Outline-Driven Development for Claude Code

Formerly known as the ODIN Claude Plugin. The repo URL stays the same; the brand has matured.

Outline-Driven Development (nicknamed ODIN) — Advanced code agent system for Claude Code with surgical precision, diagram-first engineering, and comprehensive workflow automation.

Methodology: outline-driven-development  ·  Codex CLI: odin-codex-plugin  ·  Gemini CLI: odin-gemini-cli-extension  ·  Site: outlinedriven.github.io

Overview

ODIN is a professional-grade Claude Code plugin that transforms Claude into a sophisticated code agent with comprehensive workflow automation and rigorous engineering methodology.

Key Capabilities:

• 📐 Diagram-First Engineering - Architecture, concurrency, memory, data flow, optimization
• 🎯 Surgical Code Editing - AST-based transformations with ast-grep
• 🧠 Confidence-Driven Execution - Adaptive behavior based on complexity and risk
• 🔍 Deep Investigation - Mandatory file reading before code modifications
• 🔒 Atomic Commits - Conventional Commits protocol with incremental approvals

Installation

One repository serves every harness below. There is no build step and no converter — each tool reads skills/ directly.

Prerequisites

• Claude Code installed and running
• Git (for marketplace installation)

Full Install Script (Recommended)

wget -O ~/.claude/CLAUDE.md https://raw.githubusercontent.com/OutlineDriven/odin-claude-plugin/refs/heads/main/CLAUDE.md && claude plugin marketplace add OutlineDriven/odin-claude-plugin && claude plugin install odin@odin-marketplace

Other coding agents

| Harness | Install | Extra files needed | |---|---|---| | Codex CLI | codex plugin marketplace add OutlineDriven/odin-claude-plugin then codex plugin add odin@odin-marketplace | none — Codex reads .claude-plugin/ | | Grok Build CLI | grok plugin install OutlineDriven/odin-claude-plugin | none — Grok reads Claude Code plugins with zero configuration | | Devin CLI | devin plugins install OutlineDriven/odin-claude-plugin | .devin-plugin/ | | Antigravity (agy) | agy plugin install https://github.com/OutlineDriven/odin-claude-plugin | root plugin.json | | Cursor | /add-plugin odin in Cursor Agent chat | .cursor-plugin/ | | Kimi Code CLI | /plugins install https://github.com/OutlineDriven/odin-claude-plugin then /reload | .kimi-plugin/ | | opencode | local checkout only — see below | .opencode/plugins/odin.js |

Codex and Grok need no ODIN-specific manifest: both resolve .claude-plugin/plugin.json, and Codex additionally reads .claude-plugin/marketplace.json. .agents/plugins/marketplace.json is shipped as Codex's current (non-legacy) marketplace path.

Devin documents a .claude-plugin/ fallback from 3000.3.22, but builds before that hard-fail without .devin-plugin/plugin.json, so ODIN ships one. Devin also gates plugins behind a closed beta — request access from Cognition first. Skills surface as /odin:<skill>.

opencode documents plugin[] entries for npm packages only, so there is no one-command install from this repository. Clone it and point opencode.json at the plugin file:

{
  "plugin": ["/path/to/odin-claude-plugin/.opencode/plugins/odin.js"]
}

The plugin registers skills/ on config.skills.paths and adds one command per skill, skipping the five marked disable-model-invocation because opencode has no manual-only gate.

Verified by real installs from a clean clone of this repository at 1.17.82, on Linux. Every version below was read from the installed binary with --version, not assumed:

| Harness | Installed version | Check | Result | |---|---|---|---| | Codex CLI | codex-cli 0.146.0 | plugin marketplace add + plugin add | installed, 120 skills in plugin root | | Devin CLI | devin 3000.2.17 | devin plugins install | installed, skills exposed as /odin:<skill> | | Antigravity | agy 1.1.9 | agy plugin validate . | ok, 120 skills processed | | Grok Build | grok 0.2.118 [stable] | grok plugin validate . | manifest valid, 1 skill dir + hooks | | Claude Code | 2.1.220 | claude plugin validate | passed | | opencode | not installed | module exercised directly | 114 commands, 5 manual-only excluded |

Cursor and Kimi are not tested. Both install through an in-editor slash command rather than a shell, so there is no headless path to exercise them. Their manifests match the vendors' published schemas; treat first install as unproven.

claude plugin validate --strict fails on this repo, and did so before any harness work: metadata.maintainer, .website and .support in .claude-plugin/marketplace.json are unknown to Claude Code, which ignores them at load time. Non-strict validation passes.

Verify Installation

# List available agents
/agents

# View all commands
/help

# See the installed plugins
/plugin

Core Philosophy

ODIN follows strict engineering principles:

1. Investigate Before Acting - Never speculate about code you haven't read
2. Diagram-First Design - Five mandatory diagrams before any implementation
3. Surgical Precision - Minimal, targeted changes using AST-based tools
4. Atomic Commits - One logical change per commit, properly typed
5. Confidence-Driven - Adapt behavior based on familiarity and risk
6. Tool Selection - ast-grep > native-patch > ripgrep (never sed for edits)

Five Required Diagrams

Before any non-trivial implementation:

1. Architecture - Components, interfaces, contracts, dependencies
2. Data Flow - Sources, transformations, sinks, state transitions
3. Concurrency - Threads, synchronization, happens-before relationships
4. Memory - Ownership, lifetimes, allocation patterns, safety guarantees
5. Optimization - Bottlenecks, targets, complexity bounds, resource budgets

Skills (128 total)

Skills are invokable workflows that extend ODIN with process- and domain-specific protocols. Invoke with /<skill-name>; many also trigger on natural language cues described in their frontmatter.

Planning & Exploration (22 skills)

• askme - Verbalized Sampling protocol for deep intent exploration before planning
• batch-ask-me - Walk a dependency-aware design tree in batched question rounds until shared understanding
• wayfinder - Chart a multi-session effort into a destination, mapped fog, and decision tickets on the frontier
• mutual-sync - Three-way grounding: verify user, agent, and codebase share one picture of current state before proceeding
• loop-me - Design recurring workflows through a stateful askme session and cwd specs
• to-questionnaire - Turn a knowledge gap into an async questionnaire for the person who can answer it
• clarify - Scan a request, document, or conversation for ambiguities and unstated assumptions; certainty-tiered findings with defaults
• generalize-from-cases - Derive the rule the user means from the examples they gave, with rival readings and a stated boundary
• exhaustive - Prove a decision, state, or requirement space is fully covered by enumerating every cell as covered, gap, or deferred
• shape - Shape Up shaping: appetite, breadboard, rabbit holes, no-gos for a bet on work
• to-tickets - Break a plan into tracer-bullet tickets with blocking edges, on GitHub or in .outline/
• contexts - Coordinate context sweep before coding
• domain-modeling - Build and sharpen a project's ubiquitous language, glossary, and domain decisions as you design
• init - Analyze a codebase and create or improve AGENTS.md
• brainstorm - Explore vague or ambitious ideas into a right-sized requirements-only plan
• explore - Read-only codebase exploration to map structure, symbols, and dependencies
• strategy - Sharp interview to write or maintain STRATEGY.md as the product anchor
• ideate - Generate grounded, divergent ideas from the codebase into docs/ideation
• design - Set visual and interaction direction for UI surfaces before writing code
• prototype-logic - Throwaway single-file HTML demo that answers one question about a state model
• pov - Decisive, project-grounded verdict on adopting or switching technology
• research - Gather external knowledge from authoritative sources with verified citations

Writing & Learning (10 skills)

• teach - Run a persistent cwd teaching workspace, and route to the corpus skills when the material already exists
• map-corpus - Inventory a folder of your own study material into one CORPUS.md with prerequisite-ordered concepts
• explain-concept - Make one concept clear from a chosen angle: intuition, motivation, origin, picture, or contrast
• drill - Practise a concept and get graded: scaffolded exercises, quizzes, spaced recall, and gap probes
• capstone - Scope and judge a real project sized to what the learner has actually cleared
• writing-fragments - Mine heterogeneous raw material into an append-only pile before shaping
• writing-shape - Shape a raw pile into an article paragraph by paragraph with argued formats
• writing-beats - Build a grounded article journey one candidate beat at a time
• writing-skills - Reference for writing agent-consumed documents so they run predictably: context load, hierarchy, leading words, and pruning
• book-to-skill - Distil a book or comparable source into a validated, trigger-probed agent skill

Working Posture (7 skills)

• duet - Two-party posture: user as director, agent as executor. Surfaces every fork via AskUserQuestion with structural framing and a recommended default. Eliminates the review-bottleneck and prevents codebase-understanding debt. Pair with the Duet output style.
• axiom-mode - Compact formal-logic English register using predicate claims and ASCII keywords
• ai-collab-protocols - Surface in-task AI collaboration protocols one tactic at a time
• taste - Apply distinctive judgment to prose, code, design, or decisions instead of AI mediocrity
• do-it-now - Single-pass posture: ship the whole ask now, with no phases, stubs, or follow-up remainders
• necessary-work - Gate every candidate action on the delete test: outcome unmet or unproven, or the action is rejected
• wait-what - Re-pitch a message that did not land, with context, plain-language phrasing, and the project's own vocabulary

Engineering Methodologies (16 skills)

• test-driven - TDD with 10-language support
• type-driven - Type-driven development (Idris 2, with 10-language support)
• proof-driven - Proof-driven development (Lean 4, with property-based testing fallback and 10-language support)
• contract-driven - Design-by-Contract (DbC): pre/post/invariants at API boundaries, state invariants, and trust boundaries
• validation-first-driven - State machines, invariants, and temporal properties (Quint spec) before implementation
• tests-purge-unneeded - Delete tests that don't catch real bugs (the inverse of TDD); load-bearing principles, static-vs-dynamic carve-out, language-specific examples
• spec-driven - Write a structured spec before writing code
• source-driven - Ground implementation decisions in official documentation with version-aware citations
• ground-latest - Pin versions, stacks, and practices from today's release channels instead of from recall
• doubt-driven - Subject non-trivial decisions to adversarial review before they stand
• minimalism-driven - Minimalism as enforced doctrine: null-solution start, need-gated additions, delete > edit > add as authoring-time gates
• verification-before-completion - Require fresh, fully-read run evidence before any completion claim
• lighter-checks - Size verificati