Planning Workflow

Planning Workflow — The Foundation of Agentic Development

Core Philosophy: "Planning tokens are a lot fewer and cheaper than implementation tokens."

The models are far smarter when reasoning about a detailed plan that fits within their context window. This is the key insight behind spending 80%+ of time on planning.

Outcome — When This Skill Has Delivered

You have a usable plan when all of the following hold:

• The plan is self-contained: a fresh agent who has never seen the project can read it and start implementing without asking the human for clarification.
• It is dependency-aware: every non-trivial task names what blocks it and what it unblocks, so the work decomposes cleanly into a beads graph.
• It is justified: every architectural choice and every non-obvious feature has at least one sentence on why, not just what — future agents need the rationale to make consistent local decisions.
• It has survived at least 4 review rounds by a strong reasoning model (GPT Pro Extended Reasoning is the proven choice) and reached steady-state — i.e., the most recent round produces marginal rather than structural revisions.
• It has been converted to beads with the dependency graph intact, so implementation agents can pick up ready work via br ready --json without re-reading the plan.

You have NOT delivered if any of these is true: the plan is < ~1,500 lines for a non-trivial project (under-specified); the plan is > ~10,000 lines with no decomposition (impossible to act on); the human is still being asked "what should this do?" mid-implementation (planning was abandoned, not completed); beads exist but have no dependency edges (the plan's structure was lost in conversion).

When NOT to Use This Skill

Reach for something else if:

• The change is small and local (one bug fix, one file, < ~200 LOC) → plan in chat or as a one-line TaskCreate; the planning overhead exceeds the implementation cost.
• You are doing pure research (investigating an unknown codebase, prototyping an idea) → use codebase-archaeology or idea-wizard first; planning is for execution, not discovery.
• The architecture is dictated (you are porting an existing system, following a spec, or implementing a well-defined RFC) → use porting-to-rust or testing-conformance-harnesses; the plan is the spec itself.
• You're under a hard deadline that doesn't permit 80%-on-planning (live incident, hotfix, security patch) → ship the fix, then retrofit the plan if the area needs further work.

Grounding — Sources of Truth for Plan Decisions

When a planning model proposes architecture, the proposal is a hypothesis. Ground every load-bearing claim in a verifiable source before letting it survive a review round:

• Library/framework choices: read the actual current docs (latest stable version, not the model's training-time snapshot). If the model says "use X for Y," verify X still exists, is maintained, and supports Y in the version you'd install.
• Existing-codebase claims: for any plan that touches an existing project, the model's understanding of the project structure is suspect. Grep, git log, or use codebase-archaeology to confirm structural claims before they shape the plan.
• Performance / scaling claims: never accept a number ("handles 10k req/s," "loads in <100ms") without a citation or a planned benchmark. Bare numbers in plans are guesses dressed as facts.
• Cost claims: verify against the provider's pricing page at planning time; pricing models change. Pin the plan to the specific tier you priced against.
• Cross-references to other skills/tools: if the plan says "use <tool> for X," confirm the tool's current contract supports X. Linkrot in plans is a quiet failure mode.

A plan that survived review without grounding is a plan that will surprise you in implementation. Cheap verification at planning time beats expensive rework after the code is half-written.

Validation Loop (between review rounds)

After each review round, before sending the plan back for another pass, run all four:

1. Self-containment check — pick the most obscure task in the plan, paste it alone into a fresh chat, and ask "is this implementable as written?" If no, expand.
2. Dependency-graph check — can you draw the DAG of tasks? Are there cycles? Are there orphans (tasks with no consumers)? Either is a planning bug.
3. Justification check — sample 5 random architectural decisions. Each must have a paragraph of why. If not, ask the planning model to add it.
4. Steady-state check — diff this round's plan against the previous round's. If the diff is large structural changes, you need another round. If it's typo-level polish, you're done.

If any of these fails, the plan is not ready for beads conversion — do another review round.

Why Planning Matters

• Measure twice, cut once — becomes "Check your plan N times, implement once"
• A very big, complex markdown plan is still shorter than a few substantive code files
• Front-loading human input in planning enables removing yourself from implementation
• The code will be written ridiculously quickly when you start enough agents with a solid plan

THE EXACT PROMPT — Plan Review (GPT Pro Extended Reasoning)

Carefully review this entire plan for me and come up with your best revisions in terms of better architecture, new features, changed features, etc. to make it better, more robust/reliable, more performant, more compelling/useful, etc. For each proposed change, give me your detailed analysis and rationale/justification for why it would make the project better along with the git-diff style change versus the original plan shown below:

<PASTE YOUR EXISTING COMPLETE PLAN HERE>

THE EXACT PROMPT — Integrate Revisions (Claude Code)

After GPT Pro finishes (may take 20-30 minutes), paste output into Claude Code:

OK, now integrate these revisions to the markdown plan in-place; use ultrathink and be meticulous. At the end, you can tell me which changes you wholeheartedly agree with, which you somewhat agree with, and which you disagree with:

```[Pasted text from GPT Pro]```

Process Overview

1. INITIAL PLAN (GPT Pro / Opus 4.7 in web app)
   └─► Explain goals, intent, workflows, tech stack

2. ITERATIVE REFINEMENT (GPT Pro Extended Reasoning)
   └─► 4-5 rounds of revision until steady-state

3. MULTI-MODEL BLENDING (Optional but recommended)
   └─► Gemini 3.1 Pro Deep Think, Grok4 Heavy, Opus 4.7
   └─► GPT Pro as final arbiter

4. CONVERT TO BEADS (Claude Code + Opus 4.7)
   └─► Self-contained tasks with dependency structure

5. POLISH BEADS (6+ rounds until steady-state)
   └─► Cross-model review, never oversimplify

What Makes a Great Plan

Essential Elements

1. Self-contained — Never need to refer back to external docs
2. Granular — Break complex features into specific subtasks
3. Dependency-aware — What blocks what?
4. Justified — Include reasoning, not just instructions
5. User-focused — How does each piece serve the end user?

Common Mistakes

1. Starting implementation too early — 3 hours of planning saves 30 hours of rework
2. Single-round review — You continue to get improvements even at round 6+
3. Not using GPT Pro — Extended Reasoning is uniquely good for this
4. Skeleton-first coding — One big comprehensive plan beats incremental coding
5. Losing context — Convert plans to beads so agents don't need the original

References