Gemini Prompt Cache

Use this skill when the same large instruction block - rules, a response schema, few-shot examples - is sent on every Gemini request. Caching that prefix once turns those repeated input tokens into a much cheaper cached-token rate, which matters a lot at volume. This wraps Gemini context caching: create the cache once, then point every generate call at the cached handle.

When to invoke

• User says: "cache my gemini system prompt" / "cut gemini input cost" / "context caching" / "reuse a long prompt prefix"
• Code in the conversation uses: a fixed multi-thousand-token system instruction repeated across many generate_content calls.

When NOT to invoke

• The system prompt is short, or it changes on every call (caching a one-off prefix loses money on the cache write).
• The user wants output-side savings (use [[gemini-flash-budget]] to drop thinking tokens instead).

Concrete example

User input:

I send the same 4k-token extraction ruleset to Gemini on every row. Stop paying full price for it.

Output:

# Copy assets/cache.py into your project, then:
from cache import create_cache, generate_with_cache

prefix = create_cache(BIG_RULESET, model="gemini-2.5-flash", ttl_seconds=1800)
for row in rows:
    print(generate_with_cache(prefix, row))   # ruleset billed at cached rate

The helper reads GEMINI_API_KEY from the environment and requires google-genai.

Pattern to apply

1. Separate the stable prefix (rules, schema, examples) from the per-call user text.
2. Create the cache once with caches.create, storing the returned handle and a TTL.
3. On each call, pass cached_content=<handle> so only the user text is billed at full input rate.
4. Match the TTL to your batch duration; a too-short TTL forces re-creation, a too-long one wastes storage.
5. Track the savings with [[gemini-cost-tracker]] to confirm the cache is paying off.

Reference: assets/cache.py.

Source

Distilled from production use across the author's automation projects. v1.0.0. See also: [[gemini-flash-budget]], [[gemini-cost-tracker]].

→ Build the full runnable bot with Trawlkit.