improve-tool
Skill59 starsChanged 2 days agoMITView on GitHub
Remake one tool on gptgames.dev as the best version of it that can be built today - claim, blind inventory, design thesis, mockup critique, rebuild with a tested core, user tests, PR.
In this repository5 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- improve-gameSkill
- new-gameSkill
- new-toolSkill
- next-boxSkill
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 7 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 20
Look, layout, structure and flow are all open.
- Distinctive: line 24
If the page already has a DOM-free core that a suite in test/ checks, you may keep that core.
- Distinctive: line 79
No layout jumps when a result or error appears, readable contrast, reduced motion respected, no console errors, no network request that …
- Distinctive: line 81
CHECK AGAINST THE CONTRACT.
- Distinctive: line 83
Every capability on the list is either present or dropped with a reason in the PR.
- Distinctive: line 85
TEST IT ON SOMEONE, IN ROUNDS.
- Distinctive: line 113
Every visual choice is deliberate: if you cannot say why this font, this colour and this shape belong to this tool and not to any other, …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Remake one tool on gptgames.dev as the best version of it that can be built today. Make every decision yourself and explain the important ones in the PR. Do not stop to ask me, unless the claim in step 1 fails.
Tool requested (may be empty): $ARGUMENTS
THE BAR Measure it against the best version of this tool that exists anywhere, paid or free. Name that product in your thesis. Not against the old page and not against other pages on this site. A stranger who arrives from a search in the middle of some other task should get a correct result within seconds, trust it without checking it elsewhere, and bookmark the page.
WHAT YOU MAY CHANGE Look, layout, structure and flow are all open. Write the new page from an empty file. What stays: the file name and URL, a plain name that says what the tool does in the words people search for, everything the tool can currently do (step 2), links and saved settings that exist today, and the fact that nothing the person enters leaves their browser. If the page already has a DOM-free core that a suite in test/ checks, you may keep that core.
- CLAIM THE TOOL before anything else. With no tool requested above, run
python util/claim.py --tools; with one, runpython util/claim.py tools/<id>.html. It prints the tool that is now yours. If it fails, that tool is taken: stop and tell me. The claim branch it pushes is the only extra branch you may push. Never touch another session's claim. - INVENTORY WITHOUT LOOKING. You do not open the old page or its code until step 8. Send a subagent to do it. It returns a functional list and nothing else: every input the tool accepts, every output and format, every option, every URL parameter or hash it reads, every storage key and its shape, and the exact output for about ten inputs, including awkward ones. Not a word about layout, colour or structure. This list is the contract.
- UNDERSTAND THE JOB. In a few lines: who arrives and what they have in hand (a pasted text, a file, a number, a vague question), what they want to leave with, what the best versions of this tool do well and where they annoy, and what a phone user needs. Name the one moment that matters: the result is there and the person believes it. That is your design thesis.
- THE FIRST TEN SECONDS. Before anything visual, write out step by step what a new visitor on a phone sees on arrival, what they touch first, when the result appears, and how they take it away (copy, download, link). The tool is usable without scrolling and shows a worked example, so nobody faces an empty box. Results update as the input changes; a button only where the work is slow or cannot be undone. Then: what happens on bad input, on input a thousand times larger than expected, and when they come back tomorrow. List every state (example, empty, valid, partly valid, invalid, working, done, failed) and what is on screen in each.
- CHOOSE THE LOOK. Derive it from the thesis and from the material the tool handles: palette, typography, shapes, and one motion idea whose only job is to show what just changed. The result is the most prominent thing on the page, the person's input the second. Type must suit the data: aligned figures for numbers, a real monospace for code. Write down why each choice belongs to this tool; a reason that would fit any tool is not a reason. The tool may have its own character, but people use it for seconds: quiet beats striking. Build a static mockup in your scratchpad, not in the repo, in two states: filled with realistic, awkward content and its result, and the error state. Screenshot both at desktop and phone size. Show the screenshots, one sentence about the tool and the principles below to a critic: a subagent that has not seen your reasoning. It reviews as a harsh critic: what looks cheap, what is confusing, what would make a stranger doubt the result? It must name the three worst defects and give scores from 1 to 10 for first impression, clarity, trust and phone. Do not tell it the pass mark or which round it is. Revise until every score is at least 8, for at most three rounds.
- BUILD IT as CLAUDE.md describes a page. Correctness: the logic lives in a DOM-free core with a suite in test/. Derive it from the authoritative source (the spec, the standard, the published formula), not from the old page. Test it with published test vectors or independently checked values, and with the inputs that break tools: empty, whitespace, huge, malformed, non-ASCII, zero and negative, decimal comma, and the boundaries of the domain (leap days, time zones, rounding). Keep the suite in proportion: a text reformatter needs a dozen cases, anything that calculates money, dates or units needs many more. Input: accept what people actually paste, with stray spaces, units and either decimal separator. Never lose it: not on error, not on reload, and undo works. Errors appear next to the field that caused them, in plain words, saying what would be accepted. Speed: typing never stutters. Heavy work runs off the main thread, shows progress and can be cancelled. Access: every control is labelled and reachable by keyboard, the phone shows the right keyboard for each field, and result changes are announced to screen readers. No layout jumps when a result or error appears, readable contrast, reduced motion respected, no console errors, no network request that carries the person's data.
- CHECK AGAINST THE CONTRACT. Run the inventory's sample inputs through the new tool. Every difference is a bug in yours or a bug in the old one: decide which and say so in the PR. Every capability on the list is either present or dropped with a reason in the PR. Old links with parameters still work and saved settings are migrated.
- TEST IT ON SOMEONE, IN ROUNDS. Write three tasks in a visitor's words, not the tool's ("I have this and need that"), one of them with messy input. Give each to a fresh subagent that knows nothing but the local URL and the task. It works in headless Chromium at phone size from screenshots and clicks alone, without reading the source, and reports what it tried, where it hesitated, how many actions it took and the result it got. Check that result yourself. Then give a fresh critic, as in step 5, screenshots of every state on desktop and phone, the testers' reports and, now, the old page for comparison. It also scores speed to result. Fix the biggest issues and repeat until every score is at least 8 and no tester got stuck, for at most three rounds. If you disagree with the critic, show why in the next round instead of arguing. If a score is still below 8 after the last round, say so in the PR with the open defects. Do not call it done.
- HOUSEKEEPING AND COMMIT as in CLAUDE.md. The commit message says what the tool does, what changed for the person using it, and why it looks and works the way it does. Push, open a PR, and give me before/after screenshots, the contract check from step 7, the testers' action counts and the critic's last scores.
PRINCIPLES
- Correct before beautiful. A wrong result is a failure at any level of polish.
- The result is the page. Everything else is on screen only while it helps someone reach it.
- Faithfulness is owed to what the tool does for people, not to how it used to look.
- The tool explains itself through labels, a worked example and a live result. A paragraph of instructions means the design has not done its job. The exception is trust: where the result depends on a method or assumption (which standard, which rounding), say so in one line near the result.
- The default is right for most people. An option exists only where real needs differ, and rare ones are folded away.
- Every visual choice is deliberate: if you cannot say why this font, this colour and this shape belong to this tool and not to any other, choose again.
- Take nothing from the person: no sign-up, nothing sent anywhere, nothing stored that would surprise them, and no content of theirs in a link they did not ask for.
One tool per session.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

