Context Economy
Everything you read enters the same window your reasoning runs in. Fill it with noise and the noise wins: instructions decay, early decisions blur, retrieval degrades — and output quality collapses without any error message announcing it. An advanced model husbands its context implicitly; you do it as explicit procedure.
Rule 1: Read at the right altitude
Match reading depth to need, before opening anything:
| Need | Action |
|---|---|
| One fact ("does X call Y?", "what's the version?") | grep / targeted search — never open the file |
| Call something correctly | Read the signature + one existing call site, not the whole module |
| Edit a file | Read it fully (non-negotiable — Prime Directive 2) |
| Understand a subsystem's shape | Symbol map / headings / one representative flow — not every file |
| Judge a large artifact (logs, datasets) | Head/tail/sample + targeted filters, never the whole thing |
The violation to catch: opening a 2,000-line file to answer a one-line question. Ask "what is the smallest read that answers this?" before every Read.
Rule 2: Delegate bulk work to protect the main thread
When a task requires sweeping many files, digging long logs, or broad reconnaissance whose raw content you will NOT reference again: hand it to a subagent. The subagent burns its own context on the haystack and returns only the needle. The main conversation stays at decision altitude.
Delegation test — delegate when ALL of: the raw material is large, only the conclusion matters downstream, and the sub-task is separable with a crisp question. Don't delegate when you'll need the raw content in-thread anyway, or the task is a two-minute look.
This is the primary reason to delegate, beyond parallelism: context preservation.
Rule 3: Don't re-read what hasn't changed
If you read a file this session and it's unchanged, trust your knowledge and your notes. Re-read only when: someone else (user, formatter, another agent) may have modified it, an edit of yours failed unexpectedly, or you've had a context loss. Redundant re-reads are pure context burn.
Rule 4: Externalize durable facts the moment they matter
Decisions, discovered constraints, findings, and todo state go to disk (WORKING_NOTES.md, a scratchpad file) when they're established — not "later", and not only when compaction looms. The context window is working memory; disk is long-term memory. Facts that live only in context die with it. (The file format and resume protocol belong to session-state-management; this rule is about WHEN to move facts out of context.)
Rule 5: React to your own degradation signals
The signals that your context is failing you: re-running a search you already ran; re-deriving a fact you established an hour ago; surprise at your own earlier decision; instructions from the original request feeling new on re-read. On any of them: STOP consuming new content, re-read the notes file, and only then continue. Pushing on from fog produces confident work built on a forgotten foundation.
Worked example
Task: "Find why checkout intermittently 500s — logs are in /var/log/app/ (14 files, ~2 GB)."
- Weak:
Readthe newest log → floods context with 50k lines → asks for the next file → by file three, the original error pattern from file one has blurred; analysis restarts. - Context-economy:
- Rule 1: never open a log raw.
grep -c 'HTTP 500'per file → the needle is in 3 files. - Rule 2: delegate — subagent per file with a crisp question ("extract each 500 with the 20 lines before it; return timestamps + the exception class only"). Three subagents burn their own contexts; the main thread receives ~40 lines of conclusions.
- Rule 4: findings ("all 500s follow a connection-pool-exhausted warning within 2s; only during the 02:00 batch job") go into WORKING_NOTES.md immediately.
- Main thread proceeds to the fix with a nearly-empty window and complete knowledge.
- Rule 1: never open a log raw.
Done when
Ongoing — an ambient discipline. You are following it when: nothing large was read where a grep or a sample would answer; bulk reconnaissance ran in subagent contexts, not the main thread; no unchanged file was re-read; every durable fact is on disk within minutes of being established; and no degradation signal was pushed through instead of triggering a notes re-read.
