When an AI assistant says “done,” it usually isn’t

When you have AI write code, the most draining part often isn't that it gets it wrong. It's that it clearly didn't get it right, yet tells you it's done.

This kind of problem has a name, and there are open-source skills to fix it. Someone collected 5 of them; each installs with one command and is ready to use. Let's go through them one by one, and add one the original post didn't mention.

1. It says it's done, but it isn't

unlazy, 4K stars.

The idea is straightforward: before it starts working, have it write an acceptance checklist. The checklist is called GATES.md; each gate has three parts: CHECK is a one-line shell command, EXPECT is the string that should appear on success, and EVIDENCE starts empty.

The pass rule is rigid: the process exit code must be 0, and EXPECT must match in the output; only then does the gate pass. Missing either condition means it doesn't count.

A few details are worth noting. --status is the only mode guaranteed not to execute commands; it only checks the checklist's own structure. --reverify reruns every gate that can run, including ones already marked complete. Install an optional Claude Code Stop hook, and if not all gates pass, the agent can't stop even if it wants to.

There's a candid self-limitation in the README: this binding can detect that the checklist was modified, but it can't prevent forgery; anyone who can edit the checklist can manufacture evidence that looks legitimate.

npx skills add Leonxlnx/unlazy

2. It talks too much

caveman, 114K stars. The name and docs are written in caveman speak, but the rules are anything but vague.

For the same explanation of repeated React re-renders, a normal agent uses 63 tokens; this one uses 20:

New object ref each render, so React re-renders. Wrap the prop in useMemo.

Its writing spec borrows the controlled English of aviation maintenance manuals, ASD-STE100: one idea per sentence, a 20-word cap, active voice. Code, commands, paths, and error messages must not be changed at all; numbers and units stay as they are. For safety warnings and irreversible operations, switch back to full sentences.

One easy-to-miss point: it does not change your prompt. Adobe's CAVEWOMAN paper found that compressing the user's own question into caveman speak reduces answer quality.

First, answer one question

Under the original post, Akshat asked: I understand how short answers save output tokens, but how does it save that 33% on input?

Because caveman doesn't only control output. It also compresses what the agent has to read in—logs, JSON, CSV, test output—compressing first, then feeding it to the model. In the official table, CSV goes from 28,041 tokens down to 314, and a whole Claude Code session goes from 885,793 down to 591,673, a 33.2% reduction. Across 54 sessions, all 18 questions were answered correctly.

The table also has one negative row: there is no compressor for HTML yet, and session tokens actually increased by 9.9%. That row wasn't hidden.

npx @caveman-ai/cli@latest

3. What it writes reeks of AI

stop-slop, 18.3K stars. By Hardik Pandya.

It provides three lists: banned phrases, banned structures, and sentence-level rules. At the phrase level, it calls out throat-clearing openers, emphasis crutch words, business jargon, all adverbs, and vague assertions. At the structure level, it bans binary oppositions, negative listing, dramatic sentence breaks, passive voice, and rhetorical setup.

After writing, it scores on five dimensions: directness, rhythm, trust, authenticity, and density, 10 points each. Below 35 gets sent back for a rewrite.

While writing this, I went through its phrase list and found “It's worth noting,” “in other words,” and “in a sense” all on it.

4. It replies with a wall of text

answer-me-with-html, 4.3K stars. Turns hard-to-explain answers into a single visualized HTML page, 2.6x faster than asking the model for HTML directly.

5. Every app it makes looks the same

UI UX Pro Max, 139K stars. Before starting, build a design system for the project—layout, style, color palette, typography—lock it down, then write pages.

The three repositories above also provide single-command installs; just follow their READMEs.

The sixth one the original post missed

Aron Brand added one in a reply, and I think it's harder to cure than the first five: decision slop. The agent asks too often, so your yes stops being valuable.

Frequent confirmation looks like caution, but it's really pushing the cost of judgment back onto you. Ask the same question three times, and you'll still read carefully the first time; by the tenth, you're clicking approve with your eyes closed. By the time there's a real moment when you need to stop it, that button has already become muscle memory.

This problem can't be fixed by installing a skill; you have to change the permission design. Which decisions must be made by a human, which ones it should own—draw the boundaries clearly in advance. If you can't, installing all five above won't help.

Final thoughts

These skills treat behavior, not capability. Installing them won't make the model smarter; it will only make it lie to you a few fewer times, ramble a little less, and push judgment back to you less often.

One more reminder: for any tool that claims to detect AI-sounding text, the score is only a reference. The real standard is simple: would you say this passage in your team chat?