See a Sample Token Audit Report

Bottom line: You can keep the same research quality with fewer tokens per finished brief by moving static context out of the chat body and tightening the output contract. Expect the largest win from ending the re-paste habit—not from switching models first.

Savings percentages are directional from this sample’s structure. Plan limits and pricing change; confirm them against Anthropic’s current documentation.

2. Token map: what ate the window

2.1 Anatomy of a typical turn in this sample

A representative mid-thread turn looked like this:

1. User: roughly 800–1,200 words restating company voice, ideal customer and “how we write briefs”—already in Project instructions.
2. User: a prior outline plus three source dumps “for freshness.”
3. Assistant: a long preamble and a dual-format deliverable: prose, table and checklist.
4. User: “Make it longer, add alternatives, rewrite in founder voice.”
5. Assistant: a near-full regeneration instead of a diff.

That pattern repeated with small variations about six times in 22 turns.

2.3 High-cost moments

• Turns 4–5: the same 12-page notes file was uploaded again after a short digression—functionally a context-reset tax.
• Turns 9–11: three full brief regenerations because acceptance criteria lived in the user’s head, not in the prompt.
• Turn 16: “Use our voice” plus the entire voice document, although Project instructions already contained it.
• Turns 18–20: tool and browsing asides that did not change the final brief structure.

3. Top waste drivers, ranked

Driver A — Re-pasting the operating system of the chat

Evidence: Voice and ideal-customer blocks appear in turns 2, 6, 10 and 16 as near-duplicates. Cost pattern: a large fixed prefix on many turns. Fix: keep one canonical block in Project custom instructions. In the chat, write: Follow Project brief rules v3; only exceptions: …

Do not: paste the full voice guide “just in case.”

Driver B — Document re-injection instead of pointers

Evidence: The same research notes appear twice, the second time after a side topic. Fix: on the first turn, produce a source card with a title, five bullets and quotes with line references. Later, write: Use source card #2 only; ignore prior paste.

Do not: re-paste the corpus to “refresh” Claude.

Driver C — Regeneration without a diff contract

Evidence: “Make it better, longer, more founder” produced a full rewrite. Fix: use a narrow revision pattern:

Revise BRIEF_v2 only.
Keep sections 1–3 unchanged.
Rewrite section 4 to include X and Y.
Max +20% length. No preamble. Output the full brief.

Driver D — No output contract

Evidence: Replies open with method talk and deliver both essays and wide tables. Fix: put a hard contract at the top of each task prompt:

Output:
- Title
- 8 bullets max (one line each)
- Risks (3)
- Sources cited by id
No preamble. No alternate versions unless I ask.

Driver E — Soft goals that invite exploration

Evidence: “Be more strategic” has no definition. Fix: replace adjectives with checks: audience, decision, deadline, forbidden topics and a success test.

4. What to change

Immediate: same day

1. Move static blocks into Project instructions and remove duplicate prompt snippets.
2. Add an output-contract snippet at the top of every new task—not the whole voice guide.
3. Replace full-regeneration language in your snippet library with the diff template.
4. Create source cards for the two PDFs you reuse and stop re-pasting raw text mid-thread.

This week

5. Split work into two thread types: intake and source compression; then brief drafting that sees only source cards. Do not mix them.
6. Cap assistant verbosity. If a reply has a preamble, reply once with strip preamble; same content and update the contract so it stops.
7. Track turns per finished brief and whether you hit the plan limit before week’s end.

Model and mode notes

• Prefer a strong default for drafting and a cheaper, faster pass for source compression—if your plan exposes multiple models.
• Do not switch models to fix re-paste waste; fix the prompt architecture first.
• Confirm current Project instruction length limits and attachment behavior in Anthropic’s interface and documentation.

5. “Do this Monday” plan

6. Out of scope and assumptions

• No API logs or exact token counts were provided. Rankings come from structure and repetition, not a billing export.
• No claim is made about an exact plan tier limit without Anthropic confirmation.
• Every client detail here is fictional. A real report quotes your turns and names your patterns.