Every message you send uses part of your plan's limit. A long-form video is a lot of messages. These habits are how one video fits inside my plans without hitting the wall halfway.
The expensive steps are the ones that make files: Codex rendering graphics, Claude rendering previews. The cheap steps are reading and marking. So I put an approval in front of every expensive step:
<aside> ✅
Best practice: catch changes as early as you can. A change at the storyboard costs a note. The same change after the render costs a re-render.
</aside>
A new session doesn't remember the last one. If the state lives in a file, the new session reads one file. If it lived in the chat, the new session re-reads everything to figure out where you were.
When I measured one of my videos, about 80% of Claude's tokens went to re-reading earlier context. Long sessions are where the usage goes.
| Claude Code | Codex | |
|---|---|---|
| Start fresh | /clear |
/new |
| Shrink a long session | /compact |
/compact |
| See how much you've used | /usage |
/status |
| Switch model or effort | /model, /effort |
/model |
<aside> ✅
Best practice: in Claude Code, start fresh or compact. Never let a session just keep growing.
/clear. A cut-list session doesn't need the script session in its memory./compact. It shrinks the conversation to a summary and keeps going. Add a note on what to keep, like /compact keep the cut decisions for ch03.Every message re-reads the whole session so far. That re-reading was about 80% of my Claude usage on one video.
</aside>
<aside> 💡
Tips
/usage before a big step, so you don't start a render round with 5% left.
</aside>