Every Monday you re-brief your AI like a brand-new hire. Business context, voice rules, the report format you like — typed into a blank chat from memory, because nothing carried over from Friday.
That routine is the tax you pay for using a smart model with no system around it.
A Claude Code harness ends it: a small set of plain-text files wrapped around Claude Code that hold your business context, your procedures, and your memory — so every session starts already knowing how you work.
(And if the word “harness” makes you picture something deeply technical... it’s folders with markdown files in them. You can read every line.)
One note on why this matters more than it did a year ago: when the Fable 5 model left Claude subscription plans two weeks after launch, a lot of builders discovered that the part of their setup they’d fallen in love with was rented. The files around the model were the part they could have kept. (I wrote up what an agent harness is — and what survived that week at my desk — separately.)
The full concept — what a harness is, the five layers, why one system beats twenty tools — lives in the pillar guide. This page is the build: the anatomy, the daily loop, and a first week that ends with your harness doing a real job.
The anatomy: four kinds of files
Open a working harness and you’ll find the same skeleton every time.
your-business/ ├── CLAUDE.md <- who you are, what you sell, house rules ├── .claude/ │ ├── skills/ <- your SOPs (voice, reports, publishing) │ ├── agents/ <- specialist workers for delegated jobs │ └── commands/ <- shortcuts you fire on purpose └── notes/ <- memory: decisions, logs, running context
CLAUDE.md is the instructions file — Claude Code reads it automatically at the start of every session. Yours should say what the business is, who it serves, what good work looks like, and the rules that never bend. One page is plenty to start.
Skillsare your standard operating procedures: how you sound, how the weekly report is formatted, how a draft becomes a post. Claude Code loads one whenever the task matches its description. (There’s a full walkthrough in the skills guide — your first one takes about 20 minutes.)
Agents are workers you delegate whole jobs to — a researcher, a reviewer — each with its own instructions, so the main session stays focused while they grind.
Memory is the unglamorous layer that makes the whole thing feel alive: files the AI reads and writes as it works — decisions from last week, a log of what shipped. Ask it to keep notes, and next session it remembers, because the notes are sitting right there in the folder.
All four layers are markdown. If you can write an email, you can write a harness.
The loop that does the work
This is what happens when you hand a harnessed Claude Code a job.
It reads the situation — your CLAUDE.md, the relevant skill, the notes. It picks a tool. It acts: edits the file, runs the check, drafts the post. Then it looks at what came back and decides... done, or go again.
That wheel, turning without you pushing every spoke, is the entire difference between asking an AI questions and giving it jobs.
The model supplies one turn of the wheel. The harness keeps it turning — and keeps it turning your way, because your files decide what it reads, what it checks, and what it’s not allowed to skip.
There’s a name for the craft of tuning all this: harness engineering. Searches for the term went from almost nothing to thousands a month inside a year. At the business-builder altitude, it means writing down how you work, then folding every correction back into the files.
Build with your smartest model, run with your cheapest
A pattern worth stealing from people who build these for a living: the build and the run don’t need the same brain.
Designing a harness — deciding the file structure, drafting the skills, thinking through your process — is hard reasoning work. Use the smartest model you have access to for those sessions.
Running the harness once it’s built is mostly following well-written instructions. A cheaper, faster model handles the daily loop fine — if the files are good, you barely notice the difference.
And because the whole harness is plain text, it isn’t married to any model at all. The same folder works when you point a different model at it next year. Swap the brain, keep the body.
Your first week, day by day
The path I’d hand a business builder starting from zero. Five sittings, somewhere between 20 and 90 minutes each.
- Day 1 — first session. Install Claude Code, set the safety rails, ship one tiny real thing. The first-session walkthrough covers the exact prompts; budget about 90 minutes.
- Day 2 — the instructions file.Don’t write CLAUDE.md; dictate it. Tell Claude Code: “Interview me about my business — one question at a time — then write a CLAUDE.md capturing what matters.” Then edit what it drafts. (Mine got rewritten all week. The first version described the business I wished I ran, and the drafts that came back proved it.)
- Day 3 — first skill.Voice guide first; it feeds everything else you’ll ever generate. About 20 minutes.
- Day 4 — memory. Create a notes folder and tell Claude Code to keep a running log of decisions and finished work. Then start a fresh session and watch it pick up where you left off. This is the day the system starts to feel different.
- Day 5 — a real job.Hand it something from your actual week: the newsletter draft, the client summary, the content repurpose. Each time you correct it, end with: “update the skill so that mistake can’t happen again.” Same loop the skills guide teaches: the correction goes into the files, not only the draft.
By Friday you have a small, ugly, working harness... which is exactly on schedule.
Three week-one mistakes worth skipping
I’ll save you the detours I watch new builders take.
Over-building before using. Ten skills drafted, zero jobs run. Harness files are theories about how you work until a real job tests them — build one file, use it the same day, revise what breaks.
Writing CLAUDE.md like a novel.The model reads it every single session, and a bloated one buries the three rules that matter under forty that don’t. One page. Prune it every time you add to it.
Working without a net.Claude Code edits real files — that’s the whole point, and the reason your first week should run inside a folder that’s backed up (or a copy). Set up the undo button on Day 1; the first-session walkthrough covers it.
Or start from a working harness
You don’t have to build from scratch. Copying a good harness is a legitimate way in. (For most business builders, it’s the way I’d recommend.)
The one I’d point a beginner at today is Scott Graham’s safe-agentic-workflow (cheddarfox on GitHub): a real Claude Code harness with 2,193 commits and 169 issues across 5 months of production use. It treats your AI like a small team — a planner, an architect, a reviewer, each with a job — while you drive.
Getting started is three steps: copy its .claude/ folder into your project, fill two placeholders in SETUP.md, run /start-work.
(Could be a different repo I’d recommend six months from now — this space moves fast, and I hold the pick loosely. It’s my pick today.)
The moment you copy it, it’s yours: readable, editable, version-controlled. Study how its files talk to each other and you’ll learn more about harness design in an afternoon than a week of videos would teach you.
The part of your AI setup you actually own
The Academy runs on one belief here: own your agents, swap your models.
Models are rentals — they get repriced, deprecated, and occasionally pulled outright. The harness is files in your folder. Nobody can revoke it, and every hour you put into it compounds instead of evaporating with the next model cycle.
If you’d rather not assemble yours alone: the Academy’s Armory is exactly this, pre-built — Rigs, which are battle-tested harnesses bundled around one business job each (content repurposing, funnel building, SEO audits), plus a crew to sanity-check your build and two calls a week where the half-broken version is welcome. The Academy includes the full Armory at every tier — $83/month or $497/year, with a 7-day free trial.
Either way — build it, copy it, or join and adapt one — the move is the same. Put your hours into the files you keep.
Questions people actually ask
- What is a Claude Code harness?
- A Claude Code harness is the set of plain-text files built around Claude Code that turns a general AI model into a worker that knows your business: an instructions file (CLAUDE.md), skills holding your standard procedures, agents for delegated jobs, and memory files it reads and writes. The model supplies the intelligence; the harness supplies your context, your standards, and your process.
- Do I need to know how to code to build a Claude Code harness?
- No. Every file in a basic harness is plain-language markdown — instructions written the way you'd brief a new hire. If you can describe how you work in an email, you can write harness files. The hard part is articulating your own process clearly, and that skill outlasts any particular tool.
- What files make up a Claude Code harness?
- Four kinds: CLAUDE.md at the project root (business context and house rules, read automatically every session); skills in .claude/skills/ (your SOPs); agents in .claude/agents/ (specialist workers you delegate to); and memory files — notes and logs the AI reads and writes so context survives between sessions. Slash commands in .claude/commands/ are a common fifth, for shortcuts you trigger deliberately.
- What is harness engineering?
- Harness engineering is the craft of designing and refining the system around an AI model — the instructions, skills, memory, and checks — rather than tinkering with the model itself. Searches for the term grew from almost nothing to thousands per month within a year. In practice it comes down to one habit: every time you correct the AI, fold that correction back into the files.
- Can I copy someone else's harness instead of building my own?
- Yes, and it's a strong way to start. The moment you copy a harness into your project it's yours — readable, editable, version-controlled text. A well-tested public example is safe-agentic-workflow on GitHub (2,193 commits across 5 months of production use); the AI Captains Armory offers pre-built Rigs with a crew to help you adapt them. Expect to customize either — the generic version fits nobody perfectly.
- How long does it take to build your first harness?
- About a week of short sessions: a 90-minute first sitting, then 20–60 minutes a day to add the instructions file, a first skill, and memory. The honest caveat: version one describes how you think you work, and real jobs will edit it — expect to revise the files repeatedly over your first month. That revision is normal — that's how the system compounds.
