Anvil

Spec-driven project management for solo developers using AI agents.

Forge specs into issues. Strike them into code.
All state in markdown. Zero external services.

---

Install

Any Agent (recommended)

npx skills add ppazosp/anvil

# update later
npx skills update

This installs Anvil to ~/.agents/skills/anvil/ where it's automatically discovered by Claude Code, Codex, Copilot, Cursor, Gemini CLI, OpenCode, and 18+ other agents.

Claude Code (plugin mode)

Install as a plugin to get /anvil:forge, /anvil:strike, etc. in the autocomplete menu:

/plugin marketplace add ppazosp/backpack
/plugin install anvil@backpack

Or add to ~/.claude/settings.json:

{
  "extraKnownMarketplaces": {
    "backpack": {
      "source": { "source": "github", "repo": "ppazosp/backpack" }
    }
  },
  "enabledPlugins": {
    "anvil@backpack": true
  }
}

Manual

Clone into your project and point your agent's instruction file (AGENTS.md, .github/copilot-instructions.md, etc.) to the AGENTS.md file in this repo.

Quick Start

# 1. Cast: create a standalone issue
/cast myapp "Fix login redirect" --kind mend

# 2. Forge: interrogate → spec → phases → issues
/forge myapp

# 3. Inspect: see what's ready to build
/inspect myapp

# 4. Strike: implement a feature end-to-end
/strike P1-001

# 5. Mend: fix a bug with TDD
/mend 003

# 6. Quench: manually close an issue
/quench P1-002

Commands

Command	What it does
/cast <project> <title>	Create a standalone issue .md file
/forge <project>	Spec interrogation → phase breakdown → issue .md files with parallelism diagrams
/inspect <project>	Compact status dashboard with progress bars and ready-to-launch commands
/strike <issue-id>	Full autonomous implementation: explore → plan → TDD → simplify → review → merge
/mend <issue-id>	TDD bug fix: reproduce with failing test → fix → verify → close
/quench <issue-id>	Manually complete an issue after doing the work yourself

Agents

Anvil includes 4 subagent prompts in agents/. These are dispatched during /strike and /mend workflows.

Agent	Role	File
Explorer	Traces code paths, maps architecture, documents patterns	agents/explorer.md
Architect	Designs feature architecture, produces implementation blueprints	agents/architect.md
Simplifier	Refines code for clarity and consistency, preserves functionality	agents/simplifier.md
Reviewer	Reviews for bugs, silent failures, security issues, and test coverage gaps	agents/reviewer.md

On platforms that support subagents (Claude Code, Codex), these are dispatched automatically. On platforms without subagent support, the main agent reads the prompt file and follows its instructions inline.

How It Works

Issue files

Issues are markdown files with YAML frontmatter in docs/specs/<project>/issues/:

---
id: P1-001
title: Create user schema
status: todo
kind: strike
phase: 1
heat: data-model
priority: 1
blocked_by: []
---

See docs/issue-schema.md for the full schema.

Workflow

/cast
  └── Create standalone issue .md file (bug or feature)

/forge
  ├── General spec (interrogation)
  ├── Phase breakdown (dependencies, parallelism)
  ├── Phase deep-dive (requirements, ACs)
  └── Issue .md files (YAML frontmatter + body)

/inspect
  └── Parses all issue .md files → compact dashboard

/strike or /mend
  ├── Read issue .md + check blockers
  ├── Explore codebase (explorer subagent)
  ├── Design architecture (architect subagent)
  ├── Setup worktree
  ├── TDD implementation
  ├── Simplify (simplifier subagent)
  ├── Review (reviewer subagent)
  ├── Verify (evidence-based)
  └── Merge + update issue .md

/quench
  └── Append completion summary + mark done

Parallelism

Forge organizes issues into heats (parallel workstreams). Issues in different heats touch different code and can run in parallel via separate worktrees:

Phase 1: Foundation
├── data-model heat (sequential)
│   ├── P1-001: Create schema ──▶ P1-002: Add indexes
├── auth heat (independent)
│   └── P1-003: Auth middleware
└── config heat (independent)
    └── P1-004: Environment setup

Max parallelism: 3 agents

Attribution

Anvil includes code adapted from:

• feature-dev — Copyright Anthropic, Apache 2.0
• pr-review-toolkit — Copyright Anthropic, Apache 2.0
• code-simplifier — Copyright Anthropic, Apache 2.0
• superpowers — Copyright Jesse Vincent 2025, MIT

License

MIT