The files that change how an agent works
Skills, rules, tools and runbooks, free to copy into yours.
Each one fixes something agents get wrong. Most take a minute to add.
Find the shelf you need
Where to start depends on the time you have
Three ways in; the bigger the tile, the longer the job.
Add one rule in a minute
Start with Don't flinch if your agent writes for customers.
Install one skill in ten minutes
Pick it for the job you do most.
Hand a long job to a runbook
Give your agent the steps and let it work them.
One job in four parts from one finished piece
Pull apart a piece someone finished reading, and this is what made it.
The finished piece
Write a piece lays out the steps
Runbook. The plan, pinned up before a word.
why-this-piece asks what it is for
Skill. A pause before the first word: who is this for?
but-therefore drafts it as a chain
Skill. Each event causes the next.
gutcheck shows it to buyer personas
Tool. Where a buyer looks, before a person does.
Don't flinch runs under every step
the copy says what we know, and nothing it can't back
Why I publish these
“Every file here is one my agents use.”Mackenzie Bowes
I run a crew of AI agents. Every file here was written when they kept getting the same thing wrong, and is published as they use it.
Toolbox: Skills, rules, tools and runbooks my agents work with, free to copy into yours.
All files, by the job
Each file is one short set of instructions an agent follows for one job. Every entry says who it helps and how to add it.
Skills
When you want your agent to work a way you can name.
- Get more from your agent than the first answer: Before your agent starts, it asks what the job is really for and what an expert would do, then asks for more than you first typed.
- Turn a list of events into a story that people finish: Rebuilds a devlog or post as a chain of "but" and "therefore" instead of "and then".
- Write a story your reader can use: Before you draft, it asks for one line: the situation your reader might meet and the move that lifts or lowers their standing in it.
- Make the choice, then show the result, not the method: It compares the options for each person a choice affects, including doing nothing, then shows you a short result you can check.
- Find the false belief behind a failure that keeps repeating: After a failure, it has your agent write what it did and what it believed, as axioms, principles and cruxes, then test each belief to find the one that is wrong.
- Fix a repeated mistake in structure, not with one more instruction: When the same correction comes up a second time, it picks the cheapest guard that holds it (a type, a lint rule, a hook or a script) instead of adding another line of text.
- Write error messages that say what to do next: Rewrites a vague error or validation message so it names exactly what went wrong, which rule failed and what to do next, without blaming the person.
- Test a change to your agent with a question, a figure and a judge: It turns 'is the agent better now?' into a question asked of many items of real work, with a named judge and a figure you can track.
- Fold a new requirement into a design as if it was always there: When a rule changes, it redesigns the type model or module from the start instead of adding another special case.
- Build the small tool that does the job, not the job by hand: When the work isn't trivial, it has your agent write the script that does it, rerun it on the first unit and diff the result against the hand version.
- Make your agent check the real thing before it says done: Before your agent says it is done, it runs the real program from outside and reads what it prints, instead of trusting a compile, a cached screenshot or its own word.
- Draw a page's pictures for the words that sit on them: It sets each generated scene behind its section's words, drawn for where those words go, with its own render for phones and desktops.
- Pick the format your reader can afford to take in: It ranks ways of carrying a message by how costly each is for the reader to decode, from faces and voices to writing and symbols, so you choose the one they can afford.
- Change your agent's behaviour with one small edit, not a long rulebook: A change to a skill, rule or hook ships only if it is small, one or two intentful edits, because every added line carries claims nobody checked.
- Keep your agent moving instead of asking about everything: Your agent proceeds on reversible work, settles the mechanics itself and shows what it did and why, and still asks first about anything hard to reverse, outward-facing or costly.
- Know why a piece exists before you write it: Before you write, it asks four questions in writing: the point of the piece, why a reader would care, the business goal, and what making it says about us.
Rules
When the same correction comes up twice.
- A block is a detour, not a stop: When a page, a source or a file stops you, find another way to it.
- business-is-a-discipline: Business is a discipline in reality.
- Search your notes before answering a question of fact: Before an agent answers a factual question, plans work or sets a price, it searches the team's research library and names the pages it used.
- Custom CLIs: Custom CLIs give agents large gains.
- Evidence goes beside a loaded file, not in it: A skill, rule, CLAUDE.md or brief loads into an agent's context in full, and every line of it reads as an instruction.
- Nap by default: When you clear your context, nap.
- Neighbours are peers: A neighbour is a colleague whose context runs out, not a queue.
- No eyebrows: Do not put a small label above a heading.
- Say what you know plainly in words customers read: Customer-facing copy states what we know and leaves out hedges and what we don't know, with sources on every page.
- Keep your agent off Reddit: A one-line rule that stops an agent fetching, searching or posting to Reddit by any tool.
- Number every heading and list entry in the library: Mackenzie discusses library pages by pointing at parts of them, so every part needs a number that names it.
- Build key interface parts once: A key interface part is a part that more than one view shows: main navigation, the header, status words, attention colours, a control that appears in many places.
- Ship reversible work: If you can undo a change, make the change, then show it.
- Sign your commits: Every commit you make names you.
- State machines: Design each module as a state machine.
- What your memory holds: Your memory holds three things: pointers to where things live, permissions that Mackenzie gave you, and your live experiments and plans with the agents you work with.
- You are not the user: The person your work is for is not you.
Tools
When a job repeats and a command should do it.
- aion: Asks a model that has no tools one question, with spend caps, one call per run.
- claude-nests: Gives each Claude Code agent a home folder with its name, job, diary and where it left off, loaded when you start it.
- gutcheck: Shows a screenshot or written design direction to buyer personas and tells you where they look, what they trust and what confuses them.
- inkmap: Repaints pen sketches so the ink is one colour on one paper colour, keeping line weight.
- paceline: A launch gate for agents on a Claude Code subscription: says LAUNCH or WAIT until your weekly and 5-hour limits allow.
- qmd: Searches a folder of markdown notes on your own machine, by keyword and by meaning.
- shipd: A local watcher that deploys a site's main branch when every gate passes, and rolls back when the live check fails.
- sleep: Lets a Claude Code agent save its diary and hand-off, clear its own context and wake on the hand-off. Needs a Claude Code mod host, which is not released yet.
Runbooks
When a long job needs steps in order.
- Finding the buyer's problems: How to find, verify and cull the problems a buyer has, before you decide what to offer.
- Runbook: design a directory page that a stranger scans to pick one entry: A directory is an index of many entries of one kind (skills, plugins, tools, crates) where the reader's job is to pick the one that fits the job in front of them, then leave the index.
- Extract a brief: The steps for drawing a one-page brief out of someone who cannot say directly what they need: a client (Part A) or an Aion persona (Part B).
- Runbook: greybox an image: An image greybox works like a page greybox (
greybox-gate): fix the intent cheaply, get Mackenzie's yes in his words, then hand the rendering on. - Judge and fix a site that reads as dated: How to judge whether a small-business site reads as dated to a stranger, and fix it section by section.
- Make an example site that makes an owner say "I want to look like that: The order that worked for making a made-up business page for a niche that a real owner in that niche wants.
- Product demos: how to give a demo or launch video an idea and a look: How to give a product demo or launch video an idea and a look, so that it is exciting rather than a screen recording.
- The SEO runbook: How a page gets found by people and by AI agents, in order, and how sure we are of each belief.
- Set the visual hierarchy of a page: How to set the order in which a page guides the eye: what to look at first, in what order, and how to test it.
- Runbook: write a piece: The steps for writing a post, page, email or offer that a person outside your team will read, in order.