T | TERSE | ISSUE 007 · JULY 2026 |
|
THE TERSE LETTER · 4 MIN READ You read 5 files
when you needed 1Context bloat is the #1 hidden cost in agentic coding. Put your window on a diet. |
Agentic coding tools love to read files. Helpful — until your context window is stuffed with four files you glanced at and one you actually changed. Every one of them stays in context, billed every turn, until something pushes them out. Here's how to keep the window lean. |
Reads accumulate; they don't expire.Each file read, each search result, each long command output lands in the window and stays. By turn 20 you're carrying tens of thousands of tokens of stuff you no longer need — and re-paying for all of it, every turn. |
Context size by turn 20 · typical session | → On long sessions, pruning saves around $220/year. |
|
|
Read narrow. Summarize. Start fresh.Ask for specific functions, not whole files. Have the agent summarize what it learned, then drop the raw source. And when a task is done, start a clean session instead of dragging the old context forward. |
✗ BLOATS THE WINDOW Read entire files Re-read after edits Keep stale output One endless session |
| ✓ STAYS LEAN Read by function Summarize, then drop Clear old results Fresh task = fresh chat |
|
|
Terse tracks tokens per source so you can see exactly when the window is filling with noise. |
Three things to do right now. |
1 | Ask for ranges, not files "Show me lines 40–80" beats "read this 800-line file" nine times out of ten. |
2 | Summarize and drop Have the agent restate what it learned, then clear the raw reads from context. |
3 | Start fresh between tasks A new task rarely needs the last task's context. Open a clean session. |
|
macOS · Free forever tier · 1,500 optimizations/week |
|
NEXT WEEK Why "the" is 1 token and "ChatGPT" is 2. A quick, useful tour of how tokenization actually works — and the abbreviations that quietly cut your prompts in half. |
|
Until then — stay terse. — James
Builder of Terse · terseai.org |
