Recorded reasoning explains why work should take a particular direction. It is a deliberate record of questions, evidence, decisions, and uncertainty, not a transcript of a model’s private thinking.
Use a connected agent in the project. Ask it to read existing decisions before investigating the same question again.
Ask an answerable question
State the decision and what would change it. For example, if an application already imports customer records, ask whether it needs to stream large files instead of reading them entirely into memory.
Inspect the importer and measure representative file sizes. Record the evidence, the memory limit we need to meet, and a recommendation. Do not change the importer yet.
This prompt assumes an existing importer and an agreed memory limit. Your own question should identify its assumptions just as clearly.
Record evidence and its limits
A useful finding includes the question, relevant context, observations, sources, conclusion, rationale, and next action. Link source files, documentation, or executed checks when they support the answer.
Distinguish a measurement from an estimate. A benchmark on a small fixture does not establish production performance. A document describing a library feature does not establish that your code uses it correctly.
Magistr saves findings only when the agent records them through its tools. Ask it to record consequential decisions during the work, not merely summarize them at the end of the conversation.
Follow a separate investigation
A branch gives a real question its own investigation, linked to the earlier finding that prompted it. Continue findings on that branch. When the result informs the main decision, record a conclusion that links the supporting steps and mark the investigation merged.
An investigation that stops without contributing a result can be marked abandoned with an explanation. A still-open question should remain open with a concrete next action.
You do not need a branch for every task. “Implement, test, commit” is an execution sequence, not an investigation.
Retrieve or revise a decision
Ask the agent to find the saved reasoning for the affected feature and explain which evidence still applies. Important conclusions can be pinned for retrieval. A reasoning checkpoint helps review open questions and decisions affected by revisions; it does not copy your source files.
When new evidence changes the answer, record a revision and why it changed. Keep the earlier finding as history. Update affected roadmap requirements too, so the next implementation does not follow a superseded decision.
Success means another session can retrieve the question, evidence, decision, and remaining uncertainty without relying on the original chat.
Connect reasoning to verification
A recorded explanation is not proof that a command ran. Execute a test, experiment, or benchmark through the verification workflow when the claim depends on its result. Verification explains how those results are recorded and when they become stale.