Skip to content
All guides

Fixed isn’t always verified. Record both.

For developers: separate code changes from behavior checks so the next AI has evidence.

A coding AI can change a file and still leave an important behavior untested. When another AI picks up the work—or you check it from your phone—the record should distinguish the fix from evidence that it works.

Keep the outcome and verification separate

For an upload bug, create two todos in the same list:

  • Fix the upload error. Record what changed and a commit reference when available.
  • Verify file upload. Describe the behavior that needs checking and leave it open until tested.

The split is useful when the two states can differ. You don't need to turn every small edit into a large testing checklist.

Ask your connected coding AI to record the boundary

With MCP connected, tell it:

The upload fix is implemented in the current commit. Mark the code-change item done and note the commit. Leave upload verification open: I haven't tried sending a real file yet.

Check that only the implementation item is completed. “The code looks right” should not silently turn into “an upload succeeded.”

Add evidence after checking the behavior

For example:

I uploaded a 5 MB PDF in Chrome while signed in. It completed successfully and appeared in the attachment list. Add that result to the existing upload-verification item's description. Keep that item open: image files and mobile uploads weren't tested.

If the original verification item includes those other cases, keep it open or split out the remaining work. A successful PDF test is evidence for that case, not every supported format and platform.

Continue from another AI or your phone

Ask the next connected AI:

Read the upload work in thrw. Which result is verified, under what conditions, and what remains untested? Use the recorded notes.

It should recover the commit reference, the tested file and browser, and the remaining cases. If you open thrw on your phone, the same descriptions give you that boundary without reopening the coding chat.

When a later check fails, add the observed result and next action instead of erasing the earlier successful case. The useful handoff is a record of what happened and what to do next—not a single reassuring “done.”

See why a fix and a verified result deserve separate records.