How to Actually Use AI Agent Skills

Two skill repos are pulling attention right now. Matt Pocock’s skills and Lauren Tan’s pstack. The obvious move is to install both and hope your agent gets smarter. That move fails. AI agent skills are just markdown files, and markdown you have never read is noise. The video this post covers spends a week inside both collections. The useful part is not the final list of skills. It is the method used to pick them.

Start with one skill: unslop

Unslop comes from pstack. It edits text to strip AI writing patterns and put a human voice back in. The process is four steps. Scan for the patterns, rewrite while preserving meaning, add some voice, then self-audit by asking what still reads as machine-written.

It hunts specific tells. Puffery like “pivotal moment” and “testament to.” Superficial verbs like “highlighting” and “showcasing.” Chatbot filler like “I hope this helps.” Em dash overuse, which the video points out is still leaking out of Opus. The best rule in the file is plainer than any of that. Say what it does, not how it feels. If a sentence could appear unchanged in another project’s docs, it says nothing about yours. Cut it.

The demo is what sells it. The same model answered the same question twice, once with the skill loaded and once without. One markdown file separated the two answers. That is a cheap trade for output you will actually read.

Grilling turns your agent into an interviewer

Matt Pocock’s grilling skill does the opposite of writing code. It interviews you. It maps your plan as a decision tree, then works the tree in rounds. Each round covers the decisions whose prerequisites are already settled.

The questions land hard. Who is the user and who pays? What is the actual next milestone? Which rule wins when your own guidance contradicts itself? Every question ships with example answers, so you can reply “1A, 2B, 3A” and keep moving. In the video, round two produced a question about which kind of simplicity mattered most, and the honest reaction was “I want all three.” That is the skill working. Vague plans feel fine right up until an agent builds the wrong thing.

Audit your own usage before installing AI agent skills

This is the part worth copying. Picking skills off a README is guessing. Picking them off your own history is not.

Start with a prompt, not an install command

The first move in the video is a single prompt aimed at the repo and at his own machines. Roughly: audit my usage across these three machines, compare it against the skills in this repo, and rank all of them by how well they fit me and how much they would benefit my work. Add a short reason for each. Pull history first, then split into sub-agents to do the auditing.

That last instruction carries most of the weight. One agent cannot hold a week of sessions plus forty skill files in context and still reason well. Sub-agents each read a slice and report a summary back. Note that this only works if you have real history on the machine. A fresh install has nothing to audit, so run this after a few weeks of normal work, not on day one.

Argue with the ranking

What comes back is a document, not a shopping cart. Each skill gets a fit rating, a benefit rating, and a line explaining why. In the video the top of the list included interrogate, blast radius, technical writing, arena, and prove it works.

He installed seven of them. Wizard ranked seventh and got left out, and he only noticed days later when he went looking for it and had to add it. The ranking is an opinion from a model that read your shell history. Treat it that way. Run the same audit against a second repo and compare the two lists before you commit to either.

Try a skill before it touches your machine

You do not need to install anything to test a skill. Open the SKILL.md on GitHub, copy the markdown, and paste it into your agent with a line like “I want to use this skill, I don’t have it installed, the text is below.” The agent reads it and follows it. That is the entire mechanism anyway, since installing a skill only means the model reads the file when the description triggers.

Skills that reference scripts need the script too, though you can usually tell the model where to download it. For pure markdown, the test costs two minutes and no configuration. Do this for every skill on your shortlist before any of them land on disk.

Keep what you take in one repo

The video’s setup is a personal repo called fleet that holds every skill file across his machines. He works in that repo, tells an agent to pull the parts he wants from someone else’s collection, and the change propagates. When he added a new Mac Mini, he mentioned it in the prompt and the agent covered that machine too.

You do not need that exact setup. You do need one place where your skills live. Otherwise you end up hunting through .claude and .agents directories on four computers trying to remember which one has the current version. A lot of people avoid opening those directories at all, which is the real problem. Editing your own markdown is the whole point.

The skills worth knowing about

A few others earned their place during the week of testing. Wizard walks a human through the steps only a human can do, which helps when the agent cannot reach a dashboard or run a privileged command. Blast radius finds what a change breaks elsewhere before it ships, and it refuses to trust the thread’s own write-up. It picks the one or two facts the conclusion depends on and proves them by running code. Diagnosing bugs fires on its own during debugging and sharpens root-cause explanations.

Show me your work keeps a reviewable decision trail as an append-only TSV log. One row per decision, with what, why, evidence, and result. If a decision does not fit on one line, it is not crisp yet. Arena fans out parallel attempts at the same task, reads every candidate end to end, then grafts the best ideas from each into one result. It also burns hundreds of dollars of inference, so know that going in. Writing for agents covers the case people forget, which is that agents write bad instructions for other agents unless you teach them how.

Key takeaways

• Skills are markdown files, so the value comes from reading and adapting them rather than installing the largest collection you can find.
• Unslop is the highest-leverage first install because it makes every other agent response worth reading.
• The grilling skill catches misalignment early by interviewing you about your own plan before any code gets written.
• Auditing your real machine history against a repo produces a ranked shortlist that fits your work better than anyone else’s setup.
• You can trial any text-only skill by pasting its markdown straight into your agent, so nothing needs to be installed to be tested.

Conclusion

Copying another engineer’s configuration wholesale teaches you nothing and clutters your context. These repos work better as reference material. Read the files, test a few by pasting them in, keep the ones that match the work you actually do, and drop the rest. That is how AI agent skills become a toolbox you own instead of a template you rent.