Writing Autonomous Instructions That Agents Steelman and Revise
Writing Autonomous Instructions That Agents Steelman and Revise
I write everything as a CLAUDE.md spec now and run 5 agents in parallel off it. The context pollution thing is real - and specs are the antidote.
The Context Pollution Problem
When you work with an AI agent through conversation, every message adds to the context. Corrections, tangents, clarifications, and abandoned approaches all sit in the context window alongside your actual instructions. By the time you reach the interesting part of the task, the agent is working with a noisy mix of current intent and outdated discussion.
Specs eliminate this. A CLAUDE.md file contains only the current, correct instructions. No history of wrong approaches. No "actually, ignore what I said earlier." Just the spec.
How to Write Specs That Work
Good autonomous specs have three properties. They are complete enough that an agent can start working without asking questions. They are specific enough that two agents reading the same spec would produce similar results. And they are structured enough that the agent can reference specific sections without re-reading everything.
Use headers for different concerns. Put constraints and requirements up front. Include examples of expected inputs and outputs. Define what "done" looks like.
The Steelman and Revise Pattern
The best specs include a directive telling the agent to challenge assumptions before implementing. Instead of blindly following instructions, the agent identifies potential issues - missing edge cases, conflicting requirements, ambiguous language - and flags them before writing code. This turns the agent from an executor into a collaborator that improves the spec as it works.
Running Parallel Agents
Once your spec is solid, you can hand it to multiple agents working on different parts of the same project. Each agent reads the same source of truth. No one inherits another agent's conversational context. Changes to the spec propagate to all agents on their next read.
This is how you scale from one agent to five without the coordination overhead scaling linearly.
- CLAUDE.md for Parallel Agents Guide
- Skill MD Per Folder for Parallel Agent Isolation
- Managing Parallel Claude Agents and Tracking
Fazm is an open source macOS AI agent. Open source on GitHub.