The seven cardinal sins of token waste

Every wasted dollar of Anthropic spend lands on one of these seven sins. The Mission Control Auditor grades any transcript, CLAUDE.md, or Project instructions against them and reports a letter grade.

1. Opus overuse

Reaching for Opus on a prompt that Sonnet (or even Haiku) would nail. Opus costs ~5× Sonnet on input and ~5× on output. Reserve it for prompts where you've measured Sonnet underperforming.

2. No caching

Sending a stable 4k-token system prompt on every request without cache_control. Each call pays the full input rate for tokens that could be reading at 10% of that rate.

3. Prefix poisoning

The cache only works on the exact prefix. A timestamp, a user ID, or yesterday's date sitting before your stable instructions kills the cache for everything after it. See: prompt-caching.

4. Hedge / wordy / filler bloat

"I would just like to perhaps kindly ask if you could possibly help me…" Every hedge, vague qualifier, and polite filler is a token Claude pays to read. The Prompt Critic module flags them inline.

5. System prompt over 3k tokens

A bloated system prompt makes every call expensive forever. Most production prompts compress to under 2k tokens without losing instruction fidelity. Audit yours.

6. Repeated context

Sending the same 8k-token document body on every call when caching is off. Identical to sin #2 but in the wild, this is usually a Claude.ai Project Knowledge problem: people forget they can attach a doc once and reference it forever.

7. max_tokens unset

A missing ceiling means a degenerate completion bills at full output rate. Output is 5× input — a single 4k-token runaway costs more than most full calls.

---

How the grade works

Each sin contributes 3 / 2 / 1 points by severity. Total:

Score	Grade
0	A — clean
1–2	B
3–5	C
6–8	D
9+	F

Run the Auditor on a recent transcript or your CLAUDE.md to see your grade. The same dollar-savings math drives Mission Control's Recommendations Feed.