docs menu
Docs/Quick start

Quick start

ImproveFast isn't a tool you run yourself — it's a skill your agent uses for you. Your job is to decide what to test and how to judge a good result. Your agent tries the options, scores each one, and tells you when it's found a clear winner.

1. Install the skill

bash
$ npx skills add goodfoot-io/improve.fast

Run this once in your project. Your agent (Claude Code or another skills.sh-compatible agent) picks up the skill automatically the next time it runs — see Installation for details.

2. Ask your agent to run a test

Tell your agent what you want to improve, what options you want to compare, and how to score each one. A good request names the thing you're testing and gives the agent clear scoring rules it can use the same way every time:

you, to your agent

"Use the improve-fast skill to test these three summary-prompt versions: concise, detailed, and bulleted. For each summary you produce, score it 0–1 like this: 1.0 if a human reviewer would accept it with no edits, 0.5 if it needs small edits, 0.0 if it needs a rewrite. Keep going until you're confident about a winner, then tell me which version to keep and why."

Your scoring rules are the most important part of the request. Loose rules like "score it based on quality" lead to scores that don't mean the same thing from one try to the next, which makes it take longer to find a real winner. Be specific about what a 1.0, a 0.5, and a 0.0 look like, and give your agent real examples if you have them.

3. Let the loop run

From here, your agent handles the loop on its own: it picks an option to try next (leaning toward whichever is winning so far, without ignoring the others), does the work, scores the result using your rules, and repeats. You don't need to watch every step — check in whenever you like, or just ask your agent directly:

you, to your agent

"How's the summary-prompt test going?"

Your agent will tell you how many tries it's scored so far, roughly how each option is doing, and whether a winner has shown up yet. Most tests wrap up before every option has been tried the same number of times — your agent naturally spends more tries on whichever option looks best, while still checking the others enough to be fair.

4. Get the winner and act on it

Once your agent is confident about a winner, it reports back with the numbers behind it — how many tries it scored, the average score, and how that compares to the other options. That's your cue to use the winning option in the system you're improving: update the production prompt, switch the default approach, or lock in the winning setting.

your agent, reporting back

"The test is done: detailed is the winner, averaging 0.85 across 22 scored tries versus 0.72 and 0.66 for the other two. I've updated the production summarization prompt to use the detailed version."

A few things worth knowing

  • Your scoring rules are the whole game: ImproveFast only improves what you score — nothing else. Loose or inconsistent scoring rules lead to a winner you can't really trust.
  • A winner needs enough tries: the leading option usually needs 15 or more scored tries with a clear, steady lead, up to a limit of 50 tries per option if the results stay close.
  • Tests expire in 7 days: ImproveFast is built for quick tests, not long-term tracking — if you're picking a test back up later, ask your agent to start a fresh one.