agentic-engineering: Give Coding Agents a Documentation-First Workflow

Overview

agentic-engineering is a documentation-first development harness built from shared SKILL.md playbooks for Claude Code, Codex, and other compatible coding agents. It addresses the statelessness of coding-agent sessions by making project knowledge, requirements, decisions, progress, and verification evidence live in repository documents rather than in a conversation.

The workflow connects project setup or brownfield onboarding to planning, implementation, and review. It is aimed at developers and teams using coding agents across multiple sessions, especially when work needs explicit contracts and a separate verifier. Its loop driver can advance tasks and prepare pull requests, but leaves merging to a human.

Key Features

• Ten skills cover agent instructions, PRDs, technical specifications, ADRs, implementation plans, verification, status tracking, Git workflow, documentation maintenance, and existing-project onboarding.
• Uses AGENTS.md as a shared instruction entry point, with tool-specific instruction files pointing to it.
• Specifies acceptance criteria with verification commands and evidence, and keeps approved specifications versioned.
• Records architecture decisions append-only, with trade-offs and conditions for reconsidering them.
• Tracks in-flight work and session handoffs in docs/status.md for recovery between sessions.
• Separates implementation from verification: a fresh-context verifier returns evidence-backed PASS or FAIL results.
• Provides a loop driver that creates task branches and draft PRs, and stops after repeated verification failures; it does not merge.
• Onboards existing codebases by reconstructing their current documentation chain without changing source code.

Use Cases

• A solo developer starting a project with an AI coding agent can establish shared instructions, define requirements, and plan work before implementation.
• A team running work across multiple agent sessions can use the status tracker and checkpoints to resume without relying on prior conversation context.
• A team automating task execution can route code through verification and draft PRs while retaining human review and merge control.
• A maintainer of an existing codebase can document its as-built architecture and plan future work without asking onboarding to modify code.

Getting Started

For Claude Code, add the marketplace and install the plugin:

/plugin marketplace add affectionatec/agentic-engineering
/plugin install agentic-engineering@agentic-engineering

Then use /using-agentic-engineering in a project to assess its documentation and choose a starting skill. See the README for manual installation and Codex or other-agent setup.

Alternatives

• skills: Provides composable practices such as requirements clarification and TDD, rather than a documentation-first chain for long-running work.
• claude-code-best-practice: Offers a curated guide to Claude Code workflows and configuration, rather than a portable set of linked skills with persistent project records.
• context-engineering: Demonstrates decision retrieval and output checks in a Python implementation, rather than a multi-skill workflow harness for coding agents.

Considerations

• The repository is quiet: it has had no commits in the last 90 days and has no release in a year. Teams adopting it should review the current skills and decide how they will maintain or adapt them.
• The project has two contributors, with contributions concentrated mostly in one person. That makes continuity of maintenance worth considering.
• Skills are markdown playbooks, not a standalone runtime. Use outside Claude Code and Codex may require configuring the target agent to load the skills or use them as rules.
• The workflow is intentionally gated: verification PASS is required before a PR becomes ready, and a human must review and merge it. This is a safeguard, but it means the loop is not fully autonomous.
• The README describes a verification test ratchet, but does not list a repository test suite or CI setup. Teams should assess the skills against their own project and verification commands.