[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-92383":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":15,"stars7d":15,"stars30d":16,"stars90d":15,"forks30d":15,"starsTrendScore":15,"compositeScore":17,"rankGlobal":10,"rankLanguage":10,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":19,"hasPages":19,"topics":21,"createdAt":10,"pushedAt":10,"updatedAt":32,"readmeContent":33,"aiSummary":34,"trendingCount":15,"starSnapshotCount":15,"syncStatus":35,"lastSyncTime":36,"discoverSource":37},92383,"frontier","apoorvjain25\u002Ffrontier","apoorvjain25","Make Claude Opus, Sonnet, GPT, or Gemini produce work close to what Claude Fable 5 would ship. 21 craft standards written and audited by the frontier model itself; single-response lift, or a full convergence loop with a taste gate.","",null,"JavaScript",76,14,54,0,22,45.73,"MIT License",false,"main",[22,23,24,25,26,27,28,29,30,31],"ai-agents","anthropic","claude","claude-code","code-quality","gemini","gpt","llm","prompt-engineering","skills","2026-07-22 04:02:06","\u003Cdiv align=\"center\">\n\n\u003Cimg src=\"assets\u002Fhero.png\" alt=\"frontier: quality is a procedure, not a property\" width=\"100%\">\n\n# frontier\n\n**Make Claude Opus, Sonnet, GPT, or Gemini produce work close to what Claude Fable 5 would\nship: Fable 5 itself [wrote and audited](examples\u002Fthe-distillation-run.md) these 21\nstandards, so the model you already have executes against its bar. One response gets the\nlift; the optional convergence loop and taste gate carry work that must be right.**\n\n[![License: MIT](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-MIT-2A3242.svg?labelColor=0B0E14)](LICENSE)\n[![Claude Code Skill](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FClaude%20Code-skill%20%2B%20plugin-2A3242.svg?labelColor=0B0E14)](https:\u002F\u002Fdocs.claude.com\u002Fen\u002Fdocs\u002Fclaude-code)\n[![Craft standards](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fcraft%20standards-21-2A3242.svg?labelColor=0B0E14)](frontier\u002Freferences\u002Fcraft)\n[![Judges](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fjudges-verifier%20·%20taste%20gate%20·%20panel-2A3242.svg?labelColor=0B0E14)](frontier\u002Freferences\u002Fjudges.md)\n[![Works everywhere](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Falso%20runs%20in-claude.ai%20·%20Cursor%20·%20Windsurf%20·%20aider-2A3242.svg?labelColor=0B0E14)](PROMPT.md)\n[![Target models](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Ffor-Opus%20·%20Sonnet%20·%20GPT%20·%20Gemini-2A3242.svg?labelColor=0B0E14)](#faq)\n\n```\n\u002Ffrontier \u003Cdeliverable> [quick|full|gate]\n```\n\n[Why](#quality-is-a-procedure-not-a-property) · [Output](#output-is-parseable-findings-not-vibes) ·\n[One response](#it-works-in-a-single-response) · [How](#how-it-works) · [Inside](#whats-inside) ·\n[Install](#install) · [Cost](#token-cost-bounded-on-purpose) · [Compared](#how-it-compares) ·\n[Limits](#limitations) · [FAQ](#faq)\n\n\u003C\u002Fdiv>\n\nBefore release, this system was turned on itself, and it did not pass on the first try.\n[The distillation run](examples\u002Fthe-distillation-run.md) is the adversarial audit of the\nstandards themselves; [the self-run](examples\u002Fself-run.md) publishes every gate verdict\nthis README received, failures first, with the fixes each one forced.\n\n## Quality is a procedure, not a property\n\nAsk a model to \"make it great\" and you get its training-data average: the same hero layout,\nthe same \"seamlessly leverage\" copy, the same confident report nobody verified.\nThe gap between that and frontier output is mostly not intelligence. It is method: strong\nmodels define the standard before generating, sample several attempts instead of polishing\nthe first, verify against real evidence with fresh eyes, and refuse to stop at \"looks done\".\n\nMethod can be written down. frontier is that method, packaged: one command that runs any\ntask through written quality standards, independent candidate generation, evidence-grounded\nverification sweeps, and a final taste judgment, until the work converges instead of merely\nending.\n\n## Output is parseable findings, not vibes\n\nVerifier findings arrive pinned to a location and a rubric line, in a fixed shape a script\n(or a tired human) can walk (quoted from [the sample run](examples\u002Fsample-run.md)):\n\n```\nLENS: layout\nFINDINGS:\n1. tiers section, 768px | comparison table scannable in 15s | table forces horizontal scroll at tablet width | confidence: h\n2. hero, 390x844 | claim + CTA in first viewport | CTA sits 64px below the fold on mobile | confidence: h\n3. tier cards | one primary action per view | \"Start free\" and \"Book a demo\" carry equal visual weight on the Studio tier | confidence: h\nCHECKED: rubric lines 1, 2, 6, 7 via screenshots at three widths, cropped\nNOT CHECKABLE: line 5 (FAQ content is the copy lens)\n```\n\nThe taste gate's block, from the same run:\n\n```\nGATE: fail\nFINDINGS:\n1. 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\n2. 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\nRANKING: n\u002Fa\nDISTILL:\n- marketing.md candidate: pricing toggles live where the eye decides (the tier header),\n  not above the section; a toggle seen after the price anchors monthly\n```\n\nAn empty findings list is itself a claim: it means every rubric line was actively checked\nand nothing surfaced. A full run, end to end:\n[examples\u002Fsample-run.md](examples\u002Fsample-run.md). The real thing, run on this repo:\n[examples\u002Fself-run.md](examples\u002Fself-run.md).\n\n## It works in a single response\n\nThe gap-closer needs no agent and no loop. Attach the matching craft file, and the model\nwrites the rubric, drafts against it, then runs one fresh-eyes judge pass on its own output\nand fixes what it finds, all inside one reply. Same prompt, same model, different floor: the\nstandards supply the taste the model would otherwise average away, and the judged pass\ncatches what the draft defended. [PROMPT.md](PROMPT.md) ships exactly this shape for chat\nsurfaces; `quick` mode is its Claude Code twin. The convergence loop below is the optional\nassurance tier on top, not the price of entry.\n\n## How it works\n\n```mermaid\nflowchart TD\n    P0[\"Phase 0: Scope and arm\u003Cbr\u002F>route to craft standards, write the rubric,\u003Cbr\u002F>constraint ledger, part inventory\"]\n    P0 --> P1[\"Phase 1: Candidates (creative work)\u003Cbr\u002F>3-5 independent attempts, distinct angles,\u003Cbr\u002F>taste gate ranks, winner grafts the rest\"]\n    P1 --> P2[\"Phase 2: Produce\u003Cbr\u002F>one concern per step,\u003Cbr\u002F>rubric re-read before each part\"]\n    P2 --> P3[\"Phase 3: Evidence\u003Cbr\u002F>screenshots at 3 widths, test runs, probes,\u003Cbr\u002F>frame scrubs, recomputed numbers\"]\n    P3 --> P4[\"Phase 4: Fresh-eyes sweeps\u003Cbr\u002F>one judge per lens, every finding reported,\u003Cbr\u002F>pass ledger kept\"]\n    P4 --> Q1{\"Two consecutive\u003Cbr\u002F>clean passes?\"}\n    Q1 -- \"no: fix everything\" --> P3\n    Q1 -- \"yes\" --> P5[\"Phase 5: Taste gate (high stakes)\u003Cbr\u002F>3-lens panel; DISTILL banks the call\"]\n    P5 --> DONE[\"Report: outcome, evidence,\u003Cbr\u002F>pass ledger, decisions, unverified\"]\n```\n\nThat diagram is `full` mode. `quick` collapses phases 1, 4, and 5 into a single judged\npass; the spine (rubric, produce, judge, fix) survives even in one response.\n\nThree mechanisms do the heavy lifting:\n\n| Mechanism | What it exploits |\n|---|---|\n| **Best-of-N candidates** | A model's best of five attempts sits far above its average attempt. Sampling the tail is where frontier-grade output lives. |\n| **Fresh-eyes verification** | The context that produced work defends it; a fresh context finds what the author rationalizes. Judges only find, never fix. |\n| **The taste gate + DISTILL** | Judging costs a small fraction of generating, so the strongest model available reviews everything, and every taste call it makes is converted into a permanent written rule. The system absorbs taste instead of renting it. |\n\nThe 10-minute deep dive: [docs\u002FHOW-IT-WORKS.md](docs\u002FHOW-IT-WORKS.md).\n\n## What's inside\n\n**21 craft standards**, each defining excellent in checkable numbers, with a ban list of\nmachine tells and a per-domain verification checklist:\n\n| | | |\n|---|---|---|\n| [design](frontier\u002Freferences\u002Fcraft\u002Fdesign.md) | [motion](frontier\u002Freferences\u002Fcraft\u002Fmotion.md) | [writing](frontier\u002Freferences\u002Fcraft\u002Fwriting.md) |\n| [code](frontier\u002Freferences\u002Fcraft\u002Fcode.md) | [research](frontier\u002Freferences\u002Fcraft\u002Fresearch.md) | [prompting](frontier\u002Freferences\u002Fcraft\u002Fprompting.md) |\n| [product](frontier\u002Freferences\u002Fcraft\u002Fproduct.md) | [data](frontier\u002Freferences\u002Fcraft\u002Fdata.md) | [security](frontier\u002Freferences\u002Fcraft\u002Fsecurity.md) |\n| [ops](frontier\u002Freferences\u002Fcraft\u002Fops.md) | [media](frontier\u002Freferences\u002Fcraft\u002Fmedia.md) | [marketing](frontier\u002Freferences\u002Fcraft\u002Fmarketing.md) |\n| [decisions](frontier\u002Freferences\u002Fcraft\u002Fdecisions.md) | [sales](frontier\u002Freferences\u002Fcraft\u002Fsales.md) | [teaching](frontier\u002Freferences\u002Fcraft\u002Fteaching.md) |\n| [management](frontier\u002Freferences\u002Fcraft\u002Fmanagement.md) | [storytelling](frontier\u002Freferences\u002Fcraft\u002Fstorytelling.md) | [academic](frontier\u002Freferences\u002Fcraft\u002Facademic.md) |\n| [career](frontier\u002Freferences\u002Fcraft\u002Fcareer.md) | [translation](frontier\u002Freferences\u002Fcraft\u002Ftranslation.md) | [coordination](frontier\u002Freferences\u002Fcraft\u002Fcoordination.md) |\n\nA taste of the rules (each file carries 34 to 59 of these, counted after the last edit):\n\n> **writing**: no three consecutive sentences within 3 words of the same length; scan for\n> machine-cadence tells: claim triples (\"fast, simple, and secure\"), trailing participles\n> (\"..., making it easier than ever\"), symmetric negation (\"No setup. No config. Just results.\")\n\n> **design**: the primary claim and its CTA sit fully inside the first viewport at 1440x900\n> AND 390x844; hairline borders are the ink color at 6-12% alpha, never default gray-200\n\n> **data**: the classic \"drop in the last week\" is an incomplete week; check freshness before\n> insight. A surprising number is a pipeline bug until the joins are checked.\n\n> **decisions**: the flip test: write down what evidence would change your mind; if nothing\n> would, it is not a decision, it is a commitment already made\n\n> **code**: a test counts only if it fails when the change is reverted; a test that cannot\n> fail proves nothing\n\nPlus [the protocol](frontier\u002Freferences\u002Fprotocol.md) (the ten laws, weaker-model\ncompensations, ceiling raisers, lessons recorded from a frontier model) and\n[the judges](frontier\u002Freferences\u002Fjudges.md) (fresh-eyes verifier, 3-lens taste gate, panel),\nportable to any surface.\n\n## The origin: standards that carry their own audit trail\n\nThe kit was authored and then adversarially audited by Claude Fable 5 (Anthropic's frontier\ntier) in July 2026, in the final days of its general access: 8 auditor agents in fresh\ncontexts, 7 sweeping the 21 craft files and 1 reviewing the judge prompts, agents, and\nskill. About 260\ndocumented change entries came back: vague lines became numbers, rules a literal-minded\nmodel could satisfy in letter while missing in spirit got tightened, and, most unusually,\nthe model wrote down its OWN tells as ban-list entries: the machine-cadence prose tics, the\ndefault design habits, the hedge-everything analysis patterns, the fiction cliches. The\nprompt review alone returned 42 findings, all applied. The full run, with its real ledgers:\n[examples\u002Fthe-distillation-run.md](examples\u002Fthe-distillation-run.md).\n\nThe expensive model set the bar once, in writing; clearing it no longer takes the expensive\nmodel.\n\n## Install\n\n**Claude Code, as a plugin** (skills + the two judge agents):\n\n```\n\u002Fplugin marketplace add apoorvjain25\u002Ffrontier\n\u002Fplugin install frontier@apoorvjain25\n```\n\n**Claude Code, as a plain skill**: copy the inner [`frontier\u002F`](frontier) folder to\n`~\u002F.claude\u002Fskills\u002Ffrontier\u002F`, and optionally [`agents\u002F`](agents) to `~\u002F.claude\u002Fagents\u002F`.\n\n**claude.ai and Cowork**: upload `frontier-skill.zip` as a custom skill (Settings,\nCapabilities), or paste [PROMPT.md](PROMPT.md) plus the relevant craft file into a Project.\n\n**Cursor, Windsurf, aider, raw API**: paste [PROMPT.md](PROMPT.md), attach the craft file\nmatching your domain, put your task last.\n\nDetails and troubleshooting: [docs\u002FINSTALL.md](docs\u002FINSTALL.md).\n\n## Usage\n\n| Command | What you get |\n|---------|--------------|\n| `\u002Ffrontier the pricing page` | the default is `full`: candidates, convergence loop, taste gate, report |\n| `\u002Ffrontier fix the export flow quick` | rubric, produce, ONE judge pass, fixes; no loop, no gate |\n| `\u002Ffrontier apps\u002Fweb\u002Fhero.tsx gate` | taste-judge existing work in one pass, nothing modified |\n| `\u002Ffrontier the launch email` | any of the 21 domains; routing is automatic |\n\n## Token cost: bounded, on purpose\n\nA convergence loop spends more tokens than a one-shot prompt. That is the trade:\n\n- **Modes size the spend** (defined in Usage above): as author estimates from development\n  runs, `quick` lands around 1.5-2 times a one-shot, `full` around 5-9 times, `gate` a\n  single judged pass; the one measured figure sits in the [FAQ](#faq).\n- **Hard cap**: 8 whole-deliverable passes; anything still open at the cap goes into the\n  report for you to see.\n- **Cheap judges**: verification passes cost little relative to generation, because judges\n  read and report.\n\n## How it compares\n\nRubrics, style guides, and self-critique are prior art; none of that is new here. What\nfrontier adds is the combination; the table enumerates it.\n\n| | CLAUDE.md \u002F Cursor rules | One-shot mega-prompt | [production-audit](https:\u002F\u002Fgithub.com\u002Fapoorvjain25\u002Fproduction-audit) | frontier |\n|---|---|---|---|---|\n| Scope | project conventions | one task, one pass | finding what is wrong in existing products | building new work to a standard, any domain |\n| Standards | prose preferences | implied by adjectives | a defect-class taxonomy | 21 domains in checkable numbers + ban lists |\n| Verification | none | the model says it checked | per-finding, against the code | fresh-eyes judges against rendered evidence |\n| Stop condition | n\u002Fa | the response ended | two quiet passes across its lens catalog | earned: one judged pass (quick) to two clean sweeps + gate (full) |\n| Improves over time | manual edits | no | lens PRs | DISTILL: every taste call becomes a rule |\n| Token cost | ~free | low | high: many sweep passes to convergence | 1.5-9x a one-shot (author estimate), mode-sized, capped |\n\nproduction-audit is the sibling: it tears down what exists, frontier builds what is next,\nand they share the convergence philosophy.\n\n## Limitations\n\n- **Not deterministic.** Two runs sample different candidates and can converge on different\n  results. It raises the floor and the ceiling; it does not make output reproducible.\n- **It costs real tokens.** A `full` run is a multiple of a one-shot by design. Use `quick`\n  for routine work, `gate` to judge without rebuilding; the 8-pass cap bounds a `full` run.\n- **The taste ceiling is the judge's ceiling.** A model judging its own tier plateaus below\n  a stronger model's eye; the DISTILL flywheel narrows this over time rather than erasing\n  it on day one.\n- **Weaker off Claude Code.** Chat surfaces have no subagents and no screenshot or test\n  tooling; judge passes run one after another in the same context, and unverifiable claims\n  land in UNVERIFIED instead of being checked.\n- **The standards are opinionated defaults.** They encode a specific bar (honest odd\n  numbers over round vanity stats, one accent with a locked meaning, machine-cadence tells\n  banned). Your brand tokens and your edits always win.\n\n## Where things live\n\n```\nfrontier\u002F\n├── SKILL.md                      # the procedure: 6 phases, modes, pass cap, report format\n└── references\u002F\n    ├── protocol.md               # ten laws, compensations, ceiling raisers, frontier lessons\n    ├── judges.md                 # verifier + taste gate + panel, portable and parseable\n    └── craft\u002F                    # 21 standards: numbers, ban lists, checklists\nagents\u002F\n├── verifier.md                   # fresh-eyes finder (never fixes)\n└── taste-judge.md                # 3-lens gate with DISTILL\ndocs\u002F                             # how it works, install, customizing\nexamples\u002F\n├── the-distillation-run.md       # real: 8 auditors, ~260 change entries, the ledgers\n├── self-run.md                   # real: this repo through its own gate\n└── sample-run.md                 # illustrative: a full run, annotated\nPROMPT.md                         # the whole method in one paste-able file\n.claude-plugin\u002F                   # plugin + marketplace manifests\nfrontier-skill.zip                # ready upload for claude.ai custom skills\n```\n\n## Make it yours\n\nEvery gate run emits DISTILL lines: taste calls converted into rule candidates; append the\nones you agree with. When the model repeats a failure the files miss, that is a one-line\nban-list entry with a replacement, and it upgrades every future run. Workflow and the rule\nbar: [docs\u002FCUSTOMIZING.md](docs\u002FCUSTOMIZING.md) and [CONTRIBUTING.md](CONTRIBUTING.md).\n\n## FAQ\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>Can this really get Opus, Sonnet, GPT, or Gemini close to Fable 5's output?\u003C\u002Fb>\u003C\u002Fsummary>\n\nOn verifiable work, that is the design: iteration and explicit standards lift a weaker\nmodel far more than a stronger one, which is exactly the gap being closed. The ceiling is\nreal: a model judging its own tier plateaus below a stronger model's eye\n(\u003Ca href=\"#limitations\">Limitations\u003C\u002Fa> has the rest). The receipt:\n\u003Ca href=\"examples\u002Fthe-distillation-run.md\">the distillation run\u003C\u002Fa>.\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>Is this just a big system prompt?\u003C\u002Fb>\u003C\u002Fsummary>\n\nIn its single-response form, honestly, it is close: an engineered rubric plus one mandatory\nself-judge pass whose findings must be fixed before delivery. That judged pass is the\ndifference: a system prompt hopes, this one checks. On agentic surfaces it grows into a\nprocedure: part inventories, real evidence, parseable judges in fresh contexts, and an\nearned stop you can audit. \u003Ca href=\"PROMPT.md\">PROMPT.md\u003C\u002Fa> carries both forms.\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>What does a run cost?\u003C\u002Fb>\u003C\u002Fsummary>\n\nSee \u003Ca href=\"#token-cost-bounded-on-purpose\">Token cost\u003C\u002Fa> for the mode sizing. The one\nmeasured figure so far: the first gate pass this repo ran on its own README consumed about\n64k tokens and returned 12 findings, roughly 5k tokens per defect caught before launch\n(\u003Ca href=\"examples\u002Fself-run.md\">the full pass history\u003C\u002Fa>).\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>Do I need all 21 craft files?\u003C\u002Fb>\u003C\u002Fsummary>\n\nNo. Phase 0 routes each task to the 1-3 files that apply (a pricing page reads design,\nwriting, marketing). The rest stay on disk unread. Outside Claude Code, attach just the file\nmatching your domain.\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>Does it work with models other than Claude?\u003C\u002Fb>\u003C\u002Fsummary>\n\nThe skill packaging is Claude Code native, but \u003Ca href=\"PROMPT.md\">PROMPT.md\u003C\u002Fa> plus a\ncraft file runs the same procedure in Cursor, Windsurf, aider, or a raw API call to any\ncapable model. The standards are plain text and model-independent; only the packaging and\nthe tuning notes are Claude-specific.\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>What if I disagree with a rule?\u003C\u002Fb>\u003C\u002Fsummary>\n\nRules are files, and your fork is yours: edit or delete, keeping the three-part structure\n(numbered rules, ban list with replacements, verification checklist). If the rule is wrong\nin general, open an issue with the observed failure; that is exactly how these files grew.\n\u003C\u002Fdetails>\n\n\u003Cdetails>\n\u003Csummary>\u003Cb>How is this different from production-audit?\u003C\u002Fb>\u003C\u002Fsummary>\n\nSame author, the same earned stop condition, opposite direction: production-audit inspects\nan existing product until it stops finding defects; frontier manufactures new work so it\narrives already inspected. Run frontier to build, production-audit before you ship.\n\u003C\u002Fdetails>\n\n## License\n\nMIT: run it inside a company, fork the standards to your own house rules, ship products\nbuilt under it; the only obligation is keeping the license notice. If the gate earns its\nkeep, a star helps the next person find it.\n\n---\n\n\u003Cdiv align=\"center\">\n\n*The first time it fails something you were proud of, that is the skill working.*\n\n\u003C\u002Fdiv>\n","frontier 是一个面向大语言模型输出质量提升的轻量级方法框架，通过预定义的21条由Claude Fable 5自主撰写并审计的工艺标准，引导Claude（Opus\u002FSonnet）、GPT、Gemini等模型生成更接近前沿水平的交付物。它支持单次响应提质（single-response lift）或带验证门控（taste gate）的多轮收敛循环，核心特点是标准驱动、证据锚定的自动核查与迭代优化。适用于需要高确定性交付质量的场景，如关键代码生成、技术文档撰写、合规性内容产出等对可验证性要求较高的LLM应用。",2,"2026-07-08 04:30:15","CREATED_QUERY"]