When teams distribute reusable developer skills, those skills can cause repeated model round trips by reconstructing fixed procedures as prose. Measure the cost with the vendor telemetry already available, add a small privacy-preserving hook for command context, and restructure work so the model isn’t re‑deciding what the system has already decided.
Use the vendor’s OpenTelemetry export to attribute inference round trips to specific skills via the skill.name field, then measure cost in API requests and cost_usd estimates.
Enable OTEL_LOG_TOOL_DETAILS=1 and add a small local hook that emits a privacy-preserving shape of commands joined by tool_use_id to correlate tool results with telemetry.
# Why this matters Teams that package developer skills (for example as a Claude Code skill, a Cursor plugin, or a shared rules file) create a catalog of how work should be done: service structure, commit gates, which internal library to use. Those artifacts are useful, but if a skill spells out a fixed procedure in prose, the agent will reconstruct and deliberate over that procedure every invocation. That procedural amplification shows up as extra model round trips and higher inference costs even when the context size looks small.
# What you can measure today Claude Code's OpenTelemetry export already includes event-level data that matters: api_request events carry skill.name, cost_usd (an estimate), input_tokens, output_tokens, duration_ms, and an event.sequence. Tool_result events include tool_use_id, tool_name, success, and duration_ms. Crucially, api_request events attribute inference round trips to the active skill via skill.name, so you can count round trips per skill and convert to cost estimates without building a separate pipeline.
# Capture the one missing piece: command shape
# Hook performance and placement
On macOS arm64 the in‑process approach is orders of magnitude cheaper. If a hook must be slower, prefer async execution so the agent doesn't wait for it.
# How to decide whether a skill should be procedural or declarative
# Actionable checklist
# Bottom line Measure inference cost per skill using the telemetry you already have, capture only the minimal safe command context locally, and stop paying for the model to re‑decide steps the system can decide deterministically.
When an agent’s inference bill spikes, the first instinct is almost always the same: swap to a cheaper model. Continue reading on Level Up Coding »
The recurring cost of production AI is not inference. It is re-qualification: the eval reruns, prompt retuning, and regression testing you owe every time a model changes under you. Here is what that tax actually covers, and how to budget for it before it surprises you. The post We Pinned Our Model Version to Stay Safe.
When something doesn’t make sense financially, business owners naturally turn to the accounting records for answers.Sometimes the accounting records provide the answer. Sometimes they simply tell us where the problem became visible.
A disciplined, model-driven framework for hotel brand selection, covering full fee stack analysis, validated RevPAR lift, labor market fit, PIP costs, and franchise agreement negotiation.
I went looking for a clean way to show what prompt caching actually saves an agent, and the first thing I found was a fact that's easy to miss if you only read the "up to 90% savings" headline. Caching a fresh conversation's first turn costs more than not caching it. There's no cache to read from yet, so you pay the in
Originally appeared on The Rails Tech Debt Blog . You approved an APM tool at a modest monthly rate. A year later, the renewal invoice bears little resemblance to what you signed, and nobody remembers deciding to spend that much more. APM is worth having when it’s used well: it resolves incidents faster by showing you
Open the app view to save this story, compare related coverage, and continue from the same source.