# Model tiering: standing rule for any agent stack

For an agent running on OpenRouter, a custom loop, LangGraph, CrewAI, an IDE agent, or anything
that is not Claude Code. Everything below this line is written for the agent. Paste it into the
agent's system prompt or standing instructions. Replace the three bracketed model names with the
ones you actually pay for (strongest, mid-tier, cheapest).

---

## A. How you route models

You route work to models by the kind of work, automatically. Nobody will tell you to save
tokens and there is no phrase to wait for. Cheaper-by-default is the standard, not a mode.

1. **Your main session runs on `[STRONGEST MODEL]`.** That is where you reason, plan, and
   decide. Never downgrade it.

2. **Everything you spawn defaults to `[MID-TIER MODEL]`.** That means every sub-agent, every
   worker, every tool-calling helper, every background task, every loop, every scheduled job. Do
   not pass the strongest model to a spawned job by habit. Summaries, classification, and log or
   transcript reduction go one tier lower, to `[CHEAPEST MODEL]`, because their output is checked
   by reading it against the source, not by which model wrote it.

3. **Five kinds of work keep `[STRONGEST MODEL]` with no justification needed:**
   - reasoning, planning, and strategy in your main session;
   - image and video creation (rendering, editing, composing);
   - the definition of website design (information architecture, wireframes, layout, design
     system, visual direction); implementing that design in code goes to the mid tier;
   - prose a human will read (client email, articles, social posts, newsletters, reports);
   - judging quality (QA gates, scoring, design and copy reviews), because a weaker judge
     passes weaker work.

4. **Anything else that wants `[STRONGEST MODEL]` must say why, in one line, in the spawning
   call:** `judgment: <reason>`. No line, no upgrade. Record every such override where it can be
   reviewed later. Before spawning ten helpers on the strong model, ask whether even one of them
   is doing thinking. Usually none are.

5. **Effort follows the same map as the model.** Where your models expose a reasoning or effort
   parameter, run it high in your main session and on the five jobs above, where the answer
   depends on wrestling with ambiguity. Run it medium on mid-tier implementation work. Run it low
   on reductions and summaries, where a test, a diff, or a re-read decides correctness anyway.
   High effort on a mechanical job is tokens wasted; low effort on a judgment call is a wrong
   answer arrived at quickly.

6. **Report weekly:** calls by the model that actually ran, how many were downgraded, how many
   opted up with `judgment:`.

## B. Make it mechanical in your own harness

A rule you have to remember is a rule you will break. Wherever your stack lets you, move these
out of memory and into configuration:

- **A default for spawned work.** Wherever your harness sets the model for sub-agents, workers,
  or tool-calling helpers, set it to `[MID-TIER MODEL]` once, so the cheap path is the default
  path and the strong model is the thing you opt into. On OpenRouter this is the model slug you
  pass for spawned calls; keep the strongest slug for the main session only.

- **A gate that rewrites.** If your harness supports a pre-call hook or middleware on the
  spawn step, use it: when a spawned call asks for `[STRONGEST MODEL]` without a `judgment:`
  line and is not one of the five reserved jobs above, rewrite it to `[MID-TIER MODEL]` and log the decision.
  If your harness has no hook, enforce it in the one function that creates spawned calls.

- **Pin every scripted job.** Every headless or scheduled run names its model and, where
  supported, its effort or reasoning level explicitly. Headless runs rarely inherit interactive
  defaults. Add a check that flags any new job without a pinned model.

- **Effort where the model exposes it.** Set the reasoning or effort parameter high on the main
  session, medium on mid-tier work, low on reductions. If a model does not expose one, the model
  tier alone still carries most of the saving.