Guides
Guide · 3 min · Anyone

Pick the model

Which one to use, and why the answers get worse the longer you talk.

Read this whenThe answers were good this morning and they are mush now.

Two questions get asked as one. Which model should I use is about the job. Why has it got stupid is about the conversation. Most people answer the first when the problem is the second.

Which model

Three tiers, and the names change faster than the shape does. A small fast one, a middle one, a big one. The middle is the answer to almost everything.

The ends of the range are where the money and the waiting are.

At the time of writing they are Haiku, Sonnet and Opus, and Claude Code moves between them mid conversation with /model. Learn the tiers, not the names: the names get a new number roughly twice a year and the shape has not changed once.

  • Haiku is the small one. Mechanical work where you already know the answer and you want it typed.
  • Sonnet is the middle one and it is the default for a reason. Features, debugging, building a page, ordinary refactors.
  • Opus is the big one. Reach for it when the problem is shaped like a search rather than like a task.

Shaped like a search means a bug living across three files, or an architecture decision you will live with for a year. Renaming a variable does not repay the extra thinking.

Why it gets worse

The model has no memory. Every message sends the whole conversation again, with the project rules and every file it has read. That bundle is the context window, and it has a size.

The fixed cost at the front never moves. The middle grows, and the middle is mostly things that stopped mattering an hour ago.

It is not the model getting tired. It is signal against noise: ninety messages of dead ends and deleted files, weighed the same as what you asked thirty seconds ago.

Compacting, and what it costs you

Compacting replaces the middle of that bundle with a summary of itself. You keep the decisions, you lose the exact wording, and you get the room back. In Claude Code you type /compact. It also happens on its own when the window gets close to full, a safety net rather than a strategy.

The automatic compact fires wherever you happen to be, which is never the boundary you would have picked.

Compact at a boundary you chose, not at the one the machine picked.

  1. 01Compact when you finish a thing, not halfway through one. A summary written mid-thought summarises a mess.
  2. 02Say what to keep. Compacting takes an instruction, so name the decisions and the file paths.
  3. 03Start fresh when the next task is unrelated. A summary of the wrong project costs you all day.
  4. 04Move anything permanent into CLAUDE.md. A rule you re-explain after every compact was never conversation.

A conversation is working memory. A file is long term memory. Durable facts kept in the chat will eventually be summarised away.