The 12 Claude Skills We Actually Use to Run OneWave

194 Skills on Disk, 12 That Actually Run the Company

We have published 162 Claude skills to our open-source library. They have been installed roughly 29,000 times. That sounds like a flex until you sit in our actual working directory and count what we open on a normal Tuesday, which is closer to a dozen.

This is that dozen. Not the most-starred ones, not the ones that demo well - the twelve that touch real revenue, real client work, and real Tuesday-afternoon problems. Every command below is one we typed this month.

They split three ways, and the split is the point. A skill library that only makes engineers faster is a developer tool. Ours covers selling, building, and running the business - because the person who closes the deal and the person who ships the code are, at a company our size, frequently the same person on the same day.

A skill is just a folder with a markdown file in it. That is the whole format. The leverage is not the technology - it is having written down, once, the way you want a job done, so you never re-explain it again.

Part One: The Four We Sell With

1. onewave-icp - argue with your buyer before your buyer argues with you

This one changed how we ship marketing more than anything else on the list. It role-plays OneWave's ideal customer as a three-person focus group: the bullseye buyer, the executive sponsor, and a skeptic whose default answer is no. You hand it a landing page, an offer, a price change, a cold email. It reacts in character.

The critical detail is that the personas are not invented. They are derived from our real book - active clients, closed-won patterns, and the deals we lost - so the objections that come back are objections we have genuinely heard from people who had a checkbook.

Each persona returns a gut reaction, the one thing that made them lean in, the exact place they stalled, and a verdict. Then the personas argue with each other, and the argument is where the signal lives. The sponsor loves the vision, the skeptic says show me the price, and the gap between them is the thing your page is failing to resolve.

/onewave-icp pointed at a pricing page will tell you it reads like the retainer we lost, and it will be right.

2. onewave-lead-panel - the same trick, aimed at a live prospect

Where the ICP skill reacts as your archetype, this one reacts as a specific person currently in your pipeline. It reads the live CRM, builds the persona from that lead's actual record, and tells you whether they would reply, book, or ghost - and why.

It also does the unglamorous thing we need constantly: the pre-call brief. Who they are, what they have touched, ICP fit score, and where the deal actually stands. That used to be fifteen minutes of tab-switching before every call.

3. expert-panel - opinions with sources attached

Ask any model a strategy question and you get confident vibes. This skill refuses to let an expert speak until research is done first, and that ordering constraint is the entire value.

It runs three phases:

Research - two or three agents search the web in parallel, each from a different angle, and must return specific numbers with citations
Debate - two to five domain experts receive that research and analyze it. The rules require at least one numbers person, one domain specialist, and one contrarian whose job is to find what breaks
Synthesis - consensus gets separated from disagreement, everything is benchmarked against the research, and you get two or three concrete options with a recommendation attached

The skill takes a position at the end rather than presenting options neutrally, which is the difference between a consultant and a search engine. We have used it on pricing structure, comp plans, and whether to build something in-house.

4. prospect-research-compiler - the twenty minutes before every first call

Aggregates prospect intelligence from news, social, the company site, and financial data into one brief. Unremarkable as a concept and enormously useful in practice, because the alternative is doing it manually or, more honestly, not doing it at all.

Part Two: The Four We Build With

5. agent-army - many brains, not one brain divided

Our most-used skill and the one we wrote a whole post about. It deploys a two-layer hierarchy: a commander that does the thinking, a layer of specialists that each get their own full context window, and sub-agents under every one of those.

The distinction that matters is swarm versus army. A swarm splits one context window across sub-agents - one brain, divided. An army gives every Layer 1 member its own window. You pick a size and a power level at deploy time, which sets the model mix per layer, so a mechanical find-and-replace does not get billed at Opus rates.

The part we did not expect to matter as much as it does is the phantom-completion check. Agents will report COMPLETE on work they never did. The skill cross-references every report against git diff --stat and re-dispatches the liars. Trust the diff, not the report.

6. web-army - the same structure, pointed at a website

Specialist agents for design, components, content, styling, and performance, each with sub-agents beneath. This is what we reach for on a full redesign rather than a refactor. The specialization matters more than it sounds like it should - a generic agent asked to improve a page will drift toward changing everything at once.

7. hyperframes - videos that are HTML files

You write a composition in HTML with timing attributes and render it to MP4. Because the video is code, an agent can write one, and because it is deterministic, the render is reproducible.

The explainer video embedded in our Agent Army post was made this way - an army wave scripted the scenes, wrote the composition, and one command produced the file. No timeline, no scrubbing, no editor. The longer explanation is here.

8. create-image - one model, no arguing about which one

Generates images through gpt-image-2. The reason this exists as a skill rather than a habit is consistency: without it we would drift between models and get header images that do not look like they came from the same company. The header on this post came out of it.

Part Three: The Four We Run the Business With

9. meeting-intelligence - transcripts to commitments

Takes meeting notes or a transcript and extracts decisions, action items, blockers, and sentiment, then drafts the follow-up email. The extraction of blockers specifically is the part that earns its place - those are the things everyone nods at during the call and nobody writes down.

10. contract-analyzer - read the whole thing, flag the weird parts

Reviews contracts for concerning clauses, extracts key terms, and compares against standard terms. It is not a lawyer and we do not use it as one. It is the pass that happens before the lawyer, so the expensive human is reviewing three flagged provisions instead of forty pages.

11. owai-auth - Google sign-in in one command

We build a lot of small internal tools, and every one of them needs to be locked to our domain. This skill drops Google SSO restricted to our workspace into any Next.js or static app, with the client ID and the domain restriction already baked in.

It replaced a shared password in a group chat, which is the kind of thing every company does and nobody admits to.

12. skill-creator - the one that makes the others

Creates new skills, improves existing ones, and - the underrated half - runs evals against them to measure whether a skill actually triggers when it should. A skill with a vague description never fires, and you will never notice, because nothing happens and nothing is what you would expect from a tool you forgot you installed.

Most of the twelve above started as a paragraph of instructions we had pasted into chat three times in a week. The third time is the signal. That is the whole authoring process.

How to Actually Find the Right One

With 194 skills on disk, discovery becomes the real problem. Two skills exist purely to solve it: scout reads what you are currently working on and recommends the right skill or subagent for it, and skill-navigator handles the broader "what do I even have that could help here" question.

The honest failure mode of a large skill library is not bad skills. It is good skills nobody remembers owning.

The Pattern Underneath All Twelve

Look at what these have in common and it is not the AI. Every one of them is a decision we made once, written down in a form that executes.

The ICP panel is what we learned from deals we lost. The expert panel is a rule that opinions need sources. Agent Army is an org chart. The contract analyzer is a checklist someone senior used to run by hand. None of that is technology - it is institutional knowledge that normally lives in one person's head and leaves when they do.

The companies getting real leverage out of AI are not the ones with the best prompts. They are the ones that noticed their best judgment was undocumented, and wrote it down somewhere a machine could run it.

That is also why we do not think the skill library is the product. The list above is worth roughly nothing to a company that has not first decided how it wants these jobs done. The writing-down is the work. The skill is just where it lives afterward.

Try Them

The library is free and open source:

Clone it: git clone https://github.com/OneWave-AI/claude-skills.git ~/.claude/skills
Open Claude Code or Cowork and type / to see what installed
Start with /expert-panel on a real decision you are stuck on - it is the fastest one to judge, because you can check its sources

The OneWave-specific panels will not work as-is, since they read our CRM. But they are worth reading as a template: the pattern of grounding a persona in your own closed-won data is portable to any company with a pipeline, and it is the single highest-leverage skill we have written.

If you want your team's judgment turned into skills like these - or you want to know which of your recurring decisions are worth writing down first - that is what our bootcamp does, and you can just ask us.