iamneilroberts Frontier Spend
Overview
A scarce premium-tier model is a non-renewable judgment credit. The only value that survives the cap or the cliff is what lands in durable artifacts (plans, verdicts, audits, tests) that a cheaper model executes later. Most agents get "spend on judgment, not typing" right on their own — the four rules below are the ones they miss.
The four rules (each one counters a default failure)
- Adjudicate, don't author. Do NOT have the premium model write specs/plans/ADRs from scratch — planning is its fastest token burn (practitioner reports: ~35% of a usage window on one planning pass). Instead: cheaper model drafts → an independent cross-vendor reviewer audits → premium model adjudicates the disagreements and patches. Same banked artifact, several times more of them per cap. From-scratch authoring only after a cheap draft comes back structurally wrong twice.
- Ration audits to a named budget. Comprehensive sweeps feel high-value and are the fastest way to exhaust a cap. Pick the audit count up front (default: two) and name them (e.g. golden-path adversarial review + tech-debt sweep). Everything else gets triage-grade attention or none.
- Tune the effort slider. Reduced reasoning effort for triage-grade judgment (queue ranking, routine verdicts) — low-effort premium can undercut the mid-tier model's cost. Max effort only for the named audits and plan adjudication.
- Review the premium output too. The premium model is smarter AND more confidently wrong — keep the external review gate on its artifacts, don't exempt them because "the smart model wrote it".
Native mechanism for rule 1 (Claude Code v2.1.170+): the Advisor tool
Anthropic shipped a first-party Advisor tool (code.claude.com/docs/en/advisor) that
implements "premium judges, doesn't type" directly: a cheaper executor model consults a
higher-intelligence advisor model mid-generation. The relevant pairing for a capped
premium tier is executor-main + premium-advisor (e.g. Sonnet main + Fable advisor):
premium guidance at decision points without running the premium model throughout. Prefer
this over a hand-rolled adjudication loop for mid-task decisions when it is available.
It does not replace the rest of the doctrine, and three known limits are why:
- Server-side, not hookable. You cannot gate it to specific phases — it is a prose-only rule available to any eligible agent. Rationing (rule 2) and effort tuning (rule 3) remain your manual discipline.
- It still spends the cap, unbounded. A consult on a large session burns premium budget; budget-first math still applies.
- It vanishes at the cliff. When affordable premium access ends, the premium-advisor pairing goes with it — so the durable-artifact core (bank plans/audits/tests a cheap model executes later) is what survives, unchanged. Use the Advisor tool for live decisions now; keep banking artifacts for after.
Standing hygiene (agents mostly do this unprompted; keep anyway)
- Pre-digest context: open every premium session with a brief built by a cheap model (≤2k words); never let the premium model explore a repo raw.
- One judgment deliverable per session; if it drifts mechanical, hand off and switch models mid-session.
- No long sequential checklists (known rogue failure mode) and no bulk reads.
- Budget math first: a 50% weekly cap at 2× burn ≈ 4–6 tight sessions, not "3 days".
Red flags — you are about to waste the credit
- "I'll just have the premium model draft the plan itself" → rule 1.
- "While we're at it, let's also audit X" beyond the named budget → rule 2.
- Running max effort on a triage question → rule 3.
- Shipping a premium-authored artifact unreviewed → rule 4.
- The premium session's first action is a grep/read of the repo → hygiene 1.