Identity
Who the AI is supposed to be in this project: assistant, research coordinator, code builder, writing coach, verifier, or orchestrator.
Harness design
Your harness should become yours.
Jay's starter harness is meant to be adaptable. The goal is not to preserve the file unchanged. The goal is to learn how to design the instructions that shape your AI's work.
Common question
Should we change the identity?
Yes, if the current identity does not fit the work. The starter harness is not specifically designed only for students doing research projects. It is designed to be universally capable, expandable, and adaptable.
If a student is building apps, doing debate research, preparing a competition submission, running a club, or managing a personal portfolio, the harness should eventually reflect that. A good harness teaches the AI what kind of work it is helping with and what standards matter.
What a harness actually does
It sets the root layer of the workflow.
A harness is the standing instruction file that Codex or Claude Code reads before helping you. For Codex, this is usually AGENTS.md. For Claude Code, the equivalent is usually CLAUDE.md.
Rules
What it should never do: fabricate sources, hide uncertainty, leak secrets, submit schoolwork for the student, or call work done without verification.
Workflow
How it should plan, sequence, create files, update memory, ask for clarification, and verify outputs.
Taste
How you want the work to feel: direct, practical, evidence-first, concise, creative, rigorous, or more exploratory.
Design principle
Specific enough to guide. Open enough to surprise you.
Harness design is an art. If the harness is too vague, the AI starts cold and makes too many assumptions. If the harness is too bloated, the AI spends too much attention carrying rules that may not matter for the task.
Too vague
The AI forgets your standards, repeats old mistakes, asks the same setup questions, and gives generic answers.
Too rigid
The AI follows the file mechanically, over-plans simple tasks, and loses some creative range.
Just right
The AI knows your goals, respects your constraints, and still has room to choose the right method for the moment.
Token cost
A longer harness takes up more tokens.
Tokens are the pieces of text the model reads and writes. Your harness is part of the context the model may need to read at the beginning of a task. The longer the harness, the more of the model's context window and attention it can consume.
Why that matters
- Long harnesses can leave less room for project files, research notes, transcripts, datasets, or source material.
- If a model has a smaller context window, a large harness can crowd out the actual assignment, sources, or code.
- They can make every session slightly heavier, especially on free or low-limit plans.
- They can make the AI over-weight old instructions when a simple task only needs a light touch.
- They can also be worth it when the project is high-stakes, multi-session, or needs strict integrity and verification rules.
When token cost matters more
Free plans, short context windows, heavy source uploads, long PDFs, transcripts, datasets, or tasks where every bit of context space matters.
When token cost matters less
Paid subscriptions, smaller project files, repeated workflows where the harness saves time, and serious projects where consistency is more valuable than saving a few tokens.
How to improve it
Let Codex help you edit the harness.
You do not need to become a prompt engineer in the abstract. You can improve the harness by showing Codex better examples and asking it to recommend changes.
Find a useful idea
Maybe it is a public CLAUDE.md, a GitHub repo, a YouTube explanation, or someone else's agent workflow.
Paste the URL or excerpt
Ask Codex to review it and identify which ideas might improve your current AGENTS.md.
Ask for recommendations first
Do not blindly merge. Ask what to add, what to ignore, what is too advanced, and what may increase token cost.
Approve intentionally
You choose what changes. The harness should represent your workflow, not every clever trick on the internet.
Version before editing
Save the old file before changing it. If the new harness makes the AI worse, roll back.
Prompt template
Use this when reviewing someone else's harness.
Read my current AGENTS.md. Then review this external harness idea: [paste URL or excerpt] Do not edit my harness yet. First, give me: 1. what this external harness does well, 2. what would actually improve my current workflow, 3. what would add unnecessary token bloat, 4. what might reduce creativity or flexibility, 5. your recommended changes ranked by value. After I approve, propose a clean patch.
Pruning
Review the harness every few months.
Jay finds himself pruning his harness about every three months. That is the right instinct. A harness should evolve as the student gets better, the tools change, and repeated mistakes become clear.
What to prune
- Rules you no longer need because the habit is automatic.
- Specialist agents you never actually call.
- Instructions that make every task too formal or too slow.
- Old project assumptions that no longer match the student's goals.
- Repeated wording that says the same thing three different ways.
Academic integrity
The harness should protect the student.
A strong harness does not make cheating easier. It makes ownership clearer. It should remind the AI to preserve student judgment, disclose uncertainty, verify sources, avoid fabricated citations, and never produce work the student cannot explain.