⚡ Claude Token Optimization2026-04-21 — Smoke Claude

[github-actions]

Target Workflow: smoke-claude.md

**Source (redacted) Pre-downloaded run data from /tmp/gh-aw/token-audit/ (18 runs over 7 days)
Estimated cost per run: $0.071
Total tokens per run: ~185K
Cache read rate: 77% of total tokens
Cache write rate: 22% of total tokens
LLM turns: 6.8 average (range: 5–9)
Model: claude-haiku-4-5

---

Current Configuration

Setting	Value
Tools loaded	github (pull_requests toolset), playwright (all tools), bash (all commands)
MCP tools actually used	playwright: browser_navigate, browser_snapshot, browser_evaluate, browser_close, browser_wait_for
Network groups	defaults, github, playwright
Pre-agent steps	Yes (1: mkdir for playwright logs)
Prompt size	~5,276 chars (incl. injected system block)

Per-turn token breakdown (avg across 18 runs)

Token type	Average	% of total
Cache read	143,634	77%
Cache write	39,402	22%
Input (new)	38	<1%
Output	1,520	<1%
Total	184,594	100%

---

Recommendations

1. Reduce Turn Count by Combining Operations in the Prompt

Estimated savings: ~25–35K tokens/run (~15–20%)

The workflow currently averages 6.8 turns for 4 tasks. The main driver is separate LLM turns for:

• GitHub MCP call (1 turn)
• Playwright navigate + snapshot (1–2 turns)
• Bash write + bash verify (1–2 turns)
• Safe output add_comment + add_label (1–2 turns)

The prompt should explicitly instruct the agent to batch operations per turn. Add to the prompt:

**Efficiency instructions:**
- Combine file creation and verification into a single bash call: `mkdir -p /tmp/gh-aw/agent && echo "..." > /file && cat /file`
- Use a single playwright turn: navigate AND verify the title in one step (do not take a separate snapshot unless navigate result is insufficient)
- Post the comment and label in a single combined safe-output invocation where possible

This alone could reduce turns from 6.8 → 5, saving ~2 turns × ~13K tokens/turn = ~26K tokens/run (~14%).

2. Replace Playwright Page-Title Check with Bash curl

Estimated savings: ~18–30K tokens/run (~10–16%)

The current smoke test uses Playwright to navigate to https://github.com and verify the page title contains "GitHub". This is trivially always true and consumes 1–2 LLM turns (navigate + snapshot). A curl HEAD check achieves the same signal at near-zero token cost:

Change in smoke-claude.md:

-2. **Playwright Testing**: Use playwright to navigate to https://github.com and verify the page title contains "GitHub"
+2. **Playwright Testing**: Use bash to run `curl -sI https://github.com | grep -i "content-type"` and verify the response contains `text/html`. This tests HTTPS connectivity without requiring a full browser session.

And remove the playwright tool and network group entirely:

 network:
   allowed:
     - defaults
     - github
-    - playwright
 tools:
   github:
     toolsets: [pull_requests]
-  playwright:
   bash:
     - "*"

⚠️ Tradeoff: This changes the nature of the test from browser-based to HTTP-only. If the intent is specifically to validate the Playwright MCP server, keep Playwright but apply recommendation #1 to reduce its turn count.

If Playwright is kept, at minimum ensure the test uses only 1 turn:

2. **Playwright Testing**: Navigate to https://github.com using `playwright_browser_navigate`. The tool returns the page title — verify it contains "GitHub" from the navigate response directly. Do NOT call browser_snapshot separately.

3. Add Explicit Tool Use Guidance to Reduce Extraneous MCP Calls

Estimated savings: ~5–10K tokens/run (~3–5%)

From the run data, playwright_browser_snapshot is called in only 7/18 runs — indicating inconsistent agent behavior. Explicit instructions prevent unnecessary tool calls:

Add to the prompt body:

**Tool constraints:**
- GitHub: use `list_pull_requests` with `perPage: 2, state: closed` only — do not fetch PR details or reviews
- Playwright: call `browser_navigate` only — do not call `browser_snapshot` unless navigate returns no title
- Bash: chain commands with `&&` to minimize turns

4. Pre-Compute the Playwright Log Directory in the Existing Step

Estimated savings: Negligible tokens, but removes runtime ambiguity

The current steps: already creates the playwright log dir. This is correct. No change needed here — it's already optimized.

---

Cache Analysis (Anthropic-Specific)

Metric	Value
Avg cache write/run	39,402 tokens
Avg cache read/run	143,634 tokens
Cache read : write ratio	~3.6×
Runs analyzed	18

Cache write amortization: The 39K cache write on turn 1 (system prompt + tool schemas) is reused across ~6.8 turns of cache reads per run. At a 3.6× read:write ratio, the cache pays for itself within ~2 turns — caching is working well here.

Cache cost vs benefit: For Haiku pricing, cache writes are ~1.25× the cost of regular input tokens. With 6.8 turns reusing the cache, each write token is read ~3.6× — the savings are real and caching should be kept.

Cache hit consistency: 77% cache read rate is healthy. The run-to-run variation in total tokens (119K–325K) is driven by turn count variation (5–9 turns), not cache misses.

What drives cache write size (~39K)? The system prompt (security policy, safe-outputs instructions, tool schemas) accounts for the bulk. This cannot be reduced in the workflow file itself — it's framework-injected.

---

Expected Impact

Metric	Current	Projected	Savings
Total tokens/run	~185K	~130–140K	~25–30%
Cost/run	$0.071	$0.050–0.053	~25–30%
LLM turns	6.8	5.0	−1.8 turns
Session time	~3.6m	~2.8m (est.)	~20%
Monthly cost (18 runs/week)	~$20.50	~$14–15	~$5–6/mo

---

Implementation Checklist

• Add efficiency instructions to smoke-claude.md prompt body (Rec Improve links in readme to AW project #1): combine bash ops, single playwright turn
• Decide: replace Playwright with curl (saves most) or keep Playwright with tighter instructions (Rec Secret proxying #2)
• Add tool constraint instructions to prompt body (Rec feat: add integration test for rostbuness #3)
• Recompile: gh aw compile .github/workflows/smoke-claude.md
• Run post-processor: npx tsx scripts/ci/postprocess-smoke-workflows.ts
• Open a PR and verify smoke CI passes
• Compare token usage on new runs vs ~185K baseline (check agent_usage.json in logs)

Generated by Daily Claude Token Optimization Advisor · ● 788.5K · ◷