Single-request value
Overview
When an API prices or meters per request — a cap of so many requests per minute, a free daily allowance, or a pay-per-call tariff — the unit of economy is no longer the token or the dollar but the request itself. A wasteful question is not one word too long; it is a round trip too many. Ten sub-questions asked in ten separate calls burn ten requests; under a cap they multiply cost and can stall mid-sequence — work done, answers withheld.
This skill is the craft of treating every request as precious. Its centre is the packing decision — what belongs in one request —, and outward to the habits around it: preparing before calling, structuring one response into many deliverables, deciding when a clarification earns its round trip, stopping retries, and keeping the ledger that turns limit arithmetic into a measurable discipline.
Primary sources: https://docs.anthropic.com/en/api/rate-limits and https://ai.google.dev/gemini-api/docs/rate-limits.
When to use
- When the provider limits requests per minute or per day, or prices per request.
- When a free tier's daily allowance is the binding constraint on what can be done.
- When a work session would otherwise stall mid-sequence after requests run out.
- When the same content needs several small outputs and one call could carry them all.
When NOT to use
- For prompt wording, instruction form, and model knob choices (use ac-prompt-engineer).
- For HTTP mechanics, auth, streaming, and the tool-use loop (use ac-llm-api-integration).
- For infrastructure cost attribution and unit economics of systems (use ac-cost-aware-engineering).
Subskill Index
| Leaf | Domain segment | Scope (one line) |
|---|---|---|
subskills/request-economy.md |
prompt-engineering/request-economy |
Reading the limit: requests as the binding currency, and the arithmetic of allowance. |
subskills/batching-sub-questions.md |
prompt-engineering/batching-sub-questions |
Asking in one packed request what would otherwise be ten calls. |
subskills/packing-and-context.md |
prompt-engineering/packing-and-context |
Loading one request with the evidence, the questions, and the output contract. |
subskills/single-request-multi-output.md |
prompt-engineering/single-request-multi-output |
One response structured into several deliverables by design, not parsing luck. |
subskills/clarification-economics.md |
prompt-engineering/clarification-economics |
When a clarifying question is worth a round trip, and when a staged-response contract is better. |
subskills/preparation-and-caching.md |
prompt-engineering/preparation-and-caching |
Reading local and computing cheaply before spending a request. |
subskills/no-retry-degradation.md |
prompt-engineering/no-retry-degradation |
Answering from the response already received rather than burning a retry. |
subskills/value-record.md |
prompt-engineering/value-record |
The per-request ledger: what each request was for, what it returned, what remains. |
Status and honest limits
Packing work into fewer, denser requests trades structure for economy; a packed request demands more careful drafting and its failures are larger, since a malformed multi-part request wastes the whole call rather than one small one. Nothing here recommends a provider or tariff; limit regimes change, and the record should state the allowance and date from the vendor's current published page.