The complete walkthroughs from that post

You are helping me with a task I do repeatedly. Follow these rules exactly. ROLE: <one line — e.g. "You are a support lead triaging inbound customer email."> TASK: <one line — the single thing to do with the input below.> RULES: - Use only information present in the INPUT. If something is missing, write "not stated" rather than guessing. - Do not add preamble, apologies, or an offer to help further. - Keep the total response under <N> words. OUTPUT FORMAT (return exactly this shape, nothing else): Summary: <one sentence> Action needed: <yes/no> — <what, if yes> Owner: <name or "not stated"> Draft reply: <3 sentences max> INPUT: <<INPUT>>

# pip install -U openai from openai import OpenAI client = OpenAI() # reads OPENAI_API_KEY from the environment # One place to edit. Spell these EXACTLY as you want them in the transcript. VOCAB = [ "Acme Analytics", # your product names "Northwind Trading", # your client names "Priya Raghunathan", # people whose names get mangled "ARR", "MSB", "webhook", "Postgres", # jargon and acronyms ] CONTEXT = ( "Internal product and sales meeting. Speakers use the company, client, " "person and technical names given in keywords. Transcribe those terms " "with exactly that spelling and capitalisation." ) with open("meeting.m4a", "rb") as f: result = client.audio.transcriptions.create( model="gpt-transcribe", # was: "whisper-1" file=f, prompt=CONTEXT, # unstructured context about the recording keywords=VOCAB, # literal expected terms — the real lever languages=["en"], # expected language codes ) print(result.text)

# CLAUDE.md ## Stack - TypeScript 5.x, Next.js 15 (App Router), React 19 - Postgres via Drizzle ORM - Package manager: pnpm (NOT npm — lockfile is pnpm-lock.yaml) ## Commands - Install: `pnpm install` - Dev server: `pnpm dev` (port 3000) - Test all: `pnpm test` - Test one file: `pnpm test path/to/file.test.ts` - Typecheck: `pnpm typecheck` - Lint + fix: `pnpm lint --fix` ## Layout - `app/` routes and server components - `lib/` shared logic — put new business logic here, not in components - `db/schema.ts` single source of truth for tables; migrations via `pnpm db:generate` ## Rules - Never commit directly to `main` — create a branch first. - Never install a new dependency without asking me first. - Never use `any` or `@ts-expect-error`; the build runs `tsc --strict`. - Never edit files in `generated/` — they are regenerated by `pnpm db:generate`. - Always run `pnpm typecheck && pnpm test` before saying a task is done. ## Conventions - Named exports only; no default exports. - Tests live next to source as `*.test.ts`.

import anthropic client = anthropic.Anthropic() # reads ANTHROPIC_API_KEY SYSTEM_RULES = open("rules.md").read() # your ~5k-token static rulebook def ask(question: str): return client.messages.create( model="claude-opus-5", max_tokens=1024, # tools render BEFORE system; keep the list stable and sorted by name system=[ { "type": "text", "text": SYSTEM_RULES, # frozen: no timestamps, no user IDs "cache_control": {"type": "ephemeral"}, # add "ttl": "1h" for bursty traffic } ], # everything volatile lives AFTER the breakpoint messages=[{"role": "user", "content": question}], ) first = ask("Summarize rule 3.") second = ask("Now summarize rule 7.") for label, r in (("first", first), ("second", second)): u = r.usage print( f"{label}: write={u.cache_creation_input_tokens} " f"read={u.cache_read_input_tokens} uncached={u.input_tokens}" ) # Expect: first -> write>0, read=0 # second -> write=0, read>0 (cache hit confirmed)