issue

Issue

An issue is a prompt with context. Its reader is a different agent, in a fresh session, with no access to this conversation. The issue proves that the problem exists and where; it does not prescribe the solution. The reader plans.

Input

$ARGUMENTS

• --dry-run: do everything but the publish step; print the title, the labels and the body instead.
• Starts with a number: rewrite that issue. Fetch it with gh issue view <n> --json title,body,labels,url, keep the facts that still hold, drop the rest.
• Anything else: the request for a new issue.
• Empty: ask one question, "What is wrong, or what do you want?", then continue.

Ask nothing else unless the request cannot be located in the code at all.

Protocol

1. Project constraints. Read the repo's AGENTS.md or CLAUDE.md for two things only: the validation commands, and the boundaries the reader must respect. They go in Done when and Hints.

2. Evidence. Locate the code the request touches. Read it before citing it. Each fact gets an anchor the reader can jump to: path:line plus the symbol name, or a commit hash. Run git log --oneline -10 -- <file> when history explains the current state. Search once for a duplicate: gh issue list --search "<keywords>" --state all --limit 10; link a match in Evidence and say so to the user.

3. Body. Fill the template below. Size follows the problem: a small bug fits in fifteen lines.

4. Publish. No question before it: the skill is the guard, and a caller that wants a look first passes --dry-run. Labels only from gh label list. Write the body to a temporary file outside the repo, then:

   gh issue create --title "<title>" --body-file <tmp> --label "<a>,<b>"
   gh issue edit <n> --body-file <tmp>
   

   --body-file keeps the markdown intact; inline --body breaks on backticks. Report the title and the URL.

Template

## Problem

## Evidence

## Hints

## Done when

## Out of scope

Problem. What is observed and what is wanted. No cause, no solution.