2. Ask your assistant
The terminal review only covers the rules a checker can decide. Your assistant covers the rest, using the same rules, and reports in the same format.
Keep src/user.ts staged from the previous step. Onboarding added a /redline-review command to every assistant it rendered for. It has the assistant run Redline's review with --engine embedded, which collects the rules that apply to the changed files, the diff and the checker's findings into one prompt, and then follow that prompt. If the assistant cannot run it, the command tells it to collect the same rules from AGENTS.md or CLAUDE.md and read the diff itself, so the review works either way.
In Claude Code, Copilot Chat or Cursor
Type this in the chat, in the repository:
> /redline-review
In Copilot Chat it comes from .github/prompts/redline-review.prompt.md and runs in agent mode. In Claude Code it comes from .claude/commands/, and in Cursor from .cursor/commands/.
The reply is a list of findings in the same Redline/SEVERITY [rule-id]: problem shape the terminal printed, BLOCKER first, ending with a count such as 2 BLOCKER · 2 HIGH · 0 SUGGESTION. If nothing qualifies, the reply is No Redline issues in this change. and nothing else. Expect the four findings from step 1. A model may add findings for rules the checker cannot decide, and its wording changes from run to run; the rule ids do not.
In any other chat
A browser tab works too. This copies the whole prompt to the clipboard; paste it into the chat. On Linux, pipe to xclip -selection clipboard instead of pbcopy, and on Windows to clip.
$ npx redlinegate@latest review --print-prompt | pbcopy
The assistant follows the rules while it writes
Reviewing is half of it. AGENTS.md, CLAUDE.md, .github/copilot-instructions.md and .cursor/rules/ are read by the assistant on every request, so the rules shape the code before anything is reviewed. Ask for new code:
> Write a function in src/account.ts that parses an Account { id: string; email: string } from a JSON string.
Compare its answer with src/user.ts. An assistant that has read the rules validates the parsed value, usually with a type guard or a schema, instead of writing as unknown as Account. This depends on the model, so treat it as something to look for, not a guarantee. That is why the review exists.
More on what each assistant reads: Claude, GitHub Copilot, Cursor.