Skip to content

Cut agent instructions down to what the code can't tell an agent - #156

Draft
Moritz Mazetti (Radiergummi) wants to merge 1 commit into
mainfrom
prune-agent-docs
Draft

Moritz Mazetti (Radiergummi) wants to merge 1 commit into
mainfrom
prune-agent-docs

Conversation

@Radiergummi

Copy link
Copy Markdown
Member

Recent research suggests that large agent instruction files make coding agents perform worse. This PR keeps only what an agent cannot learn from the code, config or README, and turns rules into checks where the existing tooling allows it.

Summary

  • CLAUDE.md: 357 → 17 lines.
  • Kept: lock order, auth-key handling, pinned bundled binaries, the tailscale.com / Dockerfile version match, e2e run.sh replacing the host plugin, root-only tests skipping silently, the golines/nolint trap, Guards markers.

Checks

  • None added; golangci-lint, TestLocalAPIClientMatchesBundledTailscale and secmodel TestRepository already enforce the checkable rules.

Please check

  • The long design paragraphs were not compared line by line with docs/explanation/architecture.md. Detail like the drain/fetch ordering via tailscaled log lines and the iptables listing cache may only have lived in the old file.

Found, not fixed

  • CONTRIBUTING.md says make reinstall installs ghcr.io/matchory/tslink:latest; the Makefile uses tslink:latest.

🤖 Generated with Claude Code

The design notes, structure tour, CI list and style rules repeat
docs/explanation/architecture.md, docs/testing.md, CONTRIBUTING.md and
.golangci.toml. Keep only the traps and invariants.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant