Magistr calls saved feedback a signal. It can be a bug report, a question, an idea, or an observation from trying the product. Saving a signal lets you keep it without interrupting the task currently in progress.
Use a connected agent in the relevant project. You do not need a complete solution before recording feedback.
Capture the observation
Tell the agent to save the feedback as a signal. Include the action, expected result, actual result, environment, and evidence you have.
For example, for an application that uploads documents:
Save this feedback: uploading a PDF on a narrow screen hides the Cancel button. I expected to be able to stop the upload. Investigate later; keep the current task focused.
Attach or link the relevant screenshot or reproduction steps through your agent. Avoid putting credentials or private user data in material that may be shared.
Investigate the signal
Use $signals in Codex or /signals in Claude Code, and name the feedback
to investigate. The skill reads the report, checks current code, researches
uncertain claims, and records evidence and a recommendation.
This is investigation, not authorization to implement a feature. A useful result distinguishes a reproduced defect, an unsupported assumption, and a product choice that still needs a decision.
Ask the agent to link the findings to the signal so another session can retrieve them. A suggestion made only in chat is not necessarily saved.
Decide what happens next
| Outcome | What to keep |
|---|---|
| Ready to build | A linked roadmap chunk with acceptance criteria, priority, and dependencies. |
| Needs more evidence | The unanswered question and a concrete next investigation. |
| Useful later | The signal and reason for deferring it. |
| Dismissed or already covered | The reason and a link to any existing work. |
A promoted signal should point to the work it informed. Creating a chunk does not mean the implementation has started or completed.
Follow through
Review the proposed chunk and use Working in chunks when you want implementation to begin. After delivery, retry the original scenario and check that the result addresses the observation.
If the behavior is still wrong, record the new evidence on the existing issue or linked work instead of creating an unexplained duplicate.