Plan-First Coding Agent
What it is
Plan-First Coding Agent is a workflow pattern where the initial step is not code generation but multi-agent design competition. Multiple coding agents (Codex, Claude Code, Cursor) independently produce design plans from the same requirements. A human reviews, compares, and selects the best plan — or synthesizes a hybrid — before any code is written. Execution then proceeds in defined phases with verification gates.
Why it matters
Most agent coding failures stem from rushing to implementation before requirements are clarified and architecture is agreed upon. By forcing a plan-first stage with multiple perspectives, the workflow surfaces hidden assumptions, reveals alternative approaches, and prevents costly mid-implementation pivots.
Key points
- Development workflow: requirements → multi-agent plan mode → human selection → phased execution with /goal
- Agents used simultaneously: Codex, Claude Code, Cursor in plan mode
- Complex plans split into multiple phases, each with requirements and verification methods, saved as Markdown documents
- Use /goal to feed plan files to agents for phased execution; manual review at each gate if uncertain
- Code review does not need the strongest model; GPT-5.5 level is sufficient for checking design compliance and code quality
- GPT 5.5 and Claude Opus 4.7 are not universally better than each other; actual comparison per task is required
- Plan.md as persistent checkpoint: Matt Van Horn's workflow with the Compound Engineering plugin treats plan.md as the canonical state. Parallel research agents analyze the codebase, past docs, and external best practices; the resulting plan.md includes acceptance criteria and checkboxes. If context is lost, a new session can resume from the plan file.
- Time allocation inverts: Traditional development is ~80% coding and ~20% planning. Plan-first workflows flip this to ~80% planning and ~20% coding. The human concentrates on framing and taste; the agent handles execution.
- Voice input removes transcription friction: When the receiver is a context-aware AI rather than a literal word processor, imperfect transcription is no longer a bottleneck. Spoken requirements can flow directly into plan.md.
- Parallel sessions multiply throughput: Running 4-6 independent Claude Code sessions on different tasks, each anchored to its own plan.md, lets one person advance multiple workstreams simultaneously.
Evidence across sources
| Source | Key Claim | Relevance |
|---|---|---|
| AI Briefing 2026-05-29 Morning | Three agents produce plans first, human selects best, then executes in phases | Practical workflow from builder with high engagement |
| AI Briefing 2026-06-08 Evening | Matt Van Horn's plan.md-first workflow: 80% planning, 20% coding; parallel sessions anchored to plan.md; voice input as first-class interface | Second source validating plan-first as a scalable personal workflow |
Open questions
- What is the optimal number of agents for plan competition? Two, three, or more?
- How should conflicting design recommendations be resolved when no single plan dominates?
- Does plan-first add unacceptable latency for rapid prototyping contexts?
Prompts for witness
- When have you regretted not planning before letting an agent write code?
- What criteria do you use to judge one architecture plan better than another?
- Which of your current projects would benefit from flipping the planning/coding time ratio to 80/20?
- What would it take to make plan.md the canonical checkpoint for every non-trivial change in your codebase?