Find decisions

Many LLM calls do not need to write anything. They answer "is this spam", "which team owns this" or "how urgent is it", and the code branches on the answer. Those calls fit a System One model through a question set: you get a decision and a confidence band instead of free text.

The find-decisions skill looks for those calls in your repository and drafts a spec for the best ones.

It stays on your machineCopy Anchor Link

The skill runs inside Claude Code in your repository. It reads files, writes a report and draft specs, and runs bandwise run --local when the CLI is available. It never sends your code, file paths or findings to any service, including Bandwise and TypeSafe. Example states it writes are invented, not copied from your data.

What it doesCopy Anchor Link

1. Finds LLM calls. It searches for the Anthropic and OpenAI SDKs, the Vercel AI SDK, LangChain, LiteLLM, and raw HTTP calls to chat endpoints, plus your own wrappers around them.
2. Keeps the decisions. A call counts when your code compares its output to fixed values, parses it as yes or no, compares a number to a cutoff, uses it to pick from a list, or uses it to keep or drop items.
3. Classifies each one as a Noul (does X hold), a Choice (which one) or a Score (how much, on a fixed scale), picks a pattern, and lists the state the decision needs.
4. Flags poor fits. A call that must write text, a decision that takes a person more than about 10 seconds, or a choice with no fixed option list stays on the LLM. Exact rules, arithmetic and date logic stay in code.
5. Estimates volume and cost only where the code shows it, such as a call per request, per item in a batch, or on a schedule. Otherwise it says "unknown".
6. Ranks the candidates and drafts a spec for the top three, starting from the closest template.
7. Validates each draft with bandwise run --local if the CLI is installed, and fixes lint errors.
8. Writes a short report to bandwise/find-decisions-report.md: the candidates, the poor fits, the drafts and what to verify.

Using itCopy Anchor Link

Copy the find-decisions skill folder into .claude/skills/ in your repository (a Claude Code plugin install comes when the kit is published). Then ask Claude Code:

Find the LLM calls in this repo that are really decisions, and draft Bandwise specs for the best ones.

Drafts land in bandwise/sets/<slug>.json, each with an invented example state next to it.

After the reportCopy Anchor Link

• Check each draft's option list and state fields against your code.
• Label 20 to 50 real examples per draft and tune the thresholds on them. Template thresholds are starting points.
• Branch app code on overallAction first, then on route, and keep your existing path for fallback.

The skill makes no claims about accuracy or savings. It tells you what the code shows and what to measure.

Bandwise is an independent product built on TypeSafe's System One models. It is not TypeSafe's documentation. For the System One models themselves, see docs.typesafe.ai.