Goal Prompt

Trail of Bits' skill for goal mode. Goal mode keeps an agent working until a completion condition is met, so a vague condition can keep it looping turn after turn. Goal Prompt drafts that condition as a contract the agent can actually finish, then formats it as a single /goal line for Claude Code or Codex. What it does - Writes the goal as an end state, not an activity: what to read first, the exact command that proves it's done, what must not change, and a stop bound - Closes the easy outs, such as deleting code to silence a check, passing on a subset of tests or editing the checks themselves - Keeps the goal small, and moves long detail into a GOAL.md file that the goal points to - Never invents a missing piece: anything it can't ground in your request or your repo comes back as a "Missing:" list - Narrows security-audit goals to one validated finding at a time, under a threat-model file, following Trail of Bits' codex-config guidance How it works - Ask it to write, rewrite, tighten or clean up a goal prompt, and it returns one fenced line: /goal followed by the objective - A small Python formatter (standard library only, run with uv) puts it on one line, warns when there's no stop clause and rejects anything over the 4,000-character limit instead of cutting it short How to install it In Claude Code, run /plugin marketplace add trailofbits/skills, then install goal-prompt from /plugin menu. In Codex, run codex plugin marketplace add trailofbits/skills, then codex plugin add goal-prompt@trailofbits. Good to know - It's part of Trail of Bits' open-source skills marketplace on GitHub. - It only writes the goal. You paste the /goal line into your agent yourself.