Worked Example: CacheScope (pitch)

The 3-sentence pitch and judge Q&A for the CacheScope spec in good-spec.md. Use it as a shape reference.

---

Sentence 1 — The wedge

"When you add a timestamp to your Claude system prompt, you silently burn your entire prompt cache on every call."

17 words. Names a real mistake engineers make this week.

Sentence 2 — The move

"We built CacheScope — a CLI that diffs two API calls and tells you exactly which line cost you money."

19 words. One noun (CacheScope), one verb (diffs), concrete outcome (which line cost you money). No adjectives.

Sentence 3 — The moment

"Watch this."

Then run the live demo. Paste a real prompt. Red line. "$340 last month." Done.

Total demo time: 50 seconds. Do not speak during the demo — let the red line do the work.

---

Judge Q&A prep

The standard five

1. How does this scale?

   We read request logs once and run the diff offline. Every engineer on the team gets a personal report; the analysis is cheap.

2. Why not just use Helicone or LangSmith?

   Those tools show you total spend. We show you the exact line of code that caused the spike, so you can fix it without archaeology.

3. What happens when the LLM hallucinates?

   Irrelevant — we're doing static analysis of your prompts, not running inference. There's no LLM in the hot path.

4. Who pays for this?

   The same budget that pays the API bill. We sell it as a cost-cutter, not a dev tool, so it comes out of ops spend, not engineering spend.

5. What's your moat?

   We plug into CI. Every PR shows a diff of "this change would cost you $X/month in extra cache busts". Once that diff is in the PR template, you're the default.

Hack-specific three

• Does it support streaming responses?

  Yes — we diff the request side, not the response, so streaming doesn't matter.

• What about function calling / tool use?

  Tools are cached together with the system block — we treat tool schemas as part of the cache surface and diff them the same way.

• Why only Anthropic?

  Because Anthropic's prompt caching is the most aggressive and therefore the most punishing to bust. We ship with the one provider where the ROI is most visible; other providers are week 2.

---

Dry-run checklist

• Each of the 3 sentences is under 20 words
• Sentence 3's live demo matches the spec's Demo moment exactly
• All 5 standard Q&A answered in one line
• Pitch run out loud, once, with a timer — came in at 1m 42s, under the 2-minute cap