Lesson 2.3 — The three moves

Three buttons. The craft is knowing which to reach for, and when.

TL;DR: You manage a finite, rotting desk with exactly three moves — /clear, /compact, and a fresh session + spec. Match the move to the situation.

You know the desk is finite (2.1) and that a full desk rots (2.2). Here's what you do about it.

The three moves

Reset, summarize-and-keep, or start fresh from a plan.

flowchart TD
    Q{"Desk is full — what now?"}
    Q -->|"Switching to an UNRELATED task"| C["/clear<br/>wipe the conversation,<br/>keep the code on disk"]
    Q -->|"SAME task, cluttered,<br/>still need the thread"| K["/compact [focus]<br/>summarize history,<br/>keep going"]
    Q -->|"Big/important task<br/>you've thought through"| S["fresh session + SPEC.md<br/>clean slate, one target<br/>(Lesson 2.4)"]

Compaction isn't all-or-nothing. Clearing stale tool results is "one of the safest, lightest forms of compaction" — drop the giant logs, keep the decisions ¹.

The decision rule

One lookup table for "which move now?"

Situation	Move
Done with this task / switching topics	/clear
Same task, desk cluttered, need the narrative	/compact [focus]
Big task, want a clean run from a plan	fresh session + spec
Stuck after ~2 failed fixes	/clear and restate crisply 2

Two common mistakes:

• Never resetting — treating one endless chat as the unit of work.
• /compact when you meant /clear — if you've moved to a new task, you don't want a summary of the old one on the desk; you want it gone.

🧠 Test Yourself: You're on the same feature but the desk is full of exploration you no longer need — though you still want the narrative of decisions. /clear or /compact?

Answer

/compact [focus]. /clear would wipe the thread you still need; compaction keeps decisions/bugs and drops the raw file dumps ¹.

Worked example

Right vs. wrong on a topic switch, plus mid-task clutter.

You finished a feature (tests green). Next: refactor an unrelated module.

	Action	Result
❌ Wrong	Keep typing in the same session	Old files, logs, discussion dilute the refactor
✅ Right	/clear, then: "Refactor src/payments/retry.ts to exponential backoff. Keep the public signature. Run npm test -- retry to verify."	Clean desk, sharp prompt

Mid-task clutter: deep in one bug, desk full of exploration, but you still need the narrative → /compact focus on the auth bug and the two approaches that failed → tight summary, raw dumps gone.

Per-agent mechanics

The move is universal; the button differs.

	Claude Code	Codex	Cursor
Reset	/clear	/clear	New chat / new tab
Compact	/compact [focus]	/compact (+ auto-compaction)	auto-summarizes long threads
Selective recall	@-mention files · /rewind	session resume	@Past Chats

⚠️ Watch-out: Codex auto-compacts aggressively, and auto-compaction can quietly drop a subtle constraint you stated 30 turns ago. Don't trust auto-compaction with invariants — when you compact (or it compacts for you), restate the non-negotiables in your next message ².

Your turn (exercise)

Name the move for each (answers below — don't peek):

1. Tests just passed; now you'll write docs for the same feature.
2. Debugging one issue for 15 turns; the agent is circling old wrong fixes.
3. About to start a large multi-file feature you've already thought through.
4. Finished the API work; now touching a completely separate CSS file.

Answers

1. `/compact` (same feature, keep the thread) — or `/clear` + a crisp docs prompt if you don't need the history. 2. `/clear` and restate the bug crisply (past the 2-failed-fixes line; history is now scar tissue). 3. Fresh session + a `SPEC.md` (Lesson 2.4). 4. `/clear` (unrelated task).

---

← Lesson 2.2 · next → Lesson 2.4 — The spec handoff

---

1. Effective context engineering for AI agents — Anthropic ↩↩

2. Best practices for Claude Code — Anthropic ↩↩