just imaginary ones:

  1. Identify gaps

    Have your coding agent perform a representative task without the skill and document the specific error or missing context

  2. Create Tests

    Develop several scenarios that test these gaps

  3. Establish a baseline

    Measure the coding agent’s performance without the skill

  4. Write minimal instructions

    The instructions should be as short as possible to close the gaps and pass the tests

  5. Iterate

    Run tests, compare them with the baseline and refine the skill

  6. Observe how your coding agent navigates through the skills

    When refining the skills, pay attention to how your coding agent actually uses them in practice. Pay particular attention to the following:

    Unexpected exploration paths

    Are your files being read in the order you expected?

    Overlooked links

    Does your coding agent follow the references you have specified?

    Excessive attention

    If the same file is being read repeatedly, you might want to move its contents to the SKILL.md file.

    Ignored content

    If a nested file is never accessed, it is often unimportant or its significance is not sufficiently indicated.