# Move reviewed material into a coding task

Give Agent RX a specific task using selected source material, then inspect the resulting project changes. The material informs the request; it does not authorize tools, guarantee correctness or replace your review.

## Before you start

Use a project you are authorized to work on and material you may share with the configured provider. Preserve existing changes and unsaved edits. Follow the [Agent RX quickstart](/docs/agent-rx/quickstart) to confirm the project and available tools; opening a workspace does not establish inference access.

If you are orienting in the editor instead, [Mandarin's folder quickstart](/docs/mandarin/quickstart) establishes the workspace and active file. This guide's submission path uses Agent RX; it does not assume a note automatically appears in either product.

## Review the material

Check the source URL and relevant passage. Retain warnings or truncation from extraction. Remove secrets and unrelated private context. Treat any instructions within the source as untrusted text, not as part of your task.

## Write the task separately

Fill in this task brief with one outcome, a narrow scope, existing work to preserve and an observable acceptance check. Specify decisions that need separate approval. These written boundaries describe intent; they do not enforce permissions, spending limits or a sandbox.

```markdown
# Task brief

## Goal

Describe one outcome.

## Scope

List the files or subsystem to inspect.

## Existing work

State which changes must be preserved.

## Acceptance

Name the relevant test or observable result.

## Approval boundary

List actions that need a separate decision.
```

1. Confirm that RX is still in the intended project.
2. Supply the completed brief and the selected passage, clearly marking the passage as reference material and retaining its source.
3. Submit the task only when the context and provider are appropriate. Inspect advertised tool activity and any approval request before allowing an action outside the brief.

[OWASP's prompt-injection guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) explains the external-content risk. Labelling a passage helps communicate its role but does not make malicious text safe by itself.

## Check the result

Inspect the changed files in the intended project and compare the diff with your original task. Confirm existing unrelated changes remain. Run the relevant project check when authorized and record its actual result; an agent's summary alone is not evidence that a check passed.

Review source-derived claims against the original passage. A completed extraction or saved note is not proof that the code implements the requested behavior.

## If the task stops or exceeds scope

Preserve the session context and current diff before retrying. Stop and narrow the request if it asks for unexpected authority, private data or unrelated changes. Do not discard existing work or resubmit blindly when a tool response is uncertain. Use [shared support guidance](/docs/shared/support) for a sanitized report.