HangYu8123/HarnessFlow

Harness coding workflow for codex, claude, github copilot

What it solves

HarnessFlow prevents AI coding assistants from "one-shotting" changes—making immediate edits without sufficient planning or verification. It provides a structured, multi-agent pipeline that forces the AI to analyze the codebase, challenge its own assumptions, and validate results before shipping code, reducing regressions and token waste.

How it works

It is a portable Markdown instruction pack that integrates with tools like Claude Code, Codex CLI, and GitHub Copilot. Instead of a runtime, it uses a set of request templates and workflow files stored directly in the repository.

When a user submits a filled-in template, the AI assistant follows a specific pipeline:

  1. Parallel Analysis: Multiple analyst subagents investigate the codebase from different perspectives.
  2. Synthesis: A Senior Engineer agent synthesizes these findings into a single plan.
  3. Self-Challenge: A "Devil's Advocate" pass stress-tests the plan for bad assumptions.
  4. Validation: A QA Engineer validates the implementation.
  5. Persistent Memory: Results are written to a repo_info/ directory so future requests have immediate context.

Who it’s for

Developers using AI coding assistants (specifically Claude Code, Codex CLI, and VS Code + GitHub Copilot) who want more disciplined, reliable, and token-efficient code generation and refactoring.

Highlights

  • Zero Runtime: No npm install or build steps; it is purely a set of Markdown instructions.
  • Three Workflow Modes: Offers general (thorough), fast (token-efficient), and skill (community-skill-backed) variants.
  • Nine Request Types: Specialized workflows for implementation, refactoring, debugging, codebase Q&A, correctness checks, goal-directed execution, stacked PRs, repo initialization, and iterative loops.
  • Platform Adaptive: Automatically configures entry points for different AI tools via setup scripts.
  • Proven Efficiency: Benchmarks show the fast mode can reduce code footprint and token costs by 39-50% compared to the general workflow while maintaining 100% correctness.

Related

  • Project
    zhu1090093659/spec_driven_develop

    An architecture-first workflow plugin for AI coding agents that turns large software changes into a spec-driven loop of analysis, decomposition, and adaptive execution.

  • Project
    Ancienttwo/repo-harness

    A file-backed workflow and CLI that enables AI coding agents to maintain state, follow structured plans, and provide verifiable evidence of work across sessions.

  • Project
    coleam00/skills

    A collection of modular Markdown-based skills for AI coding agents that replace bloated system prompts with task-specific procedures for planning, implementing, and validating software.

  • Project
    Chachamaru127/claude-code-harness

    A disciplined delivery loop and safety layer for AI coding agents that enforces a structured Plan-Work-Review-Release workflow to prevent agent drift.

  • Project
    waybarrios/opencode-power-pack

    A portable collection of 54 professional workflows and a secure sandbox runner for AI coding agents including Claude Code, Codex, OpenCode, and Pi.