The Planning Problem: Why AI Coding Agents Need Better Ways to Think Out Loud

Let's be honest: most of us still plan projects the same way we did before AI agents existed. A quick note, maybe a Notion doc, a Slack thread that spirals into oblivion. Then we hand that half-baked plan to a coding agent and hope for the best.

The problem isn't that AI agents are bad at coding. It's that they need structure we rarely provide—and when they fill those gaps themselves, we have no record of why certain decisions were made.

The Document That Talks Back

This is the gap inplan is trying to fill. At its core, it's a Markdown editor designed for collaborative planning between humans and AI coding agents. Think of it as a shared whiteboard where both parties can draft, question, and revise a single planning document together.

What makes it interesting isn't the editor itself—it's the comment system. You can highlight any sentence and start a threaded discussion anchored right there. Ask why the agent proposed that architecture. Request clarification on an assumption. Make a decision. That exchange stays attached to that specific line, creating a durable record you can actually navigate later.

This addresses something we rarely talk about: the contextual archaeology problem. Fast-forward three months. Your agent made a choice somewhere. Where's the reasoning? Inplan makes that traceable.

Review Before Revision

Here's a pattern we see too often: you leave an AI agent working, come back, and the plan has silently transformed. Maybe it's better. Maybe something crucial changed. Hard to say.

Inplan flips this. When the agent wants to revise the plan, it doesn't quietly overwrite the document—it generates a diff you review and apply yourself. You see exactly what changed. You can approve, reject, or iterate before anything gets locked in.

This human-in-the-loop approach matters more than it might seem. It transforms the agent from an executor into a collaborator with a paper trail.

Getting Started

The tool installs as a Claude Code skill (with support for Codex and others on the roadmap). A simple "plan X" prompt opens the editor, and you're off. It's AGPL-3.0 open-source and runs entirely local via npm.

For teams wanting live collaboration, there's a hosted option at inplan.ai. Nothing in the demo requires creating an account, which is refreshing for a SaaS product to offer.

What's Still Rough

No tool is without tradeoffs. The current version is an Electron desktop app—functional but heavier than a web editor. The diff mechanism works line-by-line, which can feel coarse for fine-grained changes. And the mental model is built around one plan document at a time. If you're managing multiple interconnected specs, you'll be reaching for other tools alongside it.

The team has been transparent about these limitations, which is encouraging. Multi-document support and richer agent interaction loops are on the roadmap.

Why This Matters for the Vibe Coding Era

"Vibe coding"—letting AI handle the implementation while you steer the vision—is exploding in popularity. But vibe coding has a dirty secret: without clear direction, you're vibe-checking outcomes against intentions you never fully articulated.

Tools like inplan suggest a different future. Not just "code faster" but "think clearer." The planning phase becomes a genuine collaboration with your AI partner, producing artifacts that serve both the immediate task and future maintainers (including future-you).

Whether inplan specifically becomes your workflow depends on your tolerance for Electron apps and single-doc mental models. But the underlying concept—structured, commentable, human-reviewed AI collaboration—points toward how we'll work with coding agents as they become more capable.

The question isn't whether AI will write more of our code. It's whether we'll build the scaffolding to make that code actually understandable.

Try it: npm i -g inplan