5 CLAUDE.md Rules That Made My AI Stop Asking and Start Doing

Source: Dev.to

After months of running Claude Code autonomously, I’ve learned that most interruptions aren’t the AI’s fault—they’re the CLAUDE.md’s fault. Here are five rules that eliminated most of my “should I do X?” questions.

1. The irreversibility rule (not the uncertainty rule)

What most CLAUDE.mds say

“Ask for clarification when uncertain.”

What actually works

“Ask only for: irreversible actions, external credentials, external visibility (publishing, sending emails), costs beyond the subscription.”

The difference is significant. Uncertainty is constant—every decision has unknowns. Irreversibility is rare—most code changes can be reverted with git reset. Switching from “ask when uncertain” to “ask only when irreversible” dropped the question count by about 80 %.

2. Explicit decision framework for common forks

What most CLAUDE.mds say

“Use your judgment.”

What actually works

| Situation                     | Action                                 |
|-------------------------------|----------------------------------------|
| Technical approach unclear   | Choose the conventional approach       |
| Two valid implementations     | Choose the simpler one                 |
| Error after 3 attempts        | Document in `blocked.md`, switch tasks |
| Ambiguous requirement         | Apply most reasonable interpretation, document assumption |

The AI already has good judgment. The table isn’t teaching it to think—it eliminates the meta‑question “should I ask or should I decide?” By stating explicitly that “unclear approach → conventional approach,” that decision is pre‑made.

3. The pending_for_human.md pattern

What stops most workflows

A step that requires browser authentication or a credential I don’t have. Without a clear protocol, the AI either loops on it or silently skips it.

What works

“If blocked: document blocker in pending_for_human.md, switch to next available task.”

The format I use:

## 2026-02-28: Smithery API key needed
**Why blocked**: Smithery MCP registry requires authentication
**Steps for human**:
1. Go to https://smithery.ai/account/api-keys
2. Create API key, add to `~/.credentials`
**Everything else done**: `smithery.yaml` added to repo, package published to npm

The key is the last line: “everything else done.” The human opens this file in the morning and sees exactly one click they need to do.

4. Explicit syntax check commands

What most CLAUDE.mds say

“Make sure the code works.”

What actually works

“After every edit:
• Python → python -m py_compile <file>
• TypeScript → tsc --noEmit
Do not move to the next file before fixing errors in the current one.”

The “do not move to the next file” part is what matters. Without it, the AI writes several files, runs the build, finds scattered errors, and has to context‑switch back to each one. With it, each file is clean before the next one starts. This is the same reason you do git add -p instead of git add ..

5. Context compression protocol

What kills autonomous sessions

Running out of the context window mid‑task and having to start over.

What works

“When context gets large: write current state to tasks/mission.md. Include what’s done, what’s next, what’s blocked, and any open questions. The next session should be able to continue from tasks/mission.md without reading the full history.”

mission.md should be written as if explaining to someone who knows the project but wasn’t in the room. Not a transcript, not a summary—a briefing. This creates a natural “checkpoint” that makes context compaction safe.

What these have in common

All five rules do the same thing: convert vague expectations into specific protocols. “Use judgment” forces a meta‑decision; “choose the simpler one” eliminates it. Every meta‑decision costs context window and interruption. Eliminating meta‑decisions is what makes autonomous operation actually work.

The hooks and templates I run in production are open source: claude-code-hooks — 16 hooks + 6 templates from 700+ hours of autonomous operation.

📖 Claude Code Production Guide (¥800) — lessons from 700+ hours of autonomous operation.

Is your Claude Code setup actually safe? Run npx cc-health-check — a free 20‑point diagnostic. Score below 80? The Claude Code Ops Kit fixes everything in one command.

What’s in your CLAUDE.md that you wish you’d added earlier?