Loading repository data…
lilacmohr / repository
Multi-agent code review prompts and personas for engineering teams, with a lightweight Python layer for composing and invoking reviews from PR number and template. Part of the MOHR AI Enablement Framework.
A structured framework for building software with AI agents. Three chapters, each covering a distinct phase of the development lifecycle — from spec to setup to the repeating delivery loop.
Built and validated on ai-radar, a personal AI news briefing pipeline developed entirely with this workflow.
Review a spec before implementation begins. Send multiple independent specialized reviewers, each through a different lens, then synthesize into a prioritized action list with a Decision Register.
Most AI code failures trace to a spec that was ambiguous or incomplete. This is the cheapest place to find and fix those gaps.
→ Start here if you have a spec (PR, doc, or design) and want to find holes before you write code.
Infrastructure that makes the delivery cycle reliable: an agent briefing document (CLAUDE.md), enforcement hooks, issue templates, and bot account governance.
The agent briefing document is the most underrated artifact in AI-assisted development. A well-written CLAUDE.md is the difference between an agent that needs constant supervision and one that can be trusted with a ticket.
→ Start here if you're beginning a new project or onboarding a project to AI-assisted development for the first time.
The red-green-retro loop: TEST ticket (write failing tests) → IMPL ticket (make them pass) → retro (update setup artifacts). Repeat.
Includes the test PR review skill (asking "what tests are missing?" before the IMPL ticket opens) and the implementation review skill (asking "is this the best implementation?" after the IMPL ticket closes).
→ Start here if you're already set up and want to run a repeatable delivery loop with AI agents.
Capture the why behind engineering decisions made during AI-assisted sessions — before the session closes and the reasoning evaporates. Includes a three-level decision granularity model, a narration-first capture workflow, a Decision Record template, and a Decision Register for tracking decisions across the project.
Also includes the cross-system probe (interrogate your mental model before touching a system you don't own) and architecture quiz (test your understanding of any system).
→ Use this when a Level 1 or Level 2 decision was made during a session, before touching a service boundary, or when onboarding to an unfamiliar system.
Intent validation: confirming the built thing matches what a human actually wanted.
Extends the delivery cycle with a four-layer pre-flight pipeline (demo agent, adversarial reviewer, real data stress agent, synthesis agent) that runs before any human opens the application. By the time a human session begins, blocking bugs are ticketed and what remains is a short brief of judgment calls only a human can make.
Includes the [CRITERIA] ticket type (define acceptance criteria before tests are
written), the [ACCEPTANCE-PREP] pipeline ticket, issue templates for all three
new ticket types, and a scorecard for tracking pipeline improvement over time.
→ Start here for any customer-facing ticket that needs more than tests: a UI
feature, a customer-visible report, an auth flow, or any feature where "does this
work?" is not the same question as "is this what we wanted?"
If you're new to the playbook: start with 02-Setup. The spec review and delivery cycle both depend on having a good CLAUDE.md and governance structure.
If you're mid-project and something feels off: run a spec review on your current design doc or a recent PR. The Ambiguity Auditor often surfaces the root cause.
If you're running issues and want structure: read the delivery cycle. The red-green-retro framing makes the loop explicit and gives the retro a concrete output.
02-setup/examples/ai-radar/ contains the full CLAUDE.md and playbook notes from the
ai-radar project — a real project built with this framework. Use it as a template and
adapt to your stack.
This framework is designed to grow with your team:
| Level | What it looks like |
|---|---|
| Level 1 | Manual copy-paste. Validated prompts, consistent structure. |
| Level 2 | Slash commands: /review-pr 42 architect. Prompt construction automated. |
| Level 3 | GitHub Actions trigger on PR label. Fully headless. |
The framework is currently at Level 1 with stubs for Level 2 slash commands.
Selected from shared topics, language and repository description—not editorial ratings.
davesleal /
Claude-powered Slack bot for multi-project dev automation. Each channel gets its own AI agent with codebase context. Quick Q&A, full coding sessions, GitHub issues, peer review, and App Store Connect monitoring — all from Slack.
JiaWeiXie /
connects Claude Code and Codex to a bounded Pi coding worker for repository-grounded analysis, planning, review, and implementation. It is designed for teams that want a second coding-agent perspective without giving that worker unrestricted shell access or ownership of Git delivery.