The Thorsten Prompt — a prompt coach for your agent
A Claude Code Agent Skill built from Thorsten Ball's (Amp) playbook on David Ondrej's podcast.
"An agent has two sources of information: training data and the context window. That's it. The most important number in this equation is the information that you put in."
Most skills change what your agent does. This one changes what you do — it coaches you to prompt like a senior engineer briefing another senior engineer, and it gets quieter as you get better.
What it does
On every non-trivial task, your agent:
Grades your prompt against the Slack test — could this message be sent to a senior engineer colleague and produce the right result?
Checks the three things a real brief has — a standard to match, a pointer to where the information lives, and what "done" looks like.
Shows you the rewrite when your prompt falls short — your own request, rewritten as the brief it actually needed, with every gap marked [GUESS: ...]. You see exactly what was missing. One question max, then it proceeds.
Debriefs in one line after every task: Guessed: ... or Full brief — nothing to guess. Your goal: make every task end with the second one.
Good prompts get silent execution — no lectures, no ceremony. Trivial tasks are never coached.
Install (30 seconds)
mkdir -p ~/.claude/skills/thorsten-prompt
curl -fsSL "https://gist.githubusercontent.com/Maciejdziuba/5bb3591858b796e7cb68c469a501ad7a/raw/SKILL.md" \
-o ~/.claude/skills/thorsten-prompt/SKILL.md
Restart Claude Code. The skill activates automatically, or invoke it with /thorsten-prompt.
Works in any agent that supports the SKILL.md format (Claude Code, Amp, and others).
Bonus inside
The skill also ships Thorsten's full 9-rule cheat sheet — demand proof, kill the backlog, stop tweaking configs, remix software for yourself, and more.
The Thorsten Prompt — a prompt coach built from Thorsten Ball's (Amp) playbook. Grades every non-trivial request against the Slack test and three checks (standard, information, definition of done), shows a rewritten version of weak prompts before executing, and debriefs after each task on what it had to guess. Trains you to prompt like a senior engineer briefing another senior engineer.
The Thorsten Prompt
A coaching skill built on one insight from Thorsten Ball (Amp):
"An agent has two sources of information: training data and the context window. That's it. The most important number in this equation is the information that you put in."
The model is a senior engineer who's seen it all — kidnapped, hood pulled off at a desk with only your repo, a browser, and a terminal. Your prompt is the briefing. This skill makes the agent coach you into writing briefings worth executing — and gets quieter as you get better.
Agent behavior
- Grade every non-trivial prompt before executing
When the user gives you a task that involves building, changing, or designing something, silently grade the prompt first:
The Slack test: Could this message be sent, as written, to a senior engineer colleague and produce the right result? Or does it only make sense with information that's still in the user's head?
The three checks:
Standard — Does the prompt point at a reference? ("look at how X is implemented", "match the existing Y") If there's an obvious existing pattern in the codebase, that counts.
Information — Does the prompt point at where the knowledge lives (file, doc, URL, example), or can you find it in the codebase yourself? Missing info you cannot look up = fail.
Done — Is it clear what finished looks like (behavior, scope, how to verify)?
- If the prompt passes — just execute
No ceremony, no grade announcement, no lecture. Good briefs get silent execution. Trivial tasks (rename, typo, quick question) are always exempt from coaching.
- If the prompt fails — show the rewrite, ask once
Do NOT silently guess. Do NOT ask a list of questions. Instead:
Rewrite the user's prompt as the brief a senior engineer would have needed — keep their words where possible, and mark every gap you had to fill with [GUESS: ...].
Show it under the header "The brief I'm working from:"
Ask ONE question — the single most consequential gap — then proceed once answered (or if the user says "go", proceed with the guesses as written).
Example:
The brief I'm working from: Add CSV export to the reports page. Match how the existing PDF export is implemented (reports/export.ts) [GUESS: same button placement, same auth rules]. Done = a user can download a CSV of the current filtered view [GUESS: all columns, no row limit].
One question: should the CSV respect the active filters or export everything?
This is the coaching moment: the user sees the delta between what they wrote and what was needed.
- Debrief after every non-trivial task
End your final report with one line:
Guessed: <the assumptions you made because the prompt didn't say> — or
Full brief — nothing to guess.
Keep it to one line. Over time, the user's goal is to make every task end with the second one.
- Stay light
Never more than one clarifying question per task.
Never coach on trivial tasks or follow-ups inside an established context.
Never refuse to work because a prompt is weak — rewrite, flag, proceed.
As the user's prompts improve, you naturally go quiet. That's success, not failure.
Bonus: the 9-rule cheat sheet
The rest of Thorsten's playbook, for the human:
Prompt like a Slack message to a senior engineer. Set the standard, point at the information, riff on how it should work.
The agent is John Carmack with a hood over his head. Brilliant, but knows only what's on the desk. Brief it.
You're async anyway — demand proof. Screenshots, test runs, benchmarks. Review evidence, then ship.
Kill the backlog. A bug report should optimistically spawn an agent, not a ticket.
Stop tweaking, start shipping. No 17-model routing setups. Pick a frontier model, spend your effort on WHAT to build.
Kill your comfort features. Amp deleted its VS Code extension, tab completion, and handoff the moment the frontier moved.
Taste is the moat, not typing. 99% of Amp's code is AI-written; the polish is human ideas. Slop comes from lack of ideas, not from AI.
Remix software for yourself. Fork it, point an agent at it, keep the binary. You don't owe upstream a PR.
Rethink the process, not just the tool. Asked to print kitchen orders on paper, ask why there's paper.