model-right-sizer-install — stamp the right-sizing mandate onto a repo

This skill installs one thing, and only one thing: a standalone, organization-agnostic mandate that every substantive task in this repo must consult the model-right-sizer agent for model-selection guidance, bookended around the work itself. It targets whichever agent-instructions file(s) the repo actually uses — CLAUDE.md, AGENTS.md, or both — see step 3 below. (It will also install the model-right-sizer plugin itself, if the agent it points at isn't already discoverable — see step 2 below — but that's making its own prerequisite true, not scope creep.) It does not install a broader development process (no worktree discipline, no review pipeline, no ticket reconciliation) — those belong to whatever flow your own org already runs, if any. This mandate is scoped narrowly on purpose so any repo, in any org, can adopt it independent of everything else.

It is the companion to model-right-sizer.md: that file is the portable agent; this skill is the portable mandate to use it. Both are meant to be dropped (or submoduled + composed, per the plugin README) into a target repo's Claude tooling — this skill is not something you run against this repo, it's something you run in the repo you want to onboard.

What "every turn" means in practice

Not every keystroke deserves a blueprint — a one-line clarifying question doesn't spend meaningful model budget. Scope "every turn" to: any turn where the assistant is about to do non-trivial work — write or edit code or docs, dispatch a sub-agent, run a multi-step tool chain, or otherwise spend a real slice of model reasoning. Trivial conversational turns are exempt, and model-right-sizer already scales its own output to the size of the task (one-line skip for a tiny change, full blueprint/report for a substantial one) — see model-right-sizer.md's "Your output shape" section. Don't fight that scaling by forcing a full blueprint on every micro-edit.

What to do

1. Confirm the target. This skill always targets the current repo — the one the user wants onboarded — never a copy elsewhere.

2. Check the agent — and the dry-run skill — are actually discoverable, and install what's missing. The mandate's "before" hook (step 3 below) now invokes model-right-sizer-dryrun directly, not just the agent, so both need to resolve. Look for model-right-sizer.md somewhere your tooling reads agent personas from (e.g. .claude/agents/model-right-sizer.md, or a plugin's agents/model-right-sizer.md) and for model-right-sizer-dryrun somewhere it reads skills from. If either is not there:

   • Tell the user the agent and/or skill isn't discoverable and that fixing it means adding the CloudZero marketplace and installing the plugin — then, with their go-ahead, run:

     /plugin marketplace add cloudzero/cloudzero-claude-marketplace
     /plugin install model-right-sizer@cloudzero

     Do not run these without confirming first — installing a plugin is a change to the user's environment, not just this repo. This installs the agent and both companion skills (model-right-sizer-dryrun, model-right-sizer-install) together. Then re-check discoverability — the plugin's agents/model-right-sizer.md and skills/model-right-sizer-dryrun/ should now resolve.
   • If the plugin-install commands aren't available in this runtime (this isn't Claude Code, or plugin installs are disabled here), fall back to telling the user the mandate will point at an agent/skill that doesn't exist yet, and offer the manual fix instead: copy plugins/model-right-sizer/agents/model-right-sizer.md into .claude/agents/ and plugins/model-right-sizer/skills/model-right-sizer-dryrun/ into .claude/skills/ (both from https://github.com/cloudzero/cloudzero-claude-marketplace) in this repo, or follow the plugin README's submodule + compose pattern if they want it to stay upgradeable independently. If only the agent gets copied and not the skill, say so explicitly in the report — the mandate's "before" hook would then be pointing at a skill that doesn't exist.
   • If the agent and skill are present or were just installed, proceed to stamp the mandate below. If either is still missing (plugin install unavailable and no manual copy yet), do NOT stamp silently: tell the user the mandate would point at something that doesn't exist yet, give the manual fix above, and ask whether to stamp anyway (as a statement of intent) or stop until both are installed. Either way, say plainly in your report whether the agent and skill were already present, freshly installed, or still missing — don't silently assume either is wired.

3. Detect which file(s) to target: CLAUDE.md, AGENTS.md, or both. Look in the repo root for each. Do not assume — check for both, every time, since a repo can adopt either convention (or migrate between them) independently of this skill.

   • Only AGENTS.md exists (no CLAUDE.md) → target AGENTS.md alone. Do not also create a CLAUDE.md — the repo has already opted into the AGENTS.md convention, respect that instead of adding a second, redundant instructions file it didn't ask for.
   • Only CLAUDE.md exists (no AGENTS.md) → target CLAUDE.md alone, as before.
   • Both exist → target both, independently — the mandate block goes in each file, so whichever one a given tool actually reads, the mandate is there.
   • Neither exists → target CLAUDE.md (create it fresh, with just this block). This is the same default this skill has always had; AGENTS.md isn't invented as a target unless the repo already has one.

   Say in your step 6 report which file(s) you found and which you targeted — don't let this decision happen silently.

4. Insert or refresh the marker-delimited block below in every file targeted in step 3 — append-only, per file: create the file if it's the target and doesn't exist yet (the "neither exists" case above); if a targeted file already exists, replace only the text between the markers in that file, leaving everything else in it untouched. The block's content is identical regardless of which file(s) it lands in.

   <!-- model-right-sizer-mandate:begin (managed by model-right-sizer-install — do not hand-edit; re-run the skill to refresh) -->
   ## MANDATE — consult model-right-sizer before and after every substantive turn
   
   **Before starting any substantive task** in this repo (writing or editing
   code or docs, dispatching a sub-agent, running a multi-step tool chain),
   run the **`model-right-sizer-dryrun`** skill against the task at hand (its
   plain-English description as the intent). That returns a **right-sizing
   blueprint** — a single JSON object, schema-conformant per
   `model-right-sizer`'s `schemas/blueprint.schema.json` — saying which model
   and effort level each piece of the work should run on, weighing
   effectiveness need against efficiency pressure and difficulty, biased
   toward the smallest model that clears the bar. Hand that JSON blueprint to
   the orchestrating session/agent driving this task, and route every
   sub-agent or model dispatch by its `blueprint_rows` / `work_routing_map`
   picks rather than by default or gut feel.
   
   **After the task closes**, consult the `model-right-sizer` agent directly
   for a **model-usage report** (this pass runs the agent itself, not the
   dry-run skill — there's a real "actual" to reconcile against now, which a
   dry run never has): recommended vs. actual model per stage, the reasoning
   behind each pick, and an order-of-magnitude cost estimate — so model spend
   on this repo stays as measurable and intentional as any other resource it
   consumes.
   
   Both passes scale to the size of the task — a trivial turn gets a one-line
   skip, a substantial one gets the full blueprint (Pass A) or report (Pass
   B). `model-right-sizer` is **read-only**: it reports and recommends, it
   never edits files or makes the call for you — you (or your own tooling)
   apply its picks.
   
   This mandate is intentionally narrow — model selection only. It does not
   presume any particular development process. If this repo already follows
   a broader flow (design review, code review, integration testing, etc.),
   fold these two consults into that flow's existing bookends rather than
   running them as a separate, disconnected step.
   
   Do not edit this block by hand — re-run the `model-right-sizer-install`
   skill (source: https://github.com/cloudzero/cloudzero-claude-marketplace)
   to refresh it, so the wording stays traceable to one source instead of
   drifting per repo.
   <!-- model-right-sizer-mandate:end -->

5. Verify, don't assume. Re-read every file targeted in step 3 after writing and confirm the markers and content landed as expected in each one — don't declare success on the strength of only one when two were targeted.

6. Report what happened: which file(s) you found (CLAUDE.md, AGENTS.md, or both) and targeted; created vs. refreshed, per file; and whether the agent and the dry-run skill were already present, freshly installed via the plugin, or — if plugin-install wasn't available and either is still missing — the one-line manual instruction to add it.

Why this stays isolated from any bigger flow

Some organizations run this same mandate as one bookend inside a much larger process (design review → right-sizing → build → code review → integration → …). That full process is out of scope for this repo by design — model-right-sizer ships as a portable, single-agent core (see the plugin README's scope discipline), and this skill mirrors that discipline: it installs only the right-sizing mandate, worded generically enough to slot into any org's existing process or stand alone with none. If you want to layer your own organization's broader flow on top, write your own overlay skill the same way the README's "Extending this agent for your own organization" section suggests overlaying the agent file itself — don't fork this skill's text to do it.

Related

• model-right-sizer.md — the agent this mandate points at, and the sole source for Pass B (the usage report).
• model-right-sizer-dryrun — the skill this mandate's "before" hook now runs directly to produce the JSON blueprint; see its own file for what it returns and why.
• README.md — "Using it" / "Importing it into another repo without duplicating it" — how to get the agent file itself into a target repo, the same pattern this skill's file follows.