← All posts

How to plan features for Claude Code

Last updated: August 30, 2026

Claude Code is fast enough to build the wrong feature before lunch. The fix is a planning workflow: decide the requirements, review the screens, and approve a tech spec before the agent writes code, so it builds what you meant, not what it inferred.

Why Claude Code needs a plan

When you prompt an AI coding agent with a sentence, every decision you didn't make becomes a decision the model makes for you: which screens exist, how errors are handled, what the data model looks like, where the edge cases are. The agent fills those gaps with plausible guesses, and plausible is not the same as right. You find out in code review, or worse, after shipping.

Planning isn't overhead on top of the agent; it's the input the agent was missing. A real plan turns "build me a pricing page" into requirements the agent can be held to. This approach is often called spec-driven development.

The four-step planning workflow

This is the loop BuildBoard runs on your Mac. Each step is a Claude Code slash command that produces a reviewable artifact: plain Markdown and HTML files on your own disk. You review and approve each one before moving on.

1 · /create-feature

Write the requirements. Describe the feature in a sentence; the skill expands it into a plan: goals, user stories, edge cases, and what "done" means. This is where you catch "wait, that's not what I want" at the cost of editing a Markdown file.

2 · /create-mock

See the screens before they exist. The skill generates HTML mockups from the plan. Open them, judge them, reject the wrong direction while it's still a throwaway file instead of wired-up code.

3 · /create-plan

Approve the technical design. With the plan and mockups as input, the skill reads your actual codebase and writes a tech spec: data model, file-by-file changes, and an ordered build queue.

4 · /build-plan

Let the agent build as designed. Claude Code implements the feature one story at a time, with the plan, mockups, and spec in context. Code review becomes "does this match what I approved?" instead of "is this what I meant?"

Keeping the loop visible

Planning falls apart when the artifacts scatter: requirements in Notion, mockups in Figma, tech notes in a doc, and none of them in the room when you prompt. BuildBoard keeps every feature on a kanban board through five stages (plan, mock, spec, build, ship) with all of its artifacts attached and in sync. Everything lives as ordinary files under ~/BuildBoard: no cloud, no accounts, versioned with your own git.

Who this workflow is for

Solo founders, indie builders, and small teams who are already the PM, the designer, and the engineer at once. If you're shipping a real product with Claude Code, not just prototyping, the planning loop is the discipline that keeps agent speed from turning into rework.

Run this workflow with BuildBoard. A local-first planning app for Claude Code: one-time purchase on the Mac App Store, no subscription, your files stay on your disk.

Download on the Mac App Store