More instructions make your agents worse.

The Lean Context Standard is a structural approach to human-agent collaboration. Stop stuffing the context window. Start curating it.

the problem

End the junk drawer

Every mistake triggers more instructions. More instructions dilute focus. A different mistake happens. The loop repeats.

Read more
CONSTITUTION MANIFEST always loaded always visible on demand
the framework

Lean architecture for AI

A constitution for hard stops. A manifest to route context. A supply pulled on demand. Three tiers. One contract.

Read more
curate reason trust delegate amplify
the human impact

The amplification loop

Fear is replaced by discipline. Trust builds. The human delegates more. Both sides get smarter every session.

Read more
adopt the standard

Start here.

The Lean Context Standard works with any agent, any repo, any tool. Three files. Five minutes to start. A lifetime to master.

New repository

Start clean

Create CONSTITUTION.md with your hard stops. Create MANIFEST.md to route tasks to context. Create .context/ and add your first supply document. The agent reads the structure on its own.

Existing repository

Retrofit in place

Already have a CLAUDE.md, AGENTS.md, or CURSORRULES? Audit it. Extract hard stops into a constitution. Move reference material into supply documents. Replace the monolith with three files.

Prompting

Tell the agent

Symlink your tool's expected file to the constitution: ln -s CONSTITUTION.md CLAUDE.md for Claude Code, AGENTS.md for Copilot, .cursorrules for Cursor. The agent finds the constitution automatically. Add one line: "Read MANIFEST.md before starting. Pull supply documents when the task matches a route. Stop when uncertain."

your-repo/
├── CONSTITUTION.md      ← hard stops, always loaded
├── MANIFEST.md          ← routing table, always visible
├── CLAUDE.md            → CONSTITUTION.md (symlink)
├── AGENTS.md            → CONSTITUTION.md (symlink)
├── .context/            ← supply, pulled on demand
│   ├── auth-architecture.md
│   ├── network-standards.md
│   └── deploy-procedures.md
└── src/
our principles

The workbench rule

derived from lean manufacturing · 5S

The context window is a workbench. Only what the current operation needs should be on it. Everything else is inventory waste. The constitution is the sorted workspace. The manifest is the shadow board. The supply is just-in-time delivery.

Pull, not push

derived from just-in-time delivery

Context loads when the task demands it, not when the session starts. Push is overproduction. Pull is lean. Pull also makes the agent's reasoning visible — the human can trace every decision back to the context that informed it.

Stop when uncertain

derived from the andon cord

Any worker can stop the production line when something is wrong. When the agent stops instead of guessing, the gap gets filled. The system gets smarter. The act of stopping is the system working, not failing.

Invisible knowledge only

the knowledge boundary

The agent can read the code — file layout, conventions, dependencies. That is inferable knowledge. Write down what it cannot discover: why decisions were made, what was tried and rejected, what constraints the code does not encode.

the evidence
−3%
Success rate with LLM-generated context files
+20%
Inference cost increase from untargeted context
+4%
Marginal gain from human-written files (with cost tradeoff)

ETH Zurich · arXiv:2602.11988 · 138 task instances · 12 repositories · 2026

Read the full thesis

The complete argument — the degradation loop, the lean manufacturing parallel, the four rules, boundary conditions, calibration, and the force multiplier.

The Lean Context Thesis →