frontier

Make Claude Opus, Sonnet, GPT, or Gemini produce work close to what Claude Fable 5 would ship: Fable 5 itself wrote and audited these 21 standards, so the model you already have executes against its bar. One response gets the lift; the optional convergence loop and taste gate carry work that must be right.

/frontier <deliverable> [quick|full|gate]

Why · Output · One response · How · Inside · Install · Cost · Compared · Limits · FAQ

Before release, this system was turned on itself, and it did not pass on the first try. The distillation run is the adversarial audit of the standards themselves; the self-run publishes every gate verdict this README received, failures first, with the fixes each one forced.

Quality is a procedure, not a property

Ask a model to "make it great" and you get its training-data average: the same hero layout, the same "seamlessly leverage" copy, the same confident report nobody verified. The gap between that and frontier output is mostly not intelligence. It is method: strong models define the standard before generating, sample several attempts instead of polishing the first, verify against real evidence with fresh eyes, and refuse to stop at "looks done".

Method can be written down. frontier is that method, packaged: one command that runs any task through written quality standards, independent candidate generation, evidence-grounded verification sweeps, and a final taste judgment, until the work converges instead of merely ending.

Output is parseable findings, not vibes

Verifier findings arrive pinned to a location and a rubric line, in a fixed shape a script (or a tired human) can walk (quoted from the sample run):

LENS: layout
FINDINGS:
1. tiers section, 768px | comparison table scannable in 15s | table forces horizontal scroll at tablet width | confidence: h
2. hero, 390x844 | claim + CTA in first viewport | CTA sits 64px below the fold on mobile | confidence: h
3. tier cards | one primary action per view | "Start free" and "Book a demo" carry equal visual weight on the Studio tier | confidence: h
CHECKED: rubric lines 1, 2, 6, 7 via screenshots at three widths, cropped
NOT CHECKABLE: line 5 (FAQ content is the copy lens)

The taste gate's block, from the same run:

GATE: fail
FINDINGS:
1. tier names | brand owner | "Starter, Growth, Studio" could be any SaaS; the product voice is trade-specific everywhere else | rename from the studio world | confidence: m
2. annual toggle | first-time audience | eye lands on the calculator, then tiers; the toggle registers on second read only | move it into the tier-card header row | confidence: h
RANKING: n/a
DISTILL:
- marketing.md candidate: pricing toggles live where the eye decides (the tier header),
  not above the section; a toggle seen after the price anchors monthly

An empty findings list is itself a claim: it means every rubric line was actively checked and nothing surfaced. A full run, end to end: examples/sample-run.md. The real thing, run on this repo: examples/self-run.md.

It works in a single response

The gap-closer needs no agent and no loop. Attach the matching craft file, and the model writes the rubric, drafts against it, then runs one fresh-eyes judge pass on its own output and fixes what it finds, all inside one reply. Same prompt, same model, different floor: the standards supply the taste the model would otherwise average away, and the judged pass catches what the draft defended. PROMPT.md ships exactly this shape for chat surfaces; quick mode is its Claude Code twin. The convergence loop below is the optional assurance tier on top, not the price of entry.

How it works

flowchart TD
    P0["Phase 0: Scope and arm<br/>route to craft standards, write the rubric,<br/>constraint ledger, part inventory"]
    P0 --> P1["Phase 1: Candidates (creative work)<br/>3-5 independent attempts, distinct angles,<br/>taste gate ranks, winner grafts the rest"]
    P1 --> P2["Phase 2: Produce<br/>one concern per step,<br/>rubric re-read before each part"]
    P2 --> P3["Phase 3: Evidence<br/>screenshots at 3 widths, test runs, probes,<br/>frame scrubs, recomputed numbers"]
    P3 --> P4["Phase 4: Fresh-eyes sweeps<br/>one judge per lens, every finding reported,<br/>pass ledger kept"]
    P4 --> Q1{"Two consecutive<br/>clean passes?"}
    Q1 -- "no: fix everything" --> P3
    Q1 -- "yes" --> P5["Phase 5: Taste gate (high stakes)<br/>3-lens panel; DISTILL banks the call"]
    P5 --> DONE["Report: outcome, evidence,<br/>pass ledger, decisions, unverified"]

That diagram is full mode. quick collapses phases 1, 4, and 5 into a single judged pass; the spine (rubric, produce, judge, fix) survives even in one response.

Three mechanisms do the heavy lifting:

The 10-minute deep dive: docs/HOW-IT-WORKS.md.

What's inside

21 craft standards, each defining excellent in checkable numbers, with a ban list of machine tells and a per-domain verification checklist:

A taste of the rules (each file carries 34 to 59 of these, counted after the last edit):

writing: no three consecutive sentences within 3 words of the same length; scan for machine-cadence tells: claim triples ("fast, simple, and secure"), trailing participles ("..., making it easier than ever"), symmetric negation ("No setup. No config. Just results.")

design: the primary claim and its CTA sit fully inside the first viewport at 1440x900 AND 390x844; hairline borders are the ink color at 6-12% alpha, never default gray-200

data: the classic "drop in the last week" is an incomplete week; check freshness before insight. A surprising number is a pipeline bug until the joins are checked.

decisions: the flip test: write down what evidence would change your mind; if nothing would, it is not a decision, it is a commitment already made

code: a test counts only if it fails when the change is reverted; a test that cannot fail proves nothing

Plus the protocol (the ten laws, weaker-model compensations, ceiling raisers, lessons recorded from a frontier model) and the judges (fresh-eyes verifier, 3-lens taste gate, panel), portable to any surface.

The origin: standards that carry their own audit trail

The kit was authored and then adversarially audited by Claude Fable 5 (Anthropic's frontier tier) in July 2026, in the final days of its general access: 8 auditor agents in fresh contexts, 7 sweeping the 21 craft files and 1 reviewing the judge prompts, agents, and skill. About 260 documented change entries came back: vague lines became numbers, rules a literal-minded model could satisfy in letter while missing in spirit got tightened, and, most unusually, the model wrote down its OWN tells as ban-list entries: the machine-cadence prose tics,