The class logbook
The Latest
Everything published on the site, day by day, newest first: guides, analysis, essays and quick thoughts in one running log. The shelves for browsing are elsewhere; this is where you see what is new.
The log
Entries 1–20 of 53
2 entries
Ask for the Smallest Useful Model Output
If your application needs a category and one reason, a page of explanation creates extra reading and processing work.Quick thoughtopenailatency2 entries
AI Tool Error Handling: Don't Fake an Empty Result
AI tool error handling must distinguish an empty catalog from unavailable data. Use separate result types and a mutation test that catches a swallowed timeout.11 min readAI tool error handlingerror handling
Measure Where the Waiting Time Actually Goes
The model may not be the slowest part of your AI workflow. Retrieval, tool calls and repeated validation attempts can consume most of the waiting time.Quick thoughtopenailatency2 entries
AI Agent Tool Design: One Job Per Tool Boundary
AI agent tool design should expose effects and permission boundaries. Compare an overloaded inventory function with explicit read and draft tools.13 min readAI agent tool designtool boundaries
Use Batch Work Only When the Deadline Fits
An overnight document-classification job and a live customer conversation have different timing needs. Choose the API workflow around the deadline.Quick thoughtopenaibatch-processing2 entries
AI Agent Latency Measurement: Find the Real Delay
AI agent latency measurement means tracing the whole user-visible path, not just timing the model call that feels easiest to optimize.11 min readAI agent latency measurementobservability
Match Batch Results by Identifier, Not Position
A batch job can return results in an order different from your input list. Joining by row position risks assigning one answer to another document.Quick thoughtopenaibatch-processing2 entries
Agent Evaluation Variance: Did A Really Beat B?
Agent evaluation variance can make one prompt look better by chance. Design paired repeated trials, report critical failures and inspect discordant outcomes.11 min readagent evaluation varianceAI evaluation
Measure Cache Hits Instead of Assuming Them
Two requests can look similar to a person while failing to reuse a cached prefix. Measure the actual usage data.Quick thoughtopenaiprompt-caching2 entries
Bounded LLM Repair Loop: Three Independent Limits
A bounded LLM repair loop needs separate attempt, deadline and work limits. Test late responses, exhausted repairs and forbidden writes with a fake clock.13 min readbounded LLM repair loopagent reliability
Put Stable Instructions Before Changing Inputs
If many API requests share the same instructions, keep that stable material together. Changing text at the start of every request can prevent useful…Quick thoughtopenaiprompt-caching2 entries
AI Cache Authorization: When Answers Outlive Access
AI cache authorization needs current permissions, revision-aware keys and a defined release point. Test revoked access before and during an answer request.12 min readAI cache authorizationaccess control
Give a Tool Failure a Shape the Agent Can Use
A tool response that says only “something went wrong” leaves the next action to guesswork. Return a contained error that distinguishes useful recovery cases.Quick thoughtopenaitool-calling2 entries
Agent Prompt Injection Boundary: A Worked Case
A retrieved note claims approval for a purchase. See why the agent prompt injection boundary must still deny the write.13 min readagent-securityprompt-injection
Treat Retrieved Text as Evidence, Not Permission
A retrieved document can contain instructions that have nothing to do with the user's task. Its presence in a tool response does not give it authority.Quick thoughtopenaitool-calling2 entries
LLM Validation Error Feedback That Won't Grant Policy Control
Design LLM validation error feedback that identifies malformed fields, limits repair attempts and keeps budget decisions under application control.12 min readLLM validation error feedbackstructured outputs
Separate a Proposed Action From an Executed One
“Prepare a refund” and “issue a refund” are different operations. Your tool interface should make that distinction visible.Quick thoughtopenaitool-calling2 entries
AI Agent Trusted Configuration: Who Owns the Budget?
AI agent trusted configuration keeps budgets and prices outside generated proposals. Test top-level and nested overrides against a closed Python schema.12 min readAI agent trusted configurationPydantic validation
Validate Tool Arguments Before Doing the Work
A model's tool call is a request for your application to act. Validate that request before passing it to a database or service.Quick thoughtopenaitool-callingContinued on the next page.
To browse rather than follow: topics, all articles and all thoughts. To learn in order: courses and the book.
