Use-case templates

A use-case template is a starting point for putting one task on the loop: a starter output schema, a starter grading guide, a golden checklist, and an escalation starter — bundled as one machine-readable file so a human or a coding agent can turn it into a real route without starting from a blank page. The templates carry no numbers about how well any model does the task; they are a shape to fill in, and what you fill in is what your goldens and grades then measure.

Nine are published today, one per task the shipped teach → calibrate → grade → optimize loop can serve:

• Ticket triage — classify a support ticket into your categories and a priority, and flag the ones a human should see.
• Document extraction — pull the fields you care about out of a document into a structured object.
• CS next action — recommend the next action on an account, grounded in its signals, and hand off the ones that need a person.
• Lead qualification — sort an inbound lead into your qualification stages and a routing segment, and flag the ones a rep should work by hand.
• Compliance review — check an item against your own policy and route the uncertain and adverse ones to a reviewer; a first-pass triage, never a compliance control.
• Catalog cleansing — normalize a catalog record into your taxonomy, correct what is clearly wrong, and flag the ones a person should check.
• Call disposition QA — from a transcript, check the logged disposition and score the call against your scorecard, handing the disputed ones to a supervisor.
• Financial classification — classify a financial record into your categories and risk flags, and hand off the ones a person must review. Regulated data — read the custody terms first.
• Medical classification — classify a clinical document into your categories and hand off the ones a qualified person must review; it sorts documents, it does not make clinical decisions. Regulated data — read the custody terms first.

The last two handle regulated data. Their fit notes quote the published content custody terms verbatim: do not send protected health information or other regulated data through ModelRig until a BAA or DPA is in place. Until then, the today-path is a zero-retention route (or content custody turned off), so ModelRig holds no content and grading is limited to the scores and metadata your own systems send.

Does your task fit? (four conditions)

Before you invest in a template, check the task against the four conditions the loop needs. A template shortens the path through them; it does not remove any.

1. Stable task distribution. Yesterday's examples still describe tomorrow's, so a golden set stays valid. If your categories or documents churn constantly, stabilize them first. 2. A rubric worth writing. Someone on your team can write down what a correct answer is. That written-down rubric is the grading guide. 3. Expert examples on hand. Someone who knows the work can grade a batch of past cases, so a grader can be calibrated against them. 4. Enough volume. The task runs often enough that calibrating a grader and optimizing a route pay back the effort.

Each template's fit note applies these four to that specific task and says what to do when one is shaky. A task that fails a condition structurally is honest to leave at the teaching stage rather than to force through the loop.

Two ways to use a template

A template is the same file whichever way you consume it. What ships in the modelrig npm package at node_modules/modelrig/templates/<use-case>.json is the machine-readable source; the pages here are the human-readable rendering of the same content.

A human copies from the docs

Open the template's page, read the fit note, and copy the pieces into your own route:

1. Take the starter schema, replace every TODO with your own taxonomy (fill each enum slot with your own labels), and save it as your route's output schema. 2. Take the starter guide criteria, sharpen each starter criterion in your team's own words, and set it as the route's grading guide. 3. Use the golden checklist to gather roughly twenty real examples that cover the classes and edges it names, and grade them as your expert set. 4. Use the escalation starter to decide what goes to a human and to set the guide's escalation band.

A coding agent feeds it through the existing tools

If you drive ModelRig from a coding agent (the shipped skill plus the modelrig-oracle MCP server), the template is consumed through the tools that already exist — no new tool is involved:

1. create_route writes a draft route bundle into your repo from the template's starter schema (files only; it never edits your source). You review the draft and fill in the TODO slots with your own taxonomy. 2. set_task_settings (the task-settings scope) authors the grading guide on that route from the template's starter guide, as-is; sharpen each starter criterion into your team's own definition as you go.

The agent's job is to present the filled-in draft and its evidence to you and get your go before anything lands — the template gives it a correct shape to start from, not a decision to make on your behalf.

What a template is not

• It is not a finished route. Every TODO is a slot only you can fill; the enum slots ship empty on purpose so an unfinished template cannot pretend to be done.
• It carries no claim about a model. Whether a model does your task well is something you measure with the goldens and grades the template helps you build, never something the template asserts.
• It does not set your pass threshold. That lives in the route's grade: block, not in the grading guide — the guide teaches the grader, the threshold turns a score into a pass or fail.