When Extended Thinking Flattened My Accuracy but Doubled the Bill — Field Notes on Measuring the Marginal Utility of Thinking Tokens
You pinned budget_tokens high 'to be safe,' accuracy barely moved, and the bill kept climbing. These field notes show how to ledger real thinking-token usage by p50/p95/hit-rate and measure how much accuracy each extra 1,000 thinking tokens actually buys.
Tightening Tool Schemas From the Arguments You See in Production
Record the arguments Claude actually passes to your tools in production, then use that distribution to add enums and patterns back into your JSON Schema. With logging code and before/after numbers.
Make Your Nightly MCP Connectors' Health Visible — A Lightweight Ledger for Solo Operators
You don't need Enterprise connector observability to see your MCP connectors' error rate and latency. Append one line per tool call, roll it up weekly, and let regressions ring a bell. A working health ledger for anyone running scheduled tasks solo.
Same Output, Different Path — Guarding Agent Trajectories with Invariants
When the default model changes, your final output can stay correct while the path your agent takes quietly shifts. Here is a trajectory regression harness built on recorded tool traces and deterministic invariants, with working code and measured numbers.
A connector failed for two nights and I never noticed — instrumenting my solo setup after observability went to public beta
The week connector observability hit public beta, I realized my one-person operation had no view into errors or latency. Here is how I wrapped my MCP connector calls in a thin meter and started reviewing it weekly.
My background session sat at running all night — adding heartbeats to the agents view
The Claude Code agents view finally lets you see every background session at once, but a running badge is not proof of progress. Here is the external heartbeat layer I built to catch a silently stalled session in minutes instead of the next morning, with real numbers and the traps I hit.
The Two Weeks My Web Monitor Said Everything Was Fine — Field Notes on Catching Silent Misses
A competitor monitor built on Cowork and Claude in Chrome can keep reporting no changes while quietly missing them. Here is how I separated fetch success from extraction success and instrumented the silent failures, with the code I actually run.
When a Claude Code Refactor Passes Every Test but Behaves Differently in Production — Catching Silent Contract Drift with a Behavior Diff Harness
Hand Claude Code a large refactor and your tests can stay green while production behavior quietly shifts. Here is how I record exception channels, log shape, init order, and return values as a signature, then diff them per commit to catch contract drift before it ships.
When Claude API Prompt Caching Quietly Stops Hitting in Production — Field Notes on TTL and Measured Savings
Prompt caching works beautifully the day you ship it, then quietly stops hitting in production. The five things that break the prefix, how to choose between 5-minute and 1-hour TTL, and how to measure real savings from usage instead of guessing.
Putting Cloudflare AI Gateway in Front of Claude Made the Numbers I Needed Disappear — Field Notes on Instrumentation
After putting Cloudflare AI Gateway in front of Claude API, here is where I actually got stung — cost attribution, semantic-cache false hits, fallback quietly lowering quality, and budgets that don't really stop anything — with the code I used to fix each.
Noticing From the Outside When a Scheduled Job Quietly Did Nothing
exit 0, but zero output. How to catch a silent no-op not from the job's own log but from an external heartbeat ledger and ground truth, written from running several sites on a nightly schedule as an indie developer.
When Your Claude API Response Cache Returns Stale Answers and Near-Miss Wrong Ones — Field Notes on Freshness and False-Hit Suppression
A Claude API response cache improves latency and cost immediately, but the problems that hurt in production are not average hit rate — they are stale hits and semantic false hits. Here is the key design, freshness management, false-hit suppression, and observability that keep a cache honest.