Worklog

Practical notes from hands-on experience with AI tools and workflows. Written by our human editors — some by hand, some with AI assistance — never generated by our pipeline.

Schematic diagram: two machines and a shared store holding the plan and the files on disk, with a check step sitting on the path between them, and the four run states pending, running, done and paused below
JULA Worklog 12 min

The Resume Flag That Lied to Me for Three Months

A resume flag is not resumability. Why our pipeline stopped trusting its own progress file and derives finished work …

Line-art schematic: one JSON schema document feeding three AI backends — SDK, CLI, and codex; the SDK path carries an enforced padlock, the CLI and codex paths converge into a parse funnel ending in valid JSON
JULA Worklog 11 min

Asking for JSON Is Not Enforcing It: One Schema, Three Backends, Three Guarantees

Only one of three AI backends enforces a JSON schema at the API layer. Why the schema in the prompt is the portable …

Line-art diagram of a flaky AI agent — a deterministic generator hands the agent a brief holding both an absolute path and a relative path, and three identical runs end in pass, pass and fail
JULA Worklog 12 min

117 Identical Failures, Zero Bugs: Anatomy of a Flaky AI Agent

117 identical errors in seven weeks of overnight agent runs, and no bug in the code — how an ambiguous path contract …

Line-art staircase of a rising quality floor: four treads labelled research, claim check, code checks and run compare, two arrows labelled human read and audit pushing it upward, and a separate circle labelled voice floating out of reach above
JULA Worklog 11 min

The LLM Evaluation Metrics We Actually Run: What 90+ Checks per Article Taught Us

An evaluation harness that grew by accident — what 90+ checks per article catch, where they stop discriminating, and the …

Line-art flow diagram of prompt caching — a request passes a cache check, hits read from cache while misses are processed by the LLM and written to cache, with the $8.12 vs ~$24 cost comparison
JULA Worklog 7 min

Prompt Caching in LLMs, Measured on Our Own Bill

One pipeline run, 7.3 million tokens, an $8.12 bill — what prompt caching actually does to LLM API costs, measured in …

Jula presenting the concept of Claude Skills — persistent agent instructions activated on demand
JULA Worklog 7 min

Understanding Claude Skills: A New Paradigm for Agentic Workflow Automation

How Claude Skills eliminate the repetition tax in AI-assisted development by codifying expertise into persistent, …