News & guides
Best Model for TypeScript: Let the Compiler Do the Grading
TypeScript gives you a free verifier, which changes which model you should pay for. How to build the tsc loop and where models still fall down.
· 9 min read
Best Model for Vision Tasks: What Images Are Actually For
Vision in a developer workflow means screenshots, diagrams and scanned documents. Which models take image input, and where visual understanding still fails.
· 8 min read
Blind Model Comparison: A Method That Survives Bias
Knowing which model wrote an answer changes how you judge it. A practical protocol for blind pairwise comparison that a small team can run in a day.
· 9 min read
Budgeting for Experimentation Without a Surprise Bill
Experiments produce the largest unexpected AI invoices because nobody set a ceiling. How to fund trying things: separate ledgers, time-boxes and kill switches.
· 12 min read
Building a Changelog Generator People Actually Read
Restating commit subjects is not a changelog. How to pick the right input, separate classification from writing, handle reverts, and keep regeneration deterministic.
· 9 min read
Building a Code Search Tool That Beats grep
Semantic code search only pays off if you beat ripgrep on the queries it fails. Chunking at symbol boundaries, hybrid retrieval, incremental indexing and honest evaluation.
· 10 min read
Building a Docs Chatbot That Refuses to Guess
A docs bot fails on the gap between what is documented and what users ask. Version-aware chunking, hybrid retrieval, forced citations, and abstention that actually fires.
· 10 min read
Building a Model Eval Harness You Will Actually Maintain
A practical design for a local evaluation runner: fixture format, isolation, scoring, cost tracking and the CI wiring that keeps it from rotting.
· 10 min read
Building a Postman Collection for OpenAI-Compatible APIs
A reusable Postman or Insomnia collection for any OpenAI-compatible endpoint — environments, streaming, tool calls, saved examples, and sharing it without leaking keys.
· 9 min read
Building a PR Summariser Reviewers Do Not Skip
Most PR summary bots restate the diff and get ignored within a fortnight. What reviewers actually need, how to select the diff, and how to keep cost per PR predictable.
· 9 min readBuilding a Prompt Library People Actually Reuse
Most prompt libraries become a graveyard of untested snippets. What makes one worth maintaining: ownership, discoverability, contracts and a deletion policy.
· 9 min read
Building an Incident Assistant On-Call Engineers Trust
Automate the first ten minutes of context gathering, not the diagnosis. Read-only tools, evidence-linked output, a latency budget, and why a wrong answer at 3am is expensive.
· 10 min readReady to try it?
Every model behind one key. Start free and upgrade when it sticks.