Docs — Manifesto
A model can read your code, search it, and reason about it; it cannot know your business, your conventions, or where your authority boundaries sit.
The Agent Manifesto is a soft harness for capable models that supplies the smallest surrounding system for that gap — context, policies, procedures, and boundaries — and it is maintained by subtraction as well as addition. It is a specification plus a small plugin for Claude Code and Codex. The framework must improve your work without making you operate the framework.
Eight principles that allow you to build the proper Soft Harness for your project: maintainable, compact, modern.
01
Record business facts, project realities, preferences, authoritative sources, and authority boundaries. Do not explain practice a capable model already understands.
Ask of every instruction:
Would a better model make this unnecessary?
02
Always-loaded files are a starting brief and a map. Task policies, procedures, specialist responsibilities, and detailed facts load only when they are relevant.
The move that is not a fix:
Splitting a large file changes nothing if every part still loads at startup.
03
State what must be true, what must not happen, and what evidence counts. Leave local reasoning and implementation choices to the model.
Where the line falls:
Say what done looks like, not how to think.
04
Create a skill, workflow, agent, contract, or context file only for a present need. Repetition justifies structure; anticipation does not.
Before adding a layer:
What need does this artifact answer today?
05
Prefer each tool's own instruction discovery, skills, agents, permissions, hooks, and configuration. Portability is shared intent across native forms.
What portability is not:
Forcing every tool into one fictional runtime.
06
Prompt instructions influence behaviour; they do not guarantee it. Non-negotiable restrictions belong in permissions, hooks, schemas, tests, and linters.
The distinction that matters:
A rule that must hold cannot be a sentence asking nicely.
07
Ask before destructive, external, costly, or authority-expanding actions. Consent covers the action approved and nothing adjacent to it.
Never inferred:
Permission to commit, push, deploy, publish, or restructure.
08
Models and tools change. Compare the active landscape with a fresh minimal baseline after a significant change, or at least every six months.
The habit this protects:
The framework is maintained by subtraction as well as addition.
A proper landscape lets you describe the work in ordinary language while the system quietly supplies the relevant context, policy, procedure, isolation, contracts, and enforcement. If operating the landscape becomes a second job, simplify it.
