Review your prompt.
Improve it in a minute.
Paste a prompt for an AI coding agent. You get a score, the few changes that matter most for that prompt, and a rewritten version to copy. The advice comes from 3,470 real prompts across 631 of our sessions, and each suggestion shows the number that earned it a place.
Stays in your browser. No model reads your prompt, nothing is sent and nothing is stored.
Your improved prompt
Edit anything. The lines in angle brackets are blanks only you can fill in.
You have edited this version. Turning a card on adds its line at the end, and turning it off removes only that line.
What it checks
15 checks, each one a thing we kept having to say in a follow-up message. It only shows the ones that apply to your prompt, and puts the four that matter most first. Every prompt starts on 10 for being written down, and the checks that apply to it earn the other 90.
A live secret is in this prompt
Prompts are stored as plain text in session history, on disk and often on a vendor's servers. A key pasted into a prompt is a key written down in several places.
From our sessions. Two live tokens reached our agent transcripts in September 2026, on the 8th and the 20th. That is why we built a guard hook for it.
Say what you want done, not whether the agent can
A question like "can you look at X" asks for an opinion. An instruction asks for a result. Vague verbs such as look at, check out or think about leave the agent to choose what finished means.
From our sessions. Our one-line prompts are a change plus the constraint that matters, with no explanation of how.
Add a done-when line
Without a check the agent stops when it feels finished, which is when it has written the code, not when the thing works. One observable check turns a request into a result.
From our sessions. 43% of our sessions that opened with a long brief finished on that one prompt, against 26% for a sentence or two. Our long briefs end every item with a done-when check.
Point at the file instead of explaining
Context pasted into a prompt is read once and forgotten. Context in a file is read every time, by every session, and you fix it in one place.
From our sessions. Our median prompt is 67 characters because the rules, the task and the handover live in files. Our first prompt of a session has a median of 221 characters, every later one 52.
Name the one thing that must not change
Agents optimise for the outcome you gave them. If something must stay as it is, a price, a public URL, an old browser, say so in one line or it is fair game.
From our sessions. Our one-line prompts that work give an outcome and one constraint, not a method.
Ask it to look at what it built
A passing test is not a good-looking page. Visual work needs a real render at real sizes, and the agent has to be told to look.
From our sessions. After building, the next prompt we send is nearly always a review: letters too small, something too loud, something broken on a phone.
Say what done means for a release
Merged is not live, and live is not working. If the prompt stops at shipping, the agent will report done at the merge and you will ask whether it is live.
From our sessions. We had to nudge a release 18 times in 14 days before we wrote this line into our rules. In the next seven days we counted, it was 5.
Do not wait on the build
Watching a pipeline is the most expensive way to do nothing. Auto-merge does the waiting for free and the agent can end its turn.
From our sessions. 20.5 of 69.6 hours of agent shell time in a fortnight went on watching CI.
Ask for the wait line
Any turn that ends on a wait looks identical to a turn that has finished. One required last line tells you which it is and what comes next.
From our sessions. "status" was the most typed word of all: 61 times in 631 sessions.
Hand over the small decisions
Every ask-me-first is a full round trip through a person. If the choice can be undone, the agent can take the sensible default and tell you what it picked.
From our sessions. In 21 days 152 of our agent turns ended on a question or an offer and waited about 103 hours between them. 46 of our replies were a bare yes, go or do it.
Plan for the step only you can do
Sign-ins, two-factor codes and console clicks cannot be done by the agent. A prompt that says so up front gets the exact click path and a clean handback instead of a stall.
From our sessions. We typed done, signed in, copied or uploaded 76 times, and 65 of them answered a message that named the exact step.
Say where leftovers go
A session ends with something left more often than not. If the prompt does not say where it goes, it goes into a chat message that nobody files.
From our sessions. In September we had to ask for the next-session prompt in 39% of sessions. After we made it part of the job it fell to 19%.
Audit first, then work in order
A list of tasks without an order gets done in whatever order feels easiest, and some of them are already finished. Sorting by status first stops you rebuilding what is live.
From our sessions. Our 41-task run began with an audit that sorted every task into done, done but unreleased, partly done and not started, with evidence, before any code.
Give an unattended run a stop line
When you are away, the agent still hits things that are yours to decide. Without a stop line it either guesses or sits waiting for hours.
From our sessions. The overnight briefs that finish cleanly name a stop before each blocker and say what to leave on the task.
Compare before choosing
Asked to pick a tool, an agent usually picks the first one it thinks of. Two or three options with a cost each, and a recommendation, takes the same effort and gives you a choice.
From our sessions. A typical prompt of ours: is that the best option? Is there a cheaper one?
Where the advice comes from
We counted every prompt one of our directors typed to Claude Code between 18 September to 12 November 2026. The full count, and what it found, is in our post on one-word prompts.
The meter under your prompt places it against our opening prompts. The cards draw on three of our other write-ups: 20.5 hours of agent time spent watching CI, the guard hooks born from incidents and tickets as files, where our leftovers go.
Isolated by design
The honest answer to "where does my prompt go" is nowhere.
- No model reads it. The reviewer is a file of rules, a handful of regular expressions and some arithmetic. You can read all of it with view source.
- Nothing leaves the page. There are no requests when you paste or type, no cookies and no storage. Open your browser's network panel and watch.
- It never saw our transcripts. Its only data is a short list of numbers we have already published and a table of prompt lengths. A test fails our build if the rules file ever holds a number we have not published, a secret-shaped string or an address.
- It redacts, never repeats. If your prompt holds something that looks like a key or a password, the reviewer names the type, removes the value from the improved version and does not show it back.
What it cannot do
- It spots patterns in the words, not what you mean. A prompt can score 90 and still ask for the wrong thing.
- It does not judge whether the task is a good idea, or run your prompt.
- Its advice is how one small studio works with one family of agents. Where it conflicts with your team's habits, trust your team.
- The cards say what we found in our own sessions. They are not a controlled study, and the post says where the numbers are thin.