a skill for your coding agent
ImproveFast is a skill your agent runs on its own. Tell it what you want to test and how to judge a good result, and it tries the options, scores each one, and tells you which one wins.
$ npx skills add goodfoot-io/improve.fastskill via skills.sh · works with your coding agent · no signup
example run — solid marks are scored tries; dashed slots are tries the losing options never needed.
the problem
You tweak a prompt, read a few outputs, and ship it. But AI outputs aren't the same every time, so one good or bad response doesn't really tell you anything. Manual A/B tests take a long time and waste effort on the option that's already losing. And most teams don't have time to build a whole testing setup just to check if a change actually helped.
A plain 50/50 test keeps sending tries to the option that's already losing. ImproveFast shifts more tries to whatever is winning as it goes — in our own test, 90 of 100 tries went to the better option automatically.
ImproveFast can tell the difference between a real improvement and plain luck. It waits until it's confident before calling a winner, so your agent isn't guessing either.
You don't need to build anything. Your agent just scores each try from 0 to 1, and ImproveFast handles the rest.
how it runs
ImproveFast gives your agent a simple loop to follow: try an option, do the work, score how well it went using rules you agreed on, and repeat. Scoring each try, again and again, lets it tell a real improvement apart from random luck — something a single side-by-side comparison can't do. You never have to think about the math. You just decide what to test and how to judge it.
01
Tell your agent which options to compare — prompt wording, strategies, or settings — and agree on how to score a result from 0 to 1.
02
The agent picks an option to try next, leaning toward whichever is winning so far without ignoring the others, and does the work.
03
Each try gets a score using the rules you agreed on. Scoring again and again is what separates a real improvement from luck.
04
Once it's sure enough, the agent reports the winner with the tries and average scores behind it. You decide whether to use it.
what changes
when it fits
worth testing
quick start
Install the skill once, then just talk to your agent — describe what you want to test and how to judge it. No account, no API keys — the agent handles the rest. See the installation guide for prerequisites and verification.
$ npx skills add goodfoot-io/improve.fast30-second setup · hosted service · free to use
faq
The leading option usually needs around 15 or more scored tries with a clear, steady lead before ImproveFast will call it — most tests wrap up well before the 50-tries-per-option limit. If your agent scores a handful of tries an hour, expect an answer within a session or two, not weeks.
Yes. Each test is kept separate by its own ID and is deleted automatically after 7 days. ImproveFast is built for quick tests, not long-term tracking, so your data doesn't stick around any longer than it needs to.
That's fine — ImproveFast works with any number of options. More options need more tries to sort out, but it spends most of its effort on the promising ones instead of splitting attention evenly across all of them.
Exactly what you and your agent agree to score — nothing more. That's why clear, consistent scoring rules matter so much: if the scores are inconsistent or vague, the result won't be trustworthy either.