T | TERSE | ISSUE 004 · JUNE 2026 |
|
THE TERSE LETTER · 4 MIN READ Your CLAUDE.md is
3× too longAnd it loads on every single turn. Trim it once, save all month. |
Your CLAUDE.md is the most-read file in your project — Claude loads it on every turn. Which means every redundant sentence in it is a tax you pay hundreds of times a day. Most are bloated with repetition, over-explanation, and rules that contradict each other. Here's how to tighten yours. |
| 01 | The three bloat patterns |
Repetition, narration, and dead rules.Rules restated three different ways. Long explanations of why when the model only needs the what. And stale instructions for files that no longer exist. All of it loads, every turn, forever. |
CLAUDE.md size · loaded every turn | → At 40 turns/day, the trim saves roughly $180/year. |
|
|
| 02 | How to trim without losing rules |
One rule, one line, stated once.Convert paragraphs to imperative bullets. Delete the rationale. Remove anything referencing files you've since deleted. If two rules overlap, merge them. The model follows tight instructions better, not worse. |
✗ BLOATED 3 paragraphs on style "The reason we do this…" Duplicate rules Dead file references |
| ✓ TIGHT One bullet per rule Imperative voice No rationale Current files only |
|
|
Terse can compress reference docs and CLAUDE.md blocks before you commit them. |
Three things to do right now. |
1 | Convert prose to bullets Every rule should be one imperative line. "Always use tabs" — not a paragraph about why tabs. |
2 | Delete the why The model needs the rule, not the history lesson. Cut every "because" clause. |
3 | Re-order for caching Stable rules at top, volatile notes at the bottom — so the cached prefix stays intact (see Issue 001). |
|
macOS · Free forever tier · 1,500 optimizations/week |
|
NEXT WEEK Markdown is eating your tokens. All those #, *, and ``` characters cost real money in long outputs. I'll show when to ask for plain text — and how much it saves. |
|
Until then — stay terse. — James
Builder of Terse · terseai.org |
