Skip to content

AGENTS.md

Latest commit

 

History

History
68 lines (49 loc) · 3 KB

AGENTS.md

File metadata and controls

68 lines (49 loc) · 3 KB

Agent Instructions

These instructions are the canonical harness-agnostic rules for this repository. Tool-specific adapters such as CLAUDE.md must defer to this file instead of duplicating policy.

Repository purpose

This repository accompanies the paper "A Phased Workflow for Operating LLM-Based Coding Agents". It contains workflow documentation, reusable skills, a workflow figure, and an optional autonomous implementation kit.

Interaction rules

  • Ask clarifying questions before making design choices.
  • Ask questions one at a time when the decision affects structure, wording, or workflow semantics.
  • Provide multiple-choice options, your top pick, and a short justification.
  • Be direct and concise.
  • Do not assume private context that is not in the repository or explicitly provided by the user.

Workflow principles

The workflow has four phases:

  1. Research
  2. Planning
  3. Task Definition
  4. Implementation

Task Definition and Implementation form a loop. If implementation exposes a bad task, return to task definition instead of patching around the issue in code.

Keep these concepts explicit in docs and examples:

  • Human effort is front-loaded.
  • Agent delegation increases as artifacts mature.
  • Context management applies across phases: write, select, compress, isolate.
  • Persistent artifacts maintain continuity: RESEARCH.md, PLAN.md, TASKS.md, ACTIVITY.md, and git commits.
  • Manual supervision and autonomous outer-loop execution are both valid implementation modes.

Canonical files

  • AGENTS.md is the canonical instruction file.
  • CLAUDE.md is a thin Claude Code adapter.
  • Root skills/ is the canonical skills directory.
  • .claude/skills/ should point to root skills/ to avoid duplicated skill content.
  • autonomous-implementation/AGENTS.md is canonical for the autonomous implementation kit.
  • autonomous-implementation/CLAUDE.md is a thin Claude Code adapter for that kit.

Safety and privacy

  • Do not read, print, commit, or persist secrets, tokens, private keys, browser session data, or credential stores.
  • Do not include personal/internal endpoints, private model names, private marketplaces, or local machine settings.
  • Do not read .env, .env.*, .secrets, secrets/**, or local override files.
  • Keep example settings sanitized.
  • Do not run or install Claude Code unless the user explicitly asks.
  • Do not execute autonomous-implementation/loop.sh unless the user explicitly asks.

Editing rules

  • Prefer simple docs and templates over extra machinery.
  • Keep the repo agent-agnostic first.
  • Remove duplicated content when possible.
  • Keep skill names stable unless the name is misleading.
  • Use concrete file paths and artifact names.
  • Avoid PRD-style filler in task definitions.

Writing style

  • Use plain technical prose.
  • Avoid inflated claims and marketing language.
  • Avoid unnecessary bolding, emojis, and decorative comments.
  • Avoid AI-writing tropes such as "delve", "leverage" as a verb, "robust", "landscape", "serves as", and repeated em-dash reframes.