Skip to content

Multi-Tool Path

Who it's for

Individuals or teams running more than one coding agent/tool and tired of divergent instructions, duplicated CLAUDE.md vs AGENTS.md drift, and "works in Cursor, fails in Claude Code."

Prerequisites

  • At least two tools in active use on the same repo
  • Permission to standardize on AGENTS.md (or equivalent) as source of truth
  • Willingness to delete tool-specific folklore that isn't load-bearing

Ordered modules

  1. CLAUDE.md & AGENTS.md — one content contract, multiple readers
  2. Context and Instruction Hierarchy — how layers compose and conflict
  3. Repository Structure — layout that every tool can navigate
  4. Docs as Agent Memory — put truth in files, not chat profiles
  5. Control Techniques — portable controls vs tool-only tricks
  6. Prompt Anatomy — structure that transfers
  7. AGENTS.md template — standardize the file
  8. Agent Handoff note — switch tools mid-task without amnesia
  9. Stuck Loops and Thrash — thrash looks different per tool; reset is the same
  10. Maintenance Prompts — keep docs true across tools

Exercises

  1. Diff CLAUDE.md vs AGENTS.md vs Cursor rules; merge into one AGENTS.md; make CLAUDE.md a pointer/symlink.
  2. Run the same feature spec through two tools; compare drive-bys and checklist failures; update standing orders to prevent the shared failure.
  3. Document which workflows are tool-specific (e.g. Claude Code plan mode, Cursor Composer) vs portable in a one-pager linked from README.
  4. Add a CI check or script that fails if CLAUDE.md and AGENTS.md diverge beyond an allowed pointer pattern.

Capstone

Deliver a multi-tool standard for one repo: single instruction source, shared templates, handoff note convention, and a short evaluation showing equivalent outcomes on a real task in two tools.

Expected outcome

Engineers can switch tools without rewriting the project's brain; instruction drift becomes a detectable defect, not a culture.

A field manual for AI-native software engineering.