Worked Example: CacheScope (spec)

A fully worked example of the one-page spec, so you can see what "filled in properly" looks like. Track: Context Engineering.

Use this as a sanity check. If your spec is vaguer than this, tighten it. If it's broader than this, cut.

---

Goal

Show backend engineers exactly which line of their prompt is busting the Anthropic cache, and what it's costing them every month.

User

Backend engineers who maintain a production Anthropic integration, have already hit $1k+/month in API bills, and suspect something is wrong but can't point at it.

Demo moment

• Judge does: pastes a real production prompt (two consecutive turns) into our CLI
• Our system does: diffs the two calls, highlights the one line that broke the cache, multiplies by observed request volume from the logs
• Judge sees: a red-highlighted line and the text "This line cost you $340 last month"

What's in (max 3)

1. CLI that accepts two consecutive API requests and diffs their prompts at the block level
2. Cost estimator that multiplies the cache-bust by the user's observed request volume
3. One-click fix suggestion (usually: "move this volatile field to the end of the system block")

What's out

• ❌ A web dashboard (tempting but eats the whole weekend)
• ❌ Auto-patching the user's own code (too risky in 24h, and judges will ask about it anyway)
• ❌ Support for OpenAI / Gemini prompts (scope creep — pick one provider and nail it)

Success in 24h

• Working demo that runs end-to-end on stage
• No blocker in the golden path
• Pitch dry-run at least once with a timer

Red flags

• ⚠️ Prompt caching behaviour differs between Claude Opus and Sonnet — pick Sonnet 4.6, document the assumption, move on
• ⚠️ Cost numbers must feel real, not hand-waved — pull from the Anthropic pricing page and cite it during the demo

---

Spec locked. Moving to Phase 3.