portswigger/agent-wrangler/brainstorm

Clarifying the 'what' (for the issue → PR autopilot)

This is the front half of the brainstorm skill, trimmed for autopilot use. You read it when an issue doesn't say what to build. Your only job here is to make the 'what' concrete enough to build responsibly — then stop and let the autopilot's own plan → build → verify → PR arc take over.

You are not designing the solution and not entering plan mode. The autopilot decides the how (approach, structure, naming, libraries) autonomously; you only pin down the what (intent, scope, success criteria) with the human.

When to use this

Only when you genuinely can't tell what to build from the issue. A Jira/GitHub issue with a clear summary and acceptance criteria does not need this — proceed. Reach for it for bare, underspecified free-text ("improve the dashboard", "make login nicer") where guessing would risk a whole run on the wrong thing.

The flow

1. Explore project context first. Read the relevant files, docs, and recent

commits so your questions are informed and you never ask what the repo already answers.

1. Check scope. If the request is really several independent pieces (e.g. "add

chat, billing, and analytics"), say so and agree with the human which one piece this run delivers — one issue → one PR. Don't try to squeeze a mega-issue into a single run.

1. Ask clarifying questions — one at a time. Prefer multiple-choice; open-ended

is fine. One question per message. Focus on purpose, constraints, and success criteria: what problem this solves, what "done" looks like, what's explicitly out of scope. Keep going until you could write a one-line spec of the change.

• These questions are how the autopilot stops for the human: each one puts the

card in needs-you and the run waits. Keep your phase chip on plan.

• Ask only what you can't responsibly assume. Don't ask about the how

(architecture, file layout, which library, naming) — pick sensible defaults for those and keep moving; the autopilot is meant to be autonomous.

1. Confirm the 'what'. Once it's clear, play it back in a sentence or two — what

you'll build and how you'll know it's right — and get a yes (or a correction).

1. Hand back. With the 'what' confirmed, drop this guide and continue the

autopilot arc from branch discipline (§2 of the parent skill) onward. No design write-up, no plan mode, no review loop, no /clear.

Principles

• One question at a time — don't overwhelm with a wall of questions.
• Multiple choice preferred — easier to answer than open-ended.
• YAGNI — clarify the change that's actually needed, not a bigger system.
• Ask only what you can't assume — default everything else and keep moving.
• Be flexible — if an answer reframes the task, adjust and re-ask as needed.