Skip to content

AGENTS.md

Latest commit

 

History

History
56 lines (42 loc) · 2.42 KB

AGENTS.md

File metadata and controls

56 lines (42 loc) · 2.42 KB

Project Agent Entrypoint

This file is the cross-tool entrypoint for this repository.

Read order

  1. Start at ai_docs/START_HERE.md
  2. Use ai_docs/INDEX.md to choose the right task path
  3. Use ai_docs/reference/verification.md before claiming completion

Working rules

  • Treat ai_docs/ as the only detailed source of truth
  • Read current/ before standards/, and use reference/ for shared commands or paths
  • Keep changes minimal, typed, and explicit
  • Update docs whenever behavior, structure, scripts, adapters, or public APIs change

Execution style

Think before editing

  • State assumptions when they affect the implementation
  • If multiple interpretations exist, surface them instead of choosing silently
  • Prefer clarifying uncertainty before editing files
  • If a simpler approach exists, say so before implementing
  • Push back when warranted instead of mechanically following a weak approach
  • If something is unclear, stop, name the confusion, and ask

Simplicity first

  • Choose the smallest change that fully solves the task
  • Do not add speculative flexibility, configuration, or abstraction
  • Prefer direct fixes over framework-like restructuring
  • Do not add features beyond what was asked
  • Do not create abstractions for single-use code
  • Do not add error handling for impossible scenarios
  • If a change grows large but could be much smaller, simplify it before shipping

Surgical diffs

  • Touch only files and lines that relate to the request
  • Match existing project style and terminology
  • Do not improve adjacent code, comments, or formatting unless required by the task
  • Do not refactor things that are not broken
  • If you notice unrelated dead code, mention it instead of deleting it
  • Remove only the imports, variables, functions, or docs made obsolete by your own change
  • Every changed line should trace directly to the user's request

Goal-driven verification

  • Turn each task into a verifiable outcome
  • For non-trivial work, keep a short plan and verification path in mind before editing
  • Use ai_docs/reference/verification.md before claiming completion
  • Translate vague requests into concrete checks whenever possible
  • Prefer testable targets such as reproducing a bug, proving invalid input fails, or confirming behavior before and after a refactor
  • Keep looping until the requested outcome is verified, not just implemented