haft

Agent Prompt Setup

haft works out of the box — the MCP tools are always available once you run init. But adding a project-level prompt file makes the agent significantly better at using haft proactively.

What is a project prompt?

Each AI tool reads a specific file from your project root at the start of every session. This file contains project-specific instructions that shape the agent's behavior:

Tool Prompt file
Claude CodeCLAUDE.md
Cursor.cursorrules
Codex CLIAGENTS.md
Gemini CLIGEMINI.md

The haft repository includes a reference CLAUDE.md that teaches the agent:

  • Communication style — direct, no validation theater, peer-level technical discourse
  • Thinking principles — weakest link analysis, reversibility checks, separation of concerns
  • When to use haft — the decision framework for quick choices vs. full FPF mode for complex ones
  • FPF commands and workflow — the recommended protocol (frame, explore, compare, decide, verify)
  • Key concepts — R_eff, evidence decay, CL penalties, claims, Transformer Mandate
  • Glossary — quick reference for FPF terminology

Install it

Download the reference file and rename it for your tool:

# Claude Code
curl -fsSL https://raw.githubusercontent.com/m0n0x41d/quint-code/main/CLAUDE.md > CLAUDE.md

# Cursor
curl -fsSL https://raw.githubusercontent.com/m0n0x41d/quint-code/main/CLAUDE.md > .cursorrules

# Codex CLI
curl -fsSL https://raw.githubusercontent.com/m0n0x41d/quint-code/main/CLAUDE.md > AGENTS.md

# Gemini CLI
curl -fsSL https://raw.githubusercontent.com/m0n0x41d/quint-code/main/CLAUDE.md > GEMINI.md

Adapt it to your project. The reference file is opinionated (no emoji, direct communication, functional core / imperative shell). Keep the haft-specific sections (FPF Mode, Commands, Glossary) and adjust the rest to match your team's style.

What it changes

Without CLAUDE.md:

  • Agent uses haft tools when you explicitly call slash commands
  • No proactive decision capture
  • No structured reasoning unless you ask for it

With the prompt file:

  • Agent understands when to suggest /h-reason for complex tasks
  • Agent records micro-decisions as notes when it notices them in conversation
  • Agent checks /h-status proactively to surface stale decisions
  • Agent applies weakest link analysis and reversibility checks naturally
  • Agent follows the recommended FPF protocol for non-trivial decisions

Keeping it up to date

The reference prompt evolves with haft releases. When you update the binary, check if the reference file has changed: 

# Compare your version with the latest
curl -fsSL https://raw.githubusercontent.com/m0n0x41d/quint-code/main/CLAUDE.md | diff - CLAUDE.md

The haft-specific sections (FPF Mode, Commands table, Glossary) are the most important to keep current. The general engineering style sections can stay customized.

Next step

How /h-reason works — learn the primary workflow.