Skill library
Seventeen skills that teach an AI exactly how I build.
TLDR
A skill is a short, sharp playbook an AI follows. These are mine, seventeen of them, covering the whole arc from a raw idea to a shipped product. They run quietly behind every build on this site. This page is the tour, not a download.
Why the model matters
A skill is only as good as the judgement written into it. These were authored with Fable 5, and that shows in the detail. A more capable model packs in more structure, more checks, and more genuine decision-making than the thin, generic skills earlier models tend to produce.
The difference is the one between a checklist and a colleague. Each skill reads like working alongside someone who has built the thing before, knows where it goes wrong, and tells you the test that proves it is right. That is what makes them worth keeping.
How they fit together
The core seven
Turns a vague idea into a single page that decides what you are building and, just as importantly, what you are not. The argument happens on paper in ten minutes instead of in code three weeks later.
The list of what is deliberately left out becomes the real spec. Everything not on it gets argued back in later.
How to run a team of AI agents like a chief executive: decide what matters, write the brief, hand it out, and check the result. The work gets delegated; the judgement does not.
Every breakdown traces to one of three things, a vague brief, no definition of done, or an unchecked deliverable, so problems are always findable.
A repeatable way to get real engineering done with Fable 5: define what done means, plan before writing, then verify in tight loops. Quality comes from structure, not from a longer prompt.
Tight verification catches drift early, while it is still cheap to fix.
Gives an interface a genuine point of view instead of the generic look AI tends to default to. One strong idea carried consistently through type, colour, spacing and motion.
It names the exact tells of generic AI design so they can be spotted and avoided on purpose.
A default stack and build order that gets a real, live URL up on day one, then improves from there. The hard decisions surface immediately instead of piling up.
A walking skeleton on a live URL beats a perfect app stuck on a laptop.
When prompting and search are not enough, this covers how to train a small, private language model: which base to pick, how to prepare the data, and when to stop.
It starts with the cheapest option and only climbs to training when the simpler rungs genuinely run out.
How to organise a knowledge base so the right note surfaces in seconds, for you or for an AI working on your behalf, instead of being buried in folders.
It is structured for the next search, not for tidy storage, so nothing useful goes missing.
The Karpathy method
Ten skills drawn from Andrej Karpathy's essays, talks and code, with every quote checked against the original source.
A six-step routine for training a model reliably, with a check at every step so a broken pipeline gets caught in minutes instead of after a wasted week.
It gives concrete tests that expose silent failures most people never notice.
Before any model code, spend real time reading the actual examples, hunting for duplicates, mislabels and leaks. In modern AI, the data is the program.
The problems that wreck a model hide in the aggregates a dashboard shows; only reading the examples reveals them.
A live AI product gets better through a loop, not a launch: ship it, collect where it fails, fix those cases, and repeat.
The aim is to automate the loop so well the product keeps improving while the team is on holiday.
Pick the one number that matters, freeze everything else, and race to improve it. Vague engineering becomes a sport with a scoreboard.
One number makes progress impossible to argue with; composite scores just hide the trade-offs.
Write code for the person who will read and change it, not for a hypothetical user who will configure it. Fewer options, sensible defaults, less to wade through.
Every config flag is a promise to support a combination you will never test.
Understand something important, like how AI attention or back-propagation works, by building the smallest version of it from an empty file.
You only really debug what you could have built yourself. Building it gives you X-ray vision into the real tools.
Skip the ten-minute explainer videos and do the harder reading. Learn on demand through real projects, and teach what you learn to find the gaps.
Effort is the signal. Most things sold as learning are entertainment in disguise.
Design AI features as a tight loop where the model proposes and the person approves, with a dial so users give it more rope as trust grows.
A giant AI change does not speed a person up, it makes them the bottleneck. Small, reviewable steps keep them fast.
A simple rule for building with AI by the stakes: free-flow for throwaways, AI-with-review for standard work, hand-written for the genuinely novel.
AI is brilliant on well-trodden ground and shaky on the new. Matching the mode to the task is the whole skill.
How to choose work that matters: a problem ten times as important is rarely ten times as hard, so aim higher, then protect the focus to actually finish it.
Work nobody hears about rounds to zero, so the write-up and the demo are half the project.
Want this method on your project?
