Skills your agents can actually use.
Tell us what your agents need to know. Our researchers read your topic area in detail (talks, books, reports) alongside any material you send us, and hand back a skill in the open Agent Skills format, tested against the questions you set.
One folder that loads in whatever your team already uses. Claude and Claude Code, ChatGPT and Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI, JetBrains Junie, Amp, Goose, Mistral Vibe, Kiro, Roo Code, Tabnine and forty-odd others.
What you get
The skill itself, the evidence that it works, and the validation behind both. Lean enough to drop straight in and start using. The reading behind it stays with us, where it belongs.
The skill
A folder in the open Agent Skills format: the SKILL.md standard Anthropic
published and the rest of the industry adopted, now supported by around forty agent
products. Written to trigger on the right task, not to sit unread, and yours to drop into
your own repository.
Your test questions, answered
You give us the questions the skill has to handle. We run them against an agent with the skill loaded and without it, and hand you both answers side by side, so you can see exactly what the skill adds on the work that matters to you.
Validated and evaluated
Every claim in the skill is checked before we send it. The skill is then evaluated following the best practices set out alongside the Agent Skills specification: graded against what a good answer has to do, and checked that it loads for the tasks it is meant for and stays out of the way on the ones it is not.
Why bring someone in for this?
Writing a skill looks easy. It is a Markdown file. That is exactly why most of them do not work. They read well, they load, and the agent behaves the same as it did before. The work is not the writing. It is knowing what is worth telling a machine, getting it out of the people who know it, and proving afterwards that it landed.
What your people know is not written down
The way your best estimator prices a job, the checks your senior engineer runs without being asked, the reason your team always does that step in that order. It lives in people and in habits, and every agent you deploy starts without it. Send us the material: the handbooks, the past deliverables, the recordings, the process nobody has updated since 2019. We turn it into something an agent can actually apply.
You have not written one before
The format is simple and the craft is not. What to put in, what to leave out because the model already knows it, how specific to be, when a rule helps and when it makes the agent worse. All of that comes from having written a few hundred of them and measured what changed.
You want the work checked
You have written some and they seem fine. Do they fire on the right prompts? Do they change the answer at all? Most people have never tested either, because until you run the same question with and without the skill there is nothing to compare.
You have one that needs to be better
It worked when you wrote it. Models move, products move, practice moves, and a skill that has drifted still loads and still answers, so the assessment is what tells you. You get a score, a diff, and proof of which changes improved an answer.
It is not the best use of your people
You could do this. It would take a week of reading, a week of drafting, and someone senior enough to tell a good source from a confident one. That is the week we are selling you back.
Why not just ask the AI to write one?
You can, and for something small you probably should. It is worth knowing what you get.
Ask a model to write a skill and it writes from what it already knows. That is exactly the material a skill cannot help with, because the agent had it already. You get fluent, sensible instructions to handle errors appropriately and follow best practice, and the agent behaves the same as it did before, because none of it was news.
The standard's own authoring guidance names this as the common failure: a skill generated from a model's general training knowledge produces "vague, generic procedures" instead of the specific patterns, edge cases and conventions that make a skill worth loading.
What a skill is actually made of is the part the model does not have.
What your business knows
Your conventions, your exceptions, the reason the obvious approach is wrong here. We go and get it, from your material and from the people who hold it.
What the field knows
We read the practitioners, not the summaries. Where the experts disagree, you get the disagreement rather than an average of it.
Proof it changed something
Your questions, run with the skill and without it, side by side. Generate a skill yourself and that comparison is the step nobody does, so nobody finds out.
Knowing what to leave out
A skill competes for the agent's attention with everything else in the window. Over-instruct and it gets worse, quietly. Most of our work is deciding what to cut.
Making it fire at the right moment
A skill that never triggers has no quality, and one that triggers on everything is a tax on every prompt. Getting that right is measured, not guessed.
The unglamorous format details
There are hard limits in the specification that break a skill quietly when exceeded. We checked our own library of 174 and found 45 over one of them. It is the kind of thing that only shows up when you are counting.
We have been doing this a long time
We were writing skills for AI systems years before the current standard existed. The first ones were Alexa skills, back when teaching a machine to do a job meant fighting an intent model and a slot parser. The format got better. The hard part did not change: deciding what the machine needs to be told, and what telling it will cost you.
We have worked in the Agent Skills format since December 2025, when the specification was published, and we build on it every day. We keep a private library of 174 skills across twelve categories: the tools we use to do our own work, maintained and versioned, not a portfolio assembled for a website. Writing them is how we learned to tell you what a skill is and is not supported by.
It also means we rarely start from nothing. Much of what a new skill needs (how to shape a trigger, where the format bites, which instructions a model quietly ignores) is already solved in there and gets applied to your job on day one. What we write for you is yours, and the library stays ours; you are buying the experience in it, not a licence to it.
Since December 2025 we have also been building agentic skills inside GreenCode, a funded multi-partner R&D programme on sustainable software. That has meant using them in anger across genuinely different problem spaces: surveying a research literature, generating architecture documentation from a live codebase, attributing energy consumption across a build, producing formal deliverables against a schedule and a template. Different domains break a skill in different ways, and iterating over months on real work is how you find out which instructions survive contact with a problem they were not written for.
What that experience actually buys you is mostly subtraction. Knowing which instruction to cut, which rule the model will ignore, which one will make it worse, and when the most valuable thing we can tell you is that your agents already handle this and your money is better spent elsewhere.
Built to speed your people up, not to replace them
The point of a good skill is that the person stays in the chair. Your team keeps the judgement, the relationships and the accountability; what changes is how much of the work around those has to be done by hand, and how long it takes to get to a first draft worth arguing with.
An agent working without a skill has to guess at how your field thinks and how your organisation works, and it guesses plausibly, which is the expensive failure, because plausible output takes longer to check than obviously wrong output. A rich skill removes the guessing: the agent works from the practice as it is actually done, so what comes back is closer to right the first time and faster to correct when it is not.
Less time on the first 80%
The research, the structure and the first pass are where the hours go and where the least judgement is needed. That is the part a skilled agent can carry.
More time on the part only they can do
Deciding what matters, weighing a trade-off, knowing what the client will not say out loud. A skill makes that work start sooner, it does not do it for them.
Your practice, held consistently
The way your best people already work, written down once and applied every time, including by whoever joined last month.
We build to a target we can hold ourselves to: good enough to be worth your expert's time. Work that arrives far enough along to be argued with rather than started. Every claim in it is checked back against the source it came from before delivery, so what lands on their desk is backed by the material.
Ready to start?
Tell us what your agents should get better at, and the questions the skill has to handle.