Harnesses: everything around the AI that makes it reliable

What it is

The AI model is the part that reads and writes. The harness is everything around it: standing instructions, tools, checks and the record of what changed.

Think of a car. The model is the engine. The harness is the rest of the car: steering, brakes, dashboard and seatbelts. A powerful engine without brakes is not a better car. Improving the harness can change everyday results as much as switching model.

A harness usually includes some of these parts:

• Standing instructions: a file the assistant reads every time, often written in Markdown (see guide 06).
• Skills and commands: saved ways of doing repeat tasks (see guide 11).
• Connectors and tools: safe access to your documents and systems (see guide 13).
• Memory: what the assistant keeps between conversations (see guide 24).
• Hooks: rules that run at a set moment, such as before a file is saved (see guide 25).
• Guardrails and permissions: what the AI may and may not do (see guide 26).
• Evaluations: tests you run before trusting a change (see guide 27).
• Version history: a record of every change to instructions and files, so you can see what changed and go back.
• A learning loop: a habit of recording mistakes and turning them into rules (see guide 24).

Why it matters

• The same model gives steadier results when it always gets the same instructions, tools and checks.
• Checks and permissions can catch many problems before they reach customers.
• Version history means a bad change can be undone quickly.
• When a new model arrives, your instructions, skills and tests carry over.

How to do it

1. Choose one repeat task where AI already helps.
2. Write one standing instructions file: purpose, audience, house style and what must never happen.
3. Save one skill for the step you repeat in that task.
4. Add one check: a short list a person ticks before the output is used.
5. Keep every version of these files, with a date and one line saying what changed.
6. Name an owner who approves changes and reviews the set-up on a fixed schedule, for example every three months.
7. Add more parts only when a real problem calls for them.

Example

Before: a team drafts customer replies with an approved assistant, each person with their own prompt. Tone varies, and one reply promises a refund nobody approved.

After: the team builds a small harness.

Standing instructions: customer_replies.md (v3)
- Audience: existing customers. Tone: warm, short, plain British English.
- Never promise refunds, discounts or dates. Write [needs approval] instead.
- Never include another customer's details.

Skill: "Reply draft" (v2), steps: read question, find answer in FAQ, draft reply.

Check before sending (person ticks):
[ ] Facts match the FAQ   [ ] No promises   [ ] Name and order number correct

Change log:
[date] v3: added "never promise dates" after a reply gave a delivery date.

A person still reads and sends every reply. When something goes wrong, the team updates one file and everyone gets the fix.

Common mistakes

• Building everything at once. Start with one file, one skill and one check.
• No owner. Files drift and contradict each other. Name an owner.
• Changing instructions without a record. Keep versions and a change note.
• Thinking the harness replaces the person. It reduces errors; a person still checks anything that leaves the team.

Try it this week

1. Pick one task you do with AI at least weekly (2 minutes).
2. Write a standing instructions file of no more than ten lines, using made-up examples only (6 minutes).
3. Write a three-item checklist for the output (3 minutes).
4. Save both with "v1" and today's date in a shared folder your team already uses (2 minutes).
5. Use them once and add one line to a change note (2 minutes).

Going further

• Version control: a system that keeps every saved version of a file, who changed it and why. Software teams use dedicated version control tools; many shared document platforms also keep version history. Either way, you can compare two versions and restore an older one.
• Keep harness files as plain text or Markdown, so people and assistants can read them and history shows which line changed.
• Assistants such as Claude, ChatGPT, Microsoft Copilot and Gemini offer projects, saved instructions or custom assistants; these are natural homes for the first parts of a harness.
• Connectors may use MCP (Model Context Protocol, an open standard for connecting AI to tools and data); review what each one can reach (see guide 13).
• Rerun your evaluations after any change (see guide 27).

Key words

• Model: the AI engine that reads and writes text.
• Harness: everything around the model that guides, limits and checks it.
• Standing instructions: a file the assistant reads at the start of every task.
• Version history: a saved record of every change, so you can compare and go back.
• Owner: the person who approves changes to a part of the harness.

Examples and figures in this guide are illustrative. AI tools change often: check what your organisation's approved tools offer today. Use approved tools only, and a person checks and decides.

Prefer to print or share it? Download this guide as a PDF (251 KB), or the complete AI Field Guide (1.7 MB).