Jules already reads AGENTS.md. Give it one that stops overpaying.

How this maps onto Jules

No skills directory. AGENTS.md is the whole mechanism.

Jules doesn't have a "skills" directory the way Claude Code does, and it isn't listed as a supported target on skills.sh either — there's nothing that auto-discovers skills/firstpass/SKILL.md on its own. What Jules does read automatically, per Jules's own getting-started docs, is a single root-level AGENTS.md: "Jules now automatically looks for a file named AGENTS.md in the root of your repository," used to help Jules "generate more relevant plans and completions." Jules's environment setup docs add that for simple projects without an explicit setup script, Jules also "refer[s] to agents.md or your readme.md file for hints to setup an environment on the fly." That's the entire documented surface — no .jules/ config directory, no per-user global file, no nested-file precedence rule. Undercut's policy has to be delivered as plain text inside that one file — see Install below for the exact mechanics.

The tier names — cheap, standard, frontier, apex — stay generic on purpose. models.md gives the rule of thumb rather than a fixed slug for Jules specifically: Jules runs on Google's own Gemini models rather than exposing a picker across vendors, so cheap resolves to the smallest/fastest model in Google's current Gemini lineup that Jules has access to, standard to its default working model, frontier to the highest-reasoning production model, and apex to a single batched tie-break call on the strongest one available. Nothing here sits between Jules and its model backend — it's instruction text Jules reads out of the repo at task time, no proxy, no network hop.

Install for Jules

Free. MIT. Nothing installs, nothing proxies your traffic.

Append it to AGENTS.md (the only path)

A root-level AGENTS.md is read in full by Jules on every task in the repo — so appending the policy text there is the only reliable way to get it in front of Jules, since it doesn't depend on Jules separately choosing to open another file, and there's no per-user or per-workspace equivalent to fall back to.

If the repo has no AGENTS.md yet, the same command creates one. Already run this policy through Codex, Copilot, or another AGENTS.md-reading client in the same repo? Jules reads the same file — nothing extra to install.

Validate it yourself

1. 1Read the published results — testing/README.md and the raw JSON in testing/results/
2. 2Reproduce for free, no API key, no spend: node src/main.js --mock
3. 3Kick off a small Jules task against a repo with the policy installed and check its plan/output for which tier it names before starting mechanical work

Zero lock-in. It's instruction text Jules reads — delete the block and you're back to status quo.

Six flags assign a tier. Three triggers are the only way up.

This is a named, versioned spec — not internal logic we could quietly change. When the rubric changes (a new flag, a different cap), the version number moves and the changelog says why.

Every unit of work gets scored against six flags, then assigned the cheapest tier that can plausibly handle it. It only moves up a tier when something objective proves it needs to — a failed check, a measured disagreement, an explicit uncertainty flag. It never moves up on a guess, and it never moves back down mid-task.

 unit of work
      |
  6-flag rubric
      |
  +---+----+
  | 0 flags|--> cheap
  | 1-2    |--> standard
  | 3+/own |--> frontier
  +---+----+
      | fail x2 / disagree / uncertain
      v
  escalate +1 tier
  (residue only)
      |
  still unresolved?
      v
  single batched
  apex tie-break

Why it holds up

Flags steer. Verification and escalation decide.

Even when the rubric flags are scored imperfectly, the escalator still routes work to the right tier almost every time — because a wrong flag only ever costs one extra cheap attempt, never a wrong answer or a big bill.

Honest limits

What this doesn't do.

Not a proxy, not a gateway, doesn't auto-flag in production — and on Jules specifically, it's not a "skill" Jules discovers on its own: it only takes effect if you've put it in the root AGENTS.md that Jules reads (see Install above), and it's not listed on skills.sh as an install target. Jules's own reliability at scoring the flags hasn't been separately measured the way Claude Code's has (see above); treat that as unverified until tested. It doesn't promise a dollar figure for your workload, and because Jules runs each task in an isolated cloud VM rather than a session you watch live, there's no interactive way to confirm mid-run which tier it picked — only after the fact, from its plan or output. Full list of general limits, with the finance/procurement note and the paid-tier disclosure, is on the main page.