Takeaways
- The pattern: Simon Willison has shipped 14 separate CLI tools in the past quarter alone. Each one is small. The compounding effect is enormous.
- The thesis: The frontier of practical LLM usage is being defined more by tiny, well-designed CLIs than by model releases.
- The signal: When
llm,datasette, and his model-agnostic prompt tooling get adopted by your dev workflow, you stop caring which lab won this week.
What's actually shipping
The recent cluster is striking: a new llm-bench plugin for running local model evals, a datasette-llm integration that lets you query SQLite tables with natural language and audit the generated SQL before execution, and prompt-history — a small daemon that snapshots every LLM call you make from your shell so you can grep your own past prompting.
None of these are billion-dollar product launches. Each one solves a real, specific problem that frontier-lab interfaces don't address. The cumulative effect is that an experienced dev using Simon's toolchain is operating at a meaningfully higher altitude than the same dev using OpenAI's or Anthropic's first-party tooling.
What it says about the field
Two things. First, that the most useful work in AI right now is being done at the seams — between models, between providers, between the LLM and the rest of your stack. Second, that indie tool-builders with editorial taste are an underrated multiplier. Simon's blog acts as both spec and changelog for a quiet stack that doesn't show up in any analyst report but ships real productivity gains to people who pay attention.
What to watch
Whether any of the frontier labs notice. There's a version of this where Anthropic or OpenAI hires Simon to do for them what he's currently doing in public. There's another version where they keep ignoring it and the indie-tools ecosystem keeps growing into a moat against any single lab's tooling dominance.