Bandwise docs
Templates

Templates

Ready-made question sets for common decisions, with example states and borderline cases.

Each template is a complete spec you can copy, edit and run. It comes with two or three example states and one borderline case per question, so you can test the wording on hard cases before you tune thresholds. Every template sends outages to review.

The template pack is part of the free Bandwise kit, licensed Apache-2.0. The find decisions skill uses it to draft specs from the LLM calls in your code.

TemplatePatternJob
Email triagecomposite_scoringDecide whether an email needs the recipient soon, and what kind of message it is.
PR safety gatecomposite_scoringDecide whether a pull request is safe to merge without a human reviewer.
Log-line pagerconfidence_routingDecide whether one log line needs a human now, so only real problems page someone.
Context prunerconfidence_routingDecide, item by item, whether an agent's context item still matters for the current task, and drop the rest unchanged.
Wake gateintent_routingDecide whether an event should wake a sleeping agent now, later, or not at all.
Done checkconfidence_routingDecide whether a coding agent has really finished the request before it stops, or has work left or an unchecked claim.
Action risk gatefan_outDecide whether a coding agent's shell command or file edit is risky enough that a person should confirm it before it runs.
Model tierintent_routingDecide how hard a new request to a coding agent is, so mechanical work can go to a cheaper model or subagent.
Security finding triageconfidence_routingDecide whether a static analysis finding is reachable from user input, and how bad it is in context.
Error triagefan_outDecide which team owns a new error tracker issue, how much it hurts users, and whether the latest release caused it.
Lead and event scoringconfidence_routingScore how ready an account is to buy from its recent product events, and pick the next sales action.
Inbound email routingintent_routingSend each email that reaches a shared address to the right queue, and flag the ones that need a human reply today.

Using a template

  1. Copy the spec into your repository as bandwise/sets/<slug>.json.
  2. Rename the state fields to match your data, and update input.schema.
  3. Rewrite the instructions and options for your case. The question id is never sent to the model, so the instructions carry the whole requirement.
  4. Run the example and borderline states with pnpm bandwise run --local.
  5. Label real examples and tune the thresholds on them. The thresholds in a template are starting points, not measured values.

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.

On this page