Planning with Claude Code

Planning is where results are actually determined

Most operators treat planning as a formality. They type a request, Claude starts working, and they cross their fingers hoping it heads in the right direction. When it doesn't, they wait for it to finish, then try to correct a deliverable that's already gone sideways.

The operators who get consistently better results do something different. They confirm the implementation details before a single file gets touched. That one habit compounds over every session, every client engagement.

How plan mode works

Claude Code has three operating modes. You cycle through them with Shift+Tab.

• Regular mode — Claude proposes edits one at a time and waits for your approval
• Auto-accept mode — Claude makes all changes without asking (prompt turns purple)
• Plan mode — Claude reads, researches, and proposes. It cannot edit or create files

Plan mode is your confirmation layer. Claude reads the relevant files, thinks through the approach, and writes out a structured plan. You review it, push back on anything that looks off, and iterate until you're aligned. Then you switch back to regular mode and execute.

[Shift+Tab × 2 → Plan mode active]

You: Review the Q3 client report files and outline
     how you'd restructure the executive summary.

Reading q3-report.md...
Reading client-brief-q3.md...

Plan:
1. Shorten the opening paragraph to 3 sentences
2. Move KPIs to a highlighted table above the fold
3. Rewrite recommendations section using client's
   stated priorities from the brief
4. ...

Does this approach look right? [y/n or feedback]

The value is in the iteration, not the plan itself. You might read Claude's plan and realize it's pulling from the wrong document, or structuring the deliverable in a way that doesn't match the client's format. Catching that before execution costs you two minutes. Catching it after costs you thirty.

> Hard rule. For any client-facing deliverable, start in plan mode. Confirm the approach. Then execute in regular mode with approval at each step. This two-pass workflow is the single most reliable way to maintain quality across multiple client engagements.

Claude's internal to-do list

Once you confirm a plan, Claude generates its own internal to-do list for the task. You'll see it appear as it works, ticking off items as each one is completed.

This system is more reliable than it looks. Claude tracks its own progress, picks up cleanly after interruptions, and self-corrects when a step is blocked. It's easy to underestimate because it's built in and automatic, but it handles complex multi-step work well.

The practical implication: you don't need to build an elaborate external task-tracking setup to get Claude to stay on track. The internal to-do list does that job for most operator workflows.

Where Escape becomes essential. With the to-do list running, you can hit Escape at any point to pause Claude mid-task. You'll see exactly where it stopped. You can redirect it, add context, or correct its approach, and it picks up from your instruction without losing the rest of the conversation. This is your real-time course-correction mechanism while the plan executes.

Other ways to feed Claude a plan

The internal to-do system works. But depending on how you and your team like to work, there are other ways to give Claude its marching orders.

GitHub issues. If your team tracks work in GitHub, you can copy an issue directly into the Claude prompt. Paste the full ticket text, turn on plan mode, and have Claude confirm how it interprets the requirements before touching anything. This works well for technical operators running client-specific code repos.

Project management tools. If you track work in Linear or a similar tool, you can copy and paste entire tickets into Claude. The same principle applies: plan mode first, confirm interpretation, then execute. Claude treats a well-written Linear ticket the same way it treats any other structured input.

Custom markdown files. You can create a simple to-dos folder in your project directory and write individual task files by hand. Tag the relevant file in your Claude prompt and Claude reads it as context before planning. Some operators prefer this because it gives them direct control over how tasks are worded and sequenced. The tradeoff is manual maintenance.

MCP servers. MCP (Model Context Protocol) is the integration layer that lets Claude connect directly to external tools. With a Linear MCP server running, for example, you can tell Claude to pull a specific ticket by ID and it reads the full context automatically, no copy-pasting required. We cover MCP in detail in the integrations section.

> Tip. Don't build a complex planning system before you need one. Claude's built-in to-do list handles the vast majority of operator workflows. Add external tooling when you have a genuine coordination need, not because it seems more organized.

A working planning workflow

Here is how a planning session looks in practice for a fractional leader managing multiple active client engagements.

Start a new session in plan mode.

$ claude
[Shift+Tab × 2 → Plan mode active]

Reference the relevant context. If you're working from a ticket or brief, paste it in or reference the file directly.

You: @client-briefs/acme-q2-brief.md
     Plan out the competitive analysis section
     for the Q2 board deck. Three slides max.

Review and push back. Claude will outline slide structure, data sources it intends to use, and framing. If anything doesn't match the client's expectations or the format they've asked for, correct it before execution starts.

Switch to regular mode and execute.

[Shift+Tab to cycle back → Regular mode]

You: Looks good. Build it out.

Use Escape to course-correct in real time. Watch the to-do list tick down. The moment something looks wrong, stop it and redirect.

The whole sequence takes a few minutes of setup. What it gives you is a client-ready deliverable on the first pass instead of a third.

> The real variable is ambiguity. The more clearly you define the implementation details up front, the fewer corrections you make after. Plan mode is just the mechanism. The discipline is doing the thinking before you let Claude run.

---

Author: FractionalSkill