A three-skill workflow that separates understanding from planning from doing — with context resets between each stage. Designed for ambitious coding tasks on real codebases where you can't hold the entire architecture in your head.
The core idea: quality outputs depend entirely on input quality. Instead of dumping a vague task into an agent and hoping for the best, you guide it through three distinct phases, each producing a structured artifact that carries context forward while the conversation history resets.
The agent explores the codebase systematically to understand the area you'll be changing. It documents what exists — components, integration points, patterns, file paths with line numbers — without suggesting changes.
Output: tasks/YYYY-MM-DD-description/research.md
Starting from the research document, the agent creates a phased implementation plan with concrete file paths, line numbers, and verifiable success criteria for each phase. No vague hand-waving — the plan should be detailed enough to execute mechanically.
Output: tasks/YYYY-MM-DD-description/plan.md
The agent executes the plan phase by phase, updating checkboxes as it goes. It stops at the end of each phase for your review. If you start a new session, it picks up from the first unchecked item.
Output: Updated plan.md with [x] checkmarks and a summary of changes per phase.
Each phase compresses its findings into a structured document, then you start a fresh session. This keeps the agent's context window clean (40-60% utilization) rather than letting it bloat with stale exploration history. The documents carry the signal forward; the noise stays behind.
Copy the three skill directories into your tool's skills folder:
your-project/
├── .claude/skills/ # or .cursor/skills/, .github/skills/, .codex/skills/
│ ├── research-codebase/
│ │ └── SKILL.md
│ ├── create-plan/
│ │ └── SKILL.md
│ └── implement-plan/
│ └── SKILL.md
# Phase 1: Research
/research-codebase
> "I need to add real-time notifications to the dashboard"
# Review research.md, then start a new session
# Phase 2: Plan
/create-plan
> "tasks/2026-03-30-realtime-notifications/research.md"
# Review plan.md, then start a new session
# Phase 3: Implement
/implement-plan
> "tasks/2026-03-30-realtime-notifications/plan.md"
# Review each phase, start new sessions between phases
This workflow is a modified version of the Research-Plan-Implement framework introduced by Dexter Horthy from HumanLayer. Dexter's original work on Advanced Context Engineering for Coding Agents laid the foundation — the key insight that frequent intentional compaction of context into structured artifacts is what makes AI coding agents effective on large, real-world codebases.
These skills are packaged following the Agent Skills open standard and published by computerlove.tech.