Context Engineering · ~5 min

The Five Moves

Every completed turn is a fork in the road. Picking the right move is the core skill of context management.

Why this, for you: the operating manual that ties Lessons 01–03 into a per-turn habit — pure priority-#1 daily-coding payoff. Lessons 01–03 told you why context degrades and where; this is what to do about it, turn by turn.

At each turn boundary you have exactly five options. Most people only ever use one — continue — and let the window rot. The skill is knowing when to reach for the other four.

① Continue (do nothing)

The correct default. Same task, context still relevant, window not heavy. Don't compact or clear preemptively — well under ~100K on a big window carries negligible rot.

② Rewind /rewind · Esc Esc

Drop everything after a good checkpoint. Use on a failed approach — the dead branch (tool calls, errors) is removed entirely instead of being argued with.

③ Clear /clear

Reset context entirely. Use when switching to an unrelated task — prior context would be noise, not signal (the kitchen-sink session).

④ Compact /compact [focus]

Replace history with a summary. Use at task-type transitions or after bulk reads. Lossy — direct it: /compact Focus on X; drop Y. (Lesson 01's move.)

⑤ Delegate subagent

Spawn a subagent for work that spews disposable output. Its window absorbs the reads; only the conclusion returns to yours.

The decision, in four questions

Same task, context still useful? → heavy? Compact. Light? Continue.

Unrelated task next? → Clear (or a new session if nothing is shared).

Wrong path / failed approach? → corrected ≤2 times? Rewind. More than twice? Clear and reprompt sharper.

Just need the conclusion of some exploration? → Delegate to a subagent.

Two non-obvious edges worth burning in. First: rewind beats correction. Stacking "no, not like that" onto a failed branch keeps the whole failed reasoning chain in context, anchoring the model to the same dead end — /rewind deletes the branch. Second: the "corrected more than twice → clear" rule is a hard stop, not a vibe. Past two corrections, the context is polluted; a fresh, sharper prompt beats a third correction.

↪ Your win: name the move at every turn boundary

Default to Continue — but consciously, not by inertia.
Hit a wall? Rewind before you correct a third time.
Switching topics? Clear. Reasoning-heavy stretch ahead? Compact first (L01).
About to unleash a 30-file exploration? Delegate it — keep the residue out of your window.

Retrieval practice — recall, don't peek

Question 1An approach failed and you've corrected it twice. Best move?

Question 2Delegate to a subagent when…

Question 3Switching to a wholly unrelated task, reach for…

Question 4 · spaced recall from Lessons 02–03The tail (recency) slot of an instruction file is best used for…

Ask me anything. Want a printable wallet-card of the four-question decision tree? Curious how delegate interacts with the lethal-trifecta posture when a subagent fetches web content? Say the word — or we can pause lessons and apply the content/ harness fix from Lesson 03.