Claude Skills by angusbuilds

Use when moving evals, backfills, nightly enrichment jobs, or embeddings runs off the synchronous API and onto a provider's batch/async tier to cut token costs and API bills by 50%, deciding whether a job is latency-tolerant enough to qualify, resubmitting a partial or expired batch, picking cache TTL or flex/priority service tiers for batch work, or checking that a "batch discount," "cost," or "resubmission" claim didn't quietly change which tokens the model reads.

Use when placing or debugging cache_control / cachePoint breakpoints for Anthropic, Bedrock, Gemini, OpenAI, or Azure prompt caching — cache_read_input_tokens staying at zero, a token bill or cost that jumped for no obvious reason, a tool-heavy turn silently missing cache, deciding where the 4 breakpoints go, choosing 5m vs 1h TTL, or hitting a Bedrock/Gemini cross-provider caching trap. Covers only placement changes that leave the token sequence the model reads untouched.

Use when configuring or auditing a response cache on an LLM gateway (Cloudflare AI Gateway, Portkey, Kong AI Gateway, Higress, Bifrost, OpenRouter, Fastly AI Accelerator, or a hand-rolled Vercel AI SDK caching middleware), when a "cache hit" might be returning someone else's answer, when choosing or reviewing a cache-key shape (full-body hash vs custom label vs partial-message extraction vs embedding similarity), when deciding whether to trust a `cf-aig-cache-status`, `X-Cache-Status`, or sim...

Use when a token cost or spend dashboard (Langfuse, Helicone, Datadog LLM Observability, OpenLLMetry, tokencost, litellm, opencode, Claude Code /usage) disagrees with what the provider actually billed, when investigating a spend spike or "ghost charge", when checking whether a requested cache TTL, Batch API discount, or BYOK key was actually honored, when auditing streaming-disconnect billing or cache read/write token accounting, or when setting up spend caps and anomaly detection without tou...

Use when picking or auditing an Anthropic prompt-cache TTL (5m vs 1h) for a long-running Claude agent loop, when cache_read_input_tokens or the cache hit rate looks lower than expected, when a proxy or gateway might be silently downgrading a requested 1h TTL to 5m, when a Claude Code session's caching cost or billing changes after subscription quota runs out mid-session, or when deciding whether ENABLE_PROMPT_CACHING_1H=1 is worth setting.

Use when comparing prompt-cache economics across two or more LLM providers before picking a vendor split or gateway route, when auditing a multi-provider cost dashboard for the right cache telemetry field per vendor, when someone hands you a provider's "caching doesn't change the output" claim and asks whether that's proven, when a Gemini bill looks 6-14x higher than expected after copying Anthropic-style cache_control markers over, when a Bedrock request keeps paying full input price despite...

Use when building a retry wrapper, idempotency layer, or queue in front of Anthropic/OpenAI/Bedrock/Vertex calls, when evaluating whether a job queue's "unique jobs" or "deduplication" feature actually prevents double-billing, when a caller reports getting back someone else's answer or a stale result instead of their own, when wiring AWS SQS FIFO, Step Functions, Bedrock batch jobs, Google Cloud Tasks, Sidekiq Unique Jobs, BullMQ, Temporal, Inngest, or a Kubernetes CronJob in front of a model...

Use when a token/LLM spend graph has a shape you don't like and you need to know within the hour, when deciding what budget alerts or hard spend caps to turn on for an org or workspace, when investigating a spike like the $1,771-in-4-hours entry against a $25 limit, when suspecting an SDK's auto-retry-on-timeout is double-billing a slow call, when auditing dev/CI traffic or a forgotten nightly eval still running on a paid key, when a cron or scheduled agent's context looks bloated with unused...