smart-zonelisted
Install: claude install-skill Udhay-Adithya/skills
# Smart zone
The **smart zone** is the span of context within which the model still reasons sharply.
Past it, the window technically still accepts tokens and the quality of the thinking in it
quietly falls off: earlier decisions get forgotten, instructions from the top of the
session stop being followed, and the same ground gets re-covered.
Working past the edge of it is the most common reason a long session produces worse work
than a short one, and it is almost always invisible from the inside. The fix is not to
push on. It is to finish the piece in hand and start again with a clean window and a good
prompt.
## Knowing where you are
The token count cannot be read directly, so work from what is observable.
**Hard signals.** A compaction or summarisation has already happened. The harness has
warned about context. The user says so. Any of these means the edge is already behind, not
ahead.
**Proxies, in rough order of weight.**
- Large files read in full, especially several, and especially anything over a few hundred
lines.
- Long tool outputs: a wide `grep`, a full test run, a verbose build log, a big diff.
- Output that got persisted to a file rather than returned inline, which means it was big.
- Many turns, with many rounds of edit-and-check.
- The same file read more than once in a session, which usually means the first read has
fallen out of working attention.
**Behavioural signals, which matter more than any count.** A decision made earlier being
re-litigated.