Code Is the Most Expensive Way to Think

I used to open Claude Code first. Idea half-formed, terminal open, typing instructions before I’d finished the thought. That habit is expensive in a specific, boring way: the most capable engine is also the most expensive one, and a half-formed idea makes it iterate against itself for a dozen rounds before it lands anywhere close to what I meant.

The Insight

Anthropic confirms the outer edges of this itself: chat is the lighter surface, while Claude Code and Cowork both run heavier — more system prompt, more file context, more tool calls and multi-turn reasoning per session. What I’ve watched happen over months of running all three is the finer point in between: code costs more than cowork on top of that, and cowork costs more than chat. That’s not a complaint about pricing — it’s a fact about the work each tier is built to do, and it only becomes a problem when the tiers get used in the wrong order.

Opening straight into code treats the most expensive engine like a thinking partner. It isn’t one. It’s an execution engine, and execution against a vague idea is just expensive guessing. Every round of “no, not quite, try this instead” is a round billed at the top tier for work that belonged at the bottom one.

The Fix

The order that actually works starts cheap and gets more expensive only as the idea earns it.

Chat first. Talk the idea all the way through — not until it sounds good, but until I could hand it to someone else and they’d know exactly what I’m building and why, no hedging. That’s the bar. A general idea that survives being explained out loud is worth carrying forward. One that doesn’t isn’t ready for the next tier yet.

Cowork next. I run it against my own Obsidian vault and let it pull real context — what already exists, what I’ve already decided, what contradicts what I think I remember. That’s the step that turns a general idea into a specific one, grounded in a knowledge base instead of my memory of it.

Only then does it become a prompt for code: a maximized, optimized set of instructions, handed to whichever engine is built for speed, because by that point speed is what’s actually needed.

I’ll admit the shortcut still tempts me. Sometimes I’ll stay in Claude Code and just talk it out there instead of switching back to chat, especially to get something small over the finish line without breaking flow. For a quick fix, that’s a fair trade. As a default, it isn’t — the focus-switch I’m avoiding costs less than the credits I burn solving a chat-sized problem at code-sized rates.

The Hard Truth

The expensive habit was never the mistake by itself. It’s the ceiling it puts in front of me, usually mid-build, on the exact session where finishing mattered. Cheap tiers first isn’t discipline for its own sake — it’s the only way to make sure the expensive one is still there when I actually need it.

Sources: Anthropic — Claude Enterprise consumption guide