Clear Strategy Saves API Tokens | Generated by AI

Home 2026.05 English 中文

---

Question: Is using /clear frequently (starting new sessions) one of the strategies for saving tokens when using Claude/Hermes agents?

Answer:

Yes, exactly. /clear is one of the most direct token-saving levers in any agentic CLI workflow. Here’s the mental model:

Why context length = token cost

Every message in a session appends to the context window sent on each subsequent API call. So a 50-turn conversation doesn’t cost 50 × (last message tokens) — it costs 50 × (cumulative history tokens). This grows quadratically in practice.

Turn 1:  100 tokens sent
Turn 5:  500 tokens sent
Turn 20: 2000 tokens sent  ← even if each message is tiny

/clear as a scalpel

When you /clear:

• Context resets to zero (or just system prompt)
• Next call costs only the new message + system prompt
• You lose history but gain a clean slate

The key insight is that most tasks don’t need full history — they only need the last 1-3 turns of context. So /clear between distinct subtasks is almost always the right call.

When to /clear aggressively

Scenario	Clear?
Switching from debugging to writing	✅ Yes
New file / new feature	✅ Yes
After a long multi-tool exploration	✅ Yes
Mid-debugging a single bug	❌ No — context is load-bearing
Continuing a multi-step plan	❌ No

Other token-saving strategies alongside /clear

Since you’re building CLI agents (ww, zz, etc.), you probably already know some of these, but the full picture:

1. /clear frequently — main lever
2. Compact system prompts — don’t stuff 2000 tokens of instructions into every session
3. Tool output truncation — if your agent reads files/runs bash, truncate stdout aggressively before injecting into context
4. Summarize-then-clear — ask Claude to summarize current state → /clear → inject summary as first message. Lossy but cheap
5. Task decomposition — break large agentic tasks into isolated subtasks, each with its own fresh session
6. Avoid re-reading unchanged files — cache file contents client-side, don’t re-inject them every turn

The summarize-then-clear pattern in code

def compress_session(client, history):
    summary = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=500,
        messages=history + [{
            "role": "user",
            "content": "Summarize the current task state in <200 words. Include: what's done, what's next, any key variables/paths/decisions."
        }]
    )
    # Now clear history and inject only the summary
    return [{
        "role": "user", 
        "content": f"[Session resumed] Context: {summary.content[0].text}"
    }]

This is what Claude Code’s “compact” feature does internally — it’s just a summarize + clear under the hood.

Bottom line

Yes, /clear is a first-class strategy. The mental habit to build: treat each /clear as a git commit boundary — finish a coherent unit of work, clear, start fresh. Don’t let context accumulate across unrelated tasks just because it’s convenient.

---

Back Donate