The server is the evidence. The skills are the habit.
A coding agent repeats what it did last time, which is why every generated app looks the same. Two small files change the default: one teaches it to ask before it draws, the other refuses to call a screen done until something has been decided.
MCP gives the agent eyes. A skill tells it when to open them.
Connecting a server does nothing on its own. The agent still opens with whatever it has seen most often, and what it has seen most often is a hero, three feature cards and a testimonial nobody gave. That is not a knowledge problem, it is a habit problem.
A skill is a habit written down: a short file in the agent's skills directory that fires on specific situations and carries the steps you would otherwise repeat in every prompt. Ours carry two things — when to look for evidence, and the checks that have to pass before the agent is allowed to say the UI is finished.
Ask first. Then pass the gate.
thinkui — the evidence habit
Tells the agent to settle the decision in front of it with a query shaped like the problem: how a paywall frames the trial, how a dense row shows a remaining balance. One decision per call, read what each reference demonstrates, stop when the question is answered.
Connect the server first, then the skill has something to ask.
thinkui-anti-slop — the gate
The part everyone skips. Four states designed before the happy path, real content instead of placeholders, one element that wins, tokens before raw values, then a render you have actually looked at — and six checks before anyone says done.
Plain text, no code
Each skill is a single SKILL.md. No dependencies, no install
script, no background process. Read it in the repository, or verify the archive against
the sha256 published in the index.
One command each, no key needed to install.
npx skills add https://thinkui.xyz --skill thinkui -y npx skills add https://thinkui.xyz --skill thinkui-anti-slop -y
npx skills add Dionisselami/thinkui-anti-slop -y for the
gate, npx skills add Dionisselami/thinkui --skill thinkui -y for the evidence
habit. Agents that discover skills by URL find both in
/.well-known/skills/index.json.The gate runs on judgement alone. The evidence step needs a key: see what a plan includes.
Six questions, answered in one line each.
A "no" is not a reason to lower the bar. It is the remaining work, and the agent has to say so instead of describing the process.
| Check | The question |
|---|---|
| Specific | Would this screen make sense for a competitor's product unchanged? |
| Evidence | Which decision did a reference settle, and what did it change? |
| States | Empty, loading, error and overflow — seen with real content? |
| Content | Any placeholder names or fake round numbers left in the markup? |
| Hierarchy | Name the one element that wins. Could a stranger name it too? |
| Honest limits | What was not rendered, not tested, or not fixed? |
What these skills will not do.
They will not make an agent a designer, and they do not contain layouts to copy. A reference is evidence for a decision, not a template — reproducing someone's screen in another product is the same slop in better clothes. What they change is narrower and more useful: the agent asks before it draws, and it tells you what it did not verify instead of presenting a confident summary.
They also do not require the server to be useful. Without a key the gate still runs; the skill says the evidence step was skipped rather than inventing a screenshot it never saw.
Questions this page gets.
Do I need a ThinkUI key to install the skills?
No. Installing is one command and the files are plain text. The anti-slop gate works with no account at all; only the evidence step needs a key, and when it is missing the skill says the step was skipped instead of inventing a reference.
Which coding agents can use these skills?
Any agent that reads a SKILL.md from its skills
directory: Claude Code, Cursor, Codex, OpenCode, Windsurf, Zed and the other clients the
skills CLI supports. The server itself is standard MCP over streamable HTTP.
What is inside the download?
One markdown file. No executable code, no dependencies, no background process. You can read the whole skill in the repository or fetch the archive and check its sha256 against the published digest.
Is the anti-slop skill the same as the server?
No. The server holds 1,034 references from real products. The skill is the habit of asking it the right question and the gate that runs afterwards. Install the skill and the agent still calls the server for evidence.
Install the habit. Keep the evidence one call away.
The gate is free to install. The corpus is what makes it specific.