Get a coding agent to do it right the first time
Coding agents do what the brief says and guess the rest: which files, when it is done, what must not change. Paste the task you are about to hand to Claude Code, Codex, Cursor or Copilot, and see what it is missing and what is risky. Get a better brief, and a starter file for the facts every task needs.
Read how to write a task for Claude Code, Codex or Cursor, and what to put in AGENTS.md and CLAUDE.md.
New to this? What it does, in plain words
A coding agent is an AI that edits code on its own. The brief is all it knows about your task.
"Fix the checkout bug" leaves it to find the bug, decide what fixed means and stop when it feels done. A brief that says how to reproduce the bug, where the code lives, what must not change and which command proves it works gets a change you can merge.
What you need
Nothing. No API key and no sign-up. When you press Check, the brief goes to thinkhard.si, is read and is forgotten.
How to use it
- Paste the task, as you would type it to the agent.
- Pick the agent, and add standing facts about the repository if you like: commands, folders, rules.
- Press Check the brief and fill the marked gaps in the better brief.
- Save the starter file in the repository so every task gets those facts.
The brief
About the repository
OptionalWhat every brief for an agent needs
An agent cannot ask you what you meant halfway through a task the way a colleague would, unless you tell it when to. So the brief carries what a colleague would ask about first: where to look, what done looks like, and what not to touch.
- You. Paste the briefAs you would type it to the agent.
- thinkhard.si engine. Know the taskA bug, a feature, a refactor, tests, an upgrade, an investigation.Instant
- thinkhard.si engine. Check its partsTwelve parts, weighed for that kind of task.
- thinkhard.si engine. Flag the risksSteps that cannot be undone, keys, vague goals, several jobs at once.
- The agent. Hand it overA better brief, and the standing facts file.
The parts, weighed by the kind of task
- A bug fix needs a way to reproduce it, where the code is, the command that proves it is fixed and a test that fails before the fix.
- A feature needs the outcome, where it goes, a done-when list, the check, tests, and what must not change.
- A refactor needs the promise that behavior does not change, the folders it covers, the tests that must stay green and what is out of scope.
- An investigation needs the question, the context, what to hand back and the rule that nothing is changed.
- An upgrade needs the check, what must keep working and when to stop and ask.
Standing facts go in one file
The commands, folders and rules that every task needs belong in a file in the repository that the agent reads at the start: CLAUDE.md for Claude Code, and AGENTS.md, a shared format that many other coding agents read. Then each brief only has to say what is new.
Questions
Does it work for every coding agent?
Yes. The parts of a good brief are the same for any agent that edits code on its own. The rewritten brief uses plain sections every agent reads, and the starter file is named for the agent you pick.
Is my brief stored?
No. When you press Check the brief, the brief and any repository facts go to thinkhard.si, are read and are forgotten. Nothing is written to disk or logs. Keys in a brief are flagged, never repeated. See the privacy page.
Why does it warn about dropping a table?
Anything that cannot be undone, such as deleting data, dropping a table or force-pushing, should come with a rule to stop and show you first. An agent that misreads the brief will otherwise do it, and an undo button does not reach a production database.
Can my agent check its own brief?
Yes. Agents that use the thinkhard.si MCP server get it as the check_agent_brief tool, so an agent can check a task before it starts, and POST /api/v1/agent-brief returns the same result as JSON.