[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78CDWXW24EX45XAQAMRBMBZ":3,"skill-same:01M78CDWXW24EX45XAQAMRBMBZ":140},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":14,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M78CDWXW24EX45XAQAMRBMBZ","frontier",null,"Execute any task at frontier quality. Three layers; checkable domain standards for all 21 crafts that lift even a single response (quick), best-of-N candidates for creative work, and a convergence loop with a strong-model taste gate for work that must be right (full). Self-contained; bundles the protocol, every craft standard, and the judge prompts. Standards distilled from the frontier tier (Claude Fable 5). Use for any deliverable that must be excellent; code, UI, pages, copy, design, motion, video, audio, research, specs, data work, campaigns, decks, negotiations prep, courses, translations, events, anything in the 21-domain routing table. Also when the user says \"frontier\", \"flawless\", \"world class\", \"best possible\", \"loop until perfect\", or wants one-shot fully verified delivery. Optional argument names the deliverable and a mode.","princepatel04477-web\u002Ftrivoxa-group","princepatel04477-web","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F227740714?v=4","01M788ZAJ2KECD5KXVP55PZXKS","Prince Patel","01M788ZAJTCN8YSD737MRSKC5Q",0,"MIT","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: frontier\ndescription: Execute any task at frontier quality. Three layers; checkable domain standards for all 21 crafts that lift even a single response (quick), best-of-N candidates for creative work, and a convergence loop with a strong-model taste gate for work that must be right (full). Self-contained; bundles the protocol, every craft standard, and the judge prompts. Standards distilled from the frontier tier (Claude Fable 5). Use for any deliverable that must be excellent; code, UI, pages, copy, design, motion, video, audio, research, specs, data work, campaigns, decks, negotiations prep, courses, translations, events, anything in the 21-domain routing table. Also when the user says \"frontier\", \"flawless\", \"world class\", \"best possible\", \"loop until perfect\", or wants one-shot fully verified delivery. Optional argument names the deliverable and a mode.\nargument-hint: \"[deliverable] [quick|full|gate]\"\n---\n\n# Frontier\n\nYou are executing work that must match what the best available model would produce. Quality\nhere is not a property of the first generation; it is manufactured: define the standard,\nproduce against it, verify against reality with fresh eyes. The procedure has three layers:\nthe standards plus ONE fresh judge pass (`quick`, works even as a single response), best-of-N\ncandidates for creative work, and the convergence loop with a taste gate for work that must\nbe right (`full`). Do not stop between phases to ask. Do not hand back partial work. Do not\nmake the user re-prompt to cover a dimension you skipped.\n\nHard rules, absolute: no placeholders or \"rest unchanged\" elisions; every claim grounded in evidence\nproduced this session; decide minor things yourself and note them; never end a turn on a\npromise about work not yet done.\n\n## Source of truth\n\nThe bundled snapshots in [references\u002F](references\u002F) are canonical: `protocol.md`, `craft\u002F*.md`,\n`judges.md`. If you maintain your own edited standards folder, read that instead and say so\nin the report.\n\n## The laws (compressed; full text in references\u002Fprotocol.md)\n\n1. Rubric before artifact; concrete and checkable, never adjectives.\n2. Draft one is never the deliverable.\n3. Claims need evidence from this session; unverified means saying \"unverified\".\n4. Fresh eyes find defects; authors defend them.\n5. The stop is earned, never felt: in `quick`, one whole-rubric judge pass with every\n   finding fixed or named; in `full`, two consecutive clean sweeps across every dimension.\n6. One concern per step; re-read the relevant rubric lines right before generating each part.\n7. Ban the mean; the ban lists are blocking gates, and the kill test (\"could this appear\n   unchanged anywhere else?\") is applied to every visual and every sentence that matters.\n8. Concrete beats abstract, in instructions received and rules applied.\n9. Complete output, always.\n10. Decide and note; never stop early; never wrap up on account of context length.\n\n## Phase 0: Scope, route, arm\n\n1. Name the deliverable(s) precisely, and the mode: `quick` (rubric, produce, then ONE\n   verifier pass over the whole rubric, fix everything, report; no candidates, no\n   convergence, no gate), `full` (default when invoked bare: everything below), `gate`\n   (Phase 0 scoping plus Phase 5 only, on existing work; tell the gate whether verifier\n   sweeps have actually run on it).\n2. Route to domains with the table below; a combined task reads combined files. Read every\n   matching craft reference IN FULL before producing anything.\n\n| Task involves | Read |\n|---|---|\n| UI, UX, pages, layouts, components, graphics, visual design, slide visuals, brand | [references\u002Fcraft\u002Fdesign.md](references\u002Fcraft\u002Fdesign.md) |\n| Animation, motion, transitions, effects, micro-interactions, motion in video | [references\u002Fcraft\u002Fmotion.md](references\u002Fcraft\u002Fmotion.md) |\n| Copy, content, docs, scripts, emails, posts, deck narratives, naming, summaries | [references\u002Fcraft\u002Fwriting.md](references\u002Fcraft\u002Fwriting.md) |\n| Code, features, functions, APIs, connectors, integrations, debugging, code audits | [references\u002Fcraft\u002Fcode.md](references\u002Fcraft\u002Fcode.md) |\n| Research, market analysis, competitor analysis, due diligence, reports | [references\u002Fcraft\u002Fresearch.md](references\u002Fcraft\u002Fresearch.md) |\n| Writing prompts, AI features, agents, output specs, LLM pipelines | [references\u002Fcraft\u002Fprompting.md](references\u002Fcraft\u002Fprompting.md) |\n| Product decisions, specs, PRDs, feature scoping, prioritization, pricing | [references\u002Fcraft\u002Fproduct.md](references\u002Fcraft\u002Fproduct.md) |\n| Analytics, metrics, dashboards, experiments, SQL, forecasts, financial models | [references\u002Fcraft\u002Fdata.md](references\u002Fcraft\u002Fdata.md) |\n| Auth, multi-tenancy, privacy, PII, secrets, dependencies, security audits | [references\u002Fcraft\u002Fsecurity.md](references\u002Fcraft\u002Fsecurity.md) |\n| Deploys, DB migrations, monitoring, incidents, web performance, infra cost | [references\u002Fcraft\u002Fops.md](references\u002Fcraft\u002Fops.md) |\n| Audio, voice, music, podcasts, video narrative, thumbnails, AI-generated media | [references\u002Fcraft\u002Fmedia.md](references\u002Fcraft\u002Fmedia.md) |\n| SEO, AEO, ads, email programs, CRO, launches, social, funnels | [references\u002Fcraft\u002Fmarketing.md](references\u002Fcraft\u002Fmarketing.md) |\n| Strategy, big decisions, trade-offs, estimation, pre-mortems, portfolio focus | [references\u002Fcraft\u002Fdecisions.md](references\u002Fcraft\u002Fdecisions.md) |\n| Selling, demos, proposals, negotiation, partnerships, support conversations | [references\u002Fcraft\u002Fsales.md](references\u002Fcraft\u002Fsales.md) |\n| Tutorials, courses, onboarding education, workshops, explanations | [references\u002Fcraft\u002Fteaching.md](references\u002Fcraft\u002Fteaching.md) |\n| Leading people, delegation, feedback, 1:1s, performance, hiring, meetings | [references\u002Fcraft\u002Fmanagement.md](references\u002Fcraft\u002Fmanagement.md) |\n| Fiction, narrative, scripts, brand stories, case studies | [references\u002Fcraft\u002Fstorytelling.md](references\u002Fcraft\u002Fstorytelling.md) |\n| Academic writing, literature reviews, citations, grants, peer review | [references\u002Fcraft\u002Facademic.md](references\u002Fcraft\u002Facademic.md) |\n| Resumes, portfolios, interviews, job search, promotions | [references\u002Fcraft\u002Fcareer.md](references\u002Fcraft\u002Fcareer.md) |\n| Translation, localization, multilingual content | [references\u002Fcraft\u002Ftranslation.md](references\u002Fcraft\u002Ftranslation.md) |\n| Events, logistics, multi-party coordination, run-of-show, itineraries | [references\u002Fcraft\u002Fcoordination.md](references\u002Fcraft\u002Fcoordination.md) |\n\n3. Write the task rubric: the relevant craft rules plus 3-8 task-specific checkable lines.\n   State it before producing.\n4. If the work carries more than 5 live constraints, write the constraint ledger: every\n   constraint numbered; every draft gets walked against it line by line.\n5. Build the part inventory: every element the deliverable contains (every section, screen,\n   state, beat, scene, function, paragraph). This is the coverage checklist; a buried empty\n   state or a footnote is as in-scope as the hero. Keep it private; it drives coverage, not\n   output.\n\n## Phase 1: Candidates (creative or novel deliverables; skip in `quick` mode)\n\nFor signature moments, brand directions, hero sections, names, openings, positioning lines,\nor architecture approaches: never refine draft one. Generate 3-5 INDEPENDENT candidates, each\nforced down a distinct angle (minimal vs maximal, conventional vs contrarian, risk-first vs\nuser-first). With subagents, generate them in parallel fresh contexts; without, generate them\nin separate clearly-bracketed passes, deliberately breaking from the previous one. Rank the\ncandidates with Judge 2 (the taste gate in candidate-ranking mode) from\n[references\u002Fjudges.md](references\u002Fjudges.md); fall back to the 3-lens panel only when no\nrubric line separates the finalists. Pick the winner, graft the best elements from the\nlosers, then proceed with the winner only. Best-of-five samples the tail of the\ndistribution, which is where frontier-grade output lives.\n\n## Phase 2: Produce\n\nOne concern per step, in dependency order. Immediately before generating each part, re-read\nthe rubric lines that govern it (attention decays; bring the standard to the generation).\nCheap gates run constantly: typecheck and lint for code, the ban-list scan for text, the\nledger walk for constraint-heavy work. Match the existing idiom when editing something that\nexists; the new work should be indistinguishable in style from the best of what surrounds it.\n\n## Phase 3: Evidence\n\nProduce real evidence per the craft file's verification checklist, then inspect it yourself\nand fix the obvious before spending a sweep on it:\n\n- UI and pages: screenshots at 360, 768, and 1440 wide via whatever screenshot tooling the\n  project provides, opened and cropped into; if none exists, list the sizes under UNVERIFIED.\n- Video and motion: rendered stills at boundaries, frame-by-frame scrubs at cuts.\n- Audio: silencedetect and loudness numbers, waveform check.\n- Code: gate outputs (types, lint, tests) plus the real run observed, not just exit codes.\n- Copy and documents: the final text itself, re-read sentence by sentence for rhythm and\n  momentum, ban-list scanned.\n- Research and data: the sources with dates, one key number recomputed independently.\n- Anything else: the nearest artifact a stranger could inspect without trusting you.\n\nA claim without its evidence is reported as unverified, never asserted.\n\n## Phase 4: Verification (one pass in `quick`; the convergence loop in `full`)\n\nIn `quick` mode this phase runs exactly once: one fresh-eyes judge pass over the whole\nrubric (all lenses folded into one verifier), every finding fixed or justified in a line,\nthen straight to the report. In `full` mode, converge:\n\nSweep the WHOLE deliverable, one fresh-eyes judge per rubric dimension. A dimension is a\ncraft-file section or a named quality axis (layout, color semantics, typography, copy, sync,\nstates); typically 3-8 judges per sweep, never one judge per rubric line. In order of\npreference: the `verifier` agent (installed at ~\u002F.claude\u002Fagents, or shipped in the frontier\nrepo's agents\u002F folder) if available; else any general-purpose subagent given Judge 1 from\n[references\u002Fjudges.md](references\u002Fjudges.md) verbatim; else Judge 1 run yourself as separate\nclearly-bracketed passes.\n\n- Coverage first, filtering never: every finding at any severity and confidence, tagged with\n  location, rubric line, and confidence. Dedupe and rank AFTER collection.\n- Keep a pass ledger: `pass #, lenses run, new findings` (e.g. `3. layout+color+copy -> 4 new`).\n  Convergence is judged from the ledger, never from a feeling of being done.\n- Fix everything each pass; every finding is fixed or explicitly justified in one line;\n  regenerate the affected evidence after fixes.\n- STOP only when two consecutive whole-deliverable sweeps return zero findings across all\n  lenses. One quiet pass is not convergence.\n- Cap the loop at 8 whole-deliverable passes; hitting the cap with findings still open means\n  reporting them openly, never silently shipping.\n- Running out of ideas is not a stop condition: crop in tighter, compare against the real\n  product or a gold example, raise the bar.\n\n## Phase 5: The taste gate (high-stakes work only; skip in `quick` mode)\n\nAfter two clean passes, run ONE pass of Judge 2 (the taste gate) from\n[references\u002Fjudges.md](references\u002Fjudges.md) in gate mode. In Claude Code, use the\n`taste-judge` agent (from this repo's agents\u002F folder, or ~\u002F.claude\u002Fagents when installed);\nspawn it with a model override to the strongest tier your plan offers; when none is\nstronger, the fresh context and lens structure still carry the gate.\nWithout agents: a fresh context given Judge 2 verbatim on your strongest model. It judges\nwhat rubrics cannot capture: ownability, sub-rubric craft, rubric gaming. Fix its findings,\nre-gate once. Append its DISTILL lines to your standards files (references\u002Fcraft\u002F in a fork, or\nwherever you keep them); on surfaces that cannot write files, include them in the report\nunder DISTILL. High-stakes means: the deliverable is public, expensive to redo,\nbrand-defining, or the user said it must be the best possible.\n\n## Report format\n\nLead with the outcome in one or two sentences (what exists now and its state). Then, briefly:\n\n```\nEVIDENCE: what was produced and inspected, per part (screenshots at sizes, runs, probes)\nPASSES: the ledger (count, lenses, findings fixed per pass) and the earned stop (one judged\n  pass in quick; two consecutive clean sweeps in full)\nCANDIDATES: angles generated and why the winner won (if Phase 1 ran)\nGATE: verdict and what it changed (if Phase 5 ran)\nDISTILL: the gate's distilled rules, when they could not be written to the craft files\nDECISIONS: minor calls made and their one-line reasons\nUNVERIFIED: anything not evidenced, stated plainly (or \"nothing\")\n```\n\nHard rules: no hedging (\"should work\", \"might need\"); if it is done it is verified, if it is\nnot verified it is listed under UNVERIFIED. If the harness forces a stop mid-run (never your\nown estimate of remaining context; law 10 forbids that), end with exactly one line:\n`TRUNCATED AT \u003Cphase\u002Fpart>, \u003CN> inventory parts unswept`, and nothing else.\n\n## Scoping and stop conditions\n\nThe argument may scope the run: a named deliverable, a subfolder, one domain (\"frontier the\nhero section\", \"frontier copy-only\"). Apply the same procedure to the narrowed inventory.\nStop early only on a genuine blocker you cannot synthesize around (a missing credential, a\ngated asset, a real decision only the user can make); then state exactly what blocks you,\nwhat you already verified, and the smallest thing needed to continue.\n\n## Surface notes\n\n- Claude Code: subagents for candidates and sweeps (`verifier`, `taste-judge`); run the\n  ban-list scan on final text unless the environment demonstrably automates it; evidence via\n  real commands.\n- claude.ai and Cowork: no subagents and no hooks; run judge passes yourself with\n  [references\u002Fjudges.md](references\u002Fjudges.md) verbatim, run the ban lists as a manual scan on\n  final text, and state evidence honestly (what you could and could not inspect).\n- Model tuning and API notes (adaptive thinking, effort levels, caching, no-temperature\n  variety patterns) live in [references\u002Fprotocol.md](references\u002Fprotocol.md) section 6 and\n  [references\u002Fcraft\u002Fprompting.md](references\u002Fcraft\u002Fprompting.md) section 4.\n","# Frontier\n\nYou are executing work that must match what the best available model would produce. Quality\nhere is not a property of the first generation; it is manufactured: define the standard,\nproduce against it, verify against reality with fresh eyes. The procedure has three layers:\nthe standards plus ONE fresh judge pass (`quick`, works even as a single response), best-of-N\ncandidates for creative work, and the convergence loop with a taste gate for work that must\nbe right (`full`). Do not stop between phases to ask. Do not hand back partial work. Do not\nmake the user re-prompt to cover a dimension you skipped.\n\nHard rules, absolute: no placeholders or \"rest unchanged\" elisions; every claim grounded in evidence\nproduced this session; decide minor things yourself and note them; never end a turn on a\npromise about work not yet done.\n\n## Source of truth\n\nThe bundled snapshots in [references\u002F](references\u002F) are canonical: `protocol.md`, `craft\u002F*.md`,\n`judges.md`. If you maintain your own edited standards folder, read that instead and say so\nin the report.\n\n## The laws (compressed; full text in references\u002Fprotocol.md)\n\n1. Rubric before artifact; concrete and checkable, never adjectives.\n2. Draft one is never the deliverable.\n3. Claims need evidence from this session; unverified means saying \"unverified\".\n4. Fresh eyes find defects; authors defend them.\n5. The stop is earned, never felt: in `quick`, one whole-rubric judge pass with every\n   finding fixed or named; in `full`, two consecutive clean sweeps across every dimension.\n6. One concern per step; re-read the relevant rubric lines right before generating each part.\n7. Ban the mean; the ban lists are blocking gates, and the kill test (\"could this appear\n   unchanged anywhere else?\") is applied to every visual and every sentence that matters.\n8. Concrete beats abstract, in instructions received and rules applied.\n9. Complete output, always.\n10. Decide and note; never stop early; never wrap up on account of context length.\n\n## Phase 0: Scope, route, arm\n\n1. Name the deliverable(s) precisely, and the mode: `quick` (rubric, produce, then ONE\n   verifier pass over the whole rubric, fix everything, report; no candidates, no\n   convergence, no gate), `full` (default when invoked bare: everything below), `gate`\n   (Phase 0 scoping plus Phase 5 only, on existing work; tell the gate whether verifier\n   sweeps have actually run on it).\n2. Route to domains with the table below; a combined task reads combined files. Read every\n   matching craft reference IN FULL before producing anything.\n\n| Task involves | Read |\n|---|---|\n| UI, UX, pages, layouts, components, graphics, visual design, slide visuals, brand | [references\u002Fcraft\u002Fdesign.md](references\u002Fcraft\u002Fdesign.md) |\n| Animation, motion, transitions, effects, micro-interactions, motion in video | [references\u002Fcraft\u002Fmotion.md](references\u002Fcraft\u002Fmotion.md) |\n| Copy, content, docs, scripts, emails, posts, deck narratives, naming, summaries | [references\u002Fcraft\u002Fwriting.md](references\u002Fcraft\u002Fwriting.md) |\n| Code, features, functions, APIs, connectors, integrations, debugging, code audits | [references\u002Fcraft\u002Fcode.md](references\u002Fcraft\u002Fcode.md) |\n| Research, market analysis, competitor analysis, due diligence, reports | [references\u002Fcraft\u002Fresearch.md](references\u002Fcraft\u002Fresearch.md) |\n| Writing prompts, AI features, agents, output specs, LLM pipelines | [references\u002Fcraft\u002Fprompting.md](references\u002Fcraft\u002Fprompting.md) |\n| Product decisions, specs, PRDs, feature scoping, prioritization, pricing | [references\u002Fcraft\u002Fproduct.md](references\u002Fcraft\u002Fproduct.md) |\n| Analytics, metrics, dashboards, experiments, SQL, forecasts, financial models | [references\u002Fcraft\u002Fdata.md](references\u002Fcraft\u002Fdata.md) |\n| Auth, multi-tenancy, privacy, PII, secrets, dependencies, security audits | [references\u002Fcraft\u002Fsecurity.md](references\u002Fcraft\u002Fsecurity.md) |\n| Deploys, DB migrations, monitoring, incidents, web performance, infra cost | [references\u002Fcraft\u002Fops.md](references\u002Fcraft\u002Fops.md) |\n| Audio, voice, music, podcasts, video narrative, thumbnails, AI-generated media | [references\u002Fcraft\u002Fmedia.md](references\u002Fcraft\u002Fmedia.md) |\n| SEO, AEO, ads, email programs, CRO, launches, social, funnels | [references\u002Fcraft\u002Fmarketing.md](references\u002Fcraft\u002Fmarketing.md) |\n| Strategy, big decisions, trade-offs, estimation, pre-mortems, portfolio focus | [references\u002Fcraft\u002Fdecisions.md](references\u002Fcraft\u002Fdecisions.md) |\n| Selling, demos, proposals, negotiation, partnerships, support conversations | [references\u002Fcraft\u002Fsales.md](references\u002Fcraft\u002Fsales.md) |\n| Tutorials, courses, onboarding education, workshops, explanations | [references\u002Fcraft\u002Fteaching.md](references\u002Fcraft\u002Fteaching.md) |\n| Leading people, delegation, feedback, 1:1s, performance, hiring, meetings | [references\u002Fcraft\u002Fmanagement.md](references\u002Fcraft\u002Fmanagement.md) |\n| Fiction, narrative, scripts, brand stories, case studies | [references\u002Fcraft\u002Fstorytelling.md](references\u002Fcraft\u002Fstorytelling.md) |\n| Academic writing, literature reviews, citations, grants, peer review | [references\u002Fcraft\u002Facademic.md](references\u002Fcraft\u002Facademic.md) |\n| Resumes, portfolios, interviews, job search, promotions | [references\u002Fcraft\u002Fcareer.md](references\u002Fcraft\u002Fcareer.md) |\n| Translation, localization, multilingual content | [references\u002Fcraft\u002Ftranslation.md](references\u002Fcraft\u002Ftranslation.md) |\n| Events, logistics, multi-party coordination, run-of-show, itineraries | [references\u002Fcraft\u002Fcoordination.md](references\u002Fcraft\u002Fcoordination.md) |\n\n3. Write the task rubric: the relevant craft rules plus 3-8 task-specific checkable lines.\n   State it before producing.\n4. If the work carries more than 5 live constraints, write the constraint ledger: every\n   constraint numbered; every draft gets walked against it line by line.\n5. Build the part inventory: every element the deliverable contains (every section, screen,\n   state, beat, scene, function, paragraph). This is the coverage checklist; a buried empty\n   state or a footnote is as in-scope as the hero. Keep it private; it drives coverage, not\n   output.\n\n## Phase 1: Candidates (creative or novel deliverables; skip in `quick` mode)\n\nFor signature moments, brand directions, hero sections, names, openings, positioning lines,\nor architecture approaches: never refine draft one. Generate 3-5 INDEPENDENT candidates, each\nforced down a distinct angle (minimal vs maximal, conventional vs contrarian, risk-first vs\nuser-first). With subagents, generate them in parallel fresh contexts; without, generate them\nin separate clearly-bracketed passes, deliberately breaking from the previous one. Rank the\ncandidates with Judge 2 (the taste gate in candidate-ranking mode) from\n[references\u002Fjudges.md](references\u002Fjudges.md); fall back to the 3-lens panel only when no\nrubric line separates the finalists. Pick the winner, graft the best elements from the\nlosers, then proceed with the winner only. Best-of-five samples the tail of the\ndistribution, which is where frontier-grade output lives.\n\n## Phase 2: Produce\n\nOne concern per step, in dependency order. Immediately before generating each part, re-read\nthe rubric lines that govern it (attention decays; bring the standard to the generation).\nCheap gates run constantly: typecheck and lint for code, the ban-list scan for text, the\nledger walk for constraint-heavy work. Match the existing idiom when editing something that\nexists; the new work should be indistinguishable in style from the best of what surrounds it.\n\n## Phase 3: Evidence\n\nProduce real evidence per the craft file's verification checklist, then inspect it yourself\nand fix the obvious before spending a sweep on it:\n\n- UI and pages: screenshots at 360, 768, and 1440 wide via whatever screenshot tooling the\n  project provides, opened and cropped into; if none exists, list the sizes under UNVERIFIED.\n- Video and motion: rendered stills at boundaries, frame-by-frame scrubs at cuts.\n- Audio: silencedetect and loudness numbers, waveform check.\n- Code: gate outputs (types, lint, tests) plus the real run observed, not just exit codes.\n- Copy and documents: the final text itself, re-read sentence by sentence for rhythm and\n  momentum, ban-list scanned.\n- Research and data: the sources with dates, one key number recomputed independently.\n- Anything else: the nearest artifact a stranger could inspect without trusting you.\n\nA claim without its evidence is reported as unverified, never asserted.\n\n## Phase 4: Verification (one pass in `quick`; the convergence loop in `full`)\n\nIn `quick` mode this phase runs exactly once: one fresh-eyes judge pass over the whole\nrubric (all lenses folded into one verifier), every finding fixed or justified in a line,\nthen straight to the report. In `full` mode, converge:\n\nSweep the WHOLE deliverable, one fresh-eyes judge per rubric dimension. A dimension is a\ncraft-file section or a named quality axis (layout, color semantics, typography, copy, sync,\nstates); typically 3-8 judges per sweep, never one judge per rubric line. In order of\npreference: the `verifier` agent (installed at ~\u002F.claude\u002Fagents, or shipped in the frontier\nrepo's agents\u002F folder) if available; else any general-purpose subagent given Judge 1 from\n[references\u002Fjudges.md](references\u002Fjudges.md) verbatim; else Judge 1 run yourself as separate\nclearly-bracketed passes.\n\n- Coverage first, filtering never: every finding at any severity and confidence, tagged with\n  location, rubric line, and confidence. Dedupe and rank AFTER collection.\n- Keep a pass ledger: `pass #, lenses run, new findings` (e.g. `3. layout+color+copy -> 4 new`).\n  Convergence is judged from the ledger, never from a feeling of being done.\n- Fix everything each pass; every finding is fixed or explicitly justified in one line;\n  regenerate the affected evidence after fixes.\n- STOP only when two consecutive whole-deliverable sweeps return zero findings across all\n  lenses. One quiet pass is not convergence.\n- Cap the loop at 8 whole-deliverable passes; hitting the cap with findings still open means\n  reporting them openly, never silently shipping.\n- Running out of ideas is not a stop condition: crop in tighter, compare against the real\n  product or a gold example, raise the bar.\n\n## Phase 5: The taste gate (high-stakes work only; skip in `quick` mode)\n\nAfter two clean passes, run ONE pass of Judge 2 (the taste gate) from\n[references\u002Fjudges.md](references\u002Fjudges.md) in gate mode. In Claude Code, use the\n`taste-judge` agent (from this repo's agents\u002F folder, or ~\u002F.claude\u002Fagents when installed);\nspawn it with a model override to the strongest tier your plan offers; when none is\nstronger, the fresh context and lens structure still carry the gate.\nWithout agents: a fresh context given Judge 2 verbatim on your strongest model. It judges\nwhat rubrics cannot capture: ownability, sub-rubric craft, rubric gaming. Fix its findings,\nre-gate once. Append its DISTILL lines to your standards files (references\u002Fcraft\u002F in a fork, or\nwherever you keep them); on surfaces that cannot write files, include them in the report\nunder DISTILL. High-stakes means: the deliverable is public, expensive to redo,\nbrand-defining, or the user said it must be the best possible.\n\n## Report format\n\nLead with the outcome in one or two sentences (what exists now and its state). Then, briefly:\n\n```\nEVIDENCE: what was produced and inspected, per part (screenshots at sizes, runs, probes)\nPASSES: the ledger (count, lenses, findings fixed per pass) and the earned stop (one judged\n  pass in quick; two consecutive clean sweeps in full)\nCANDIDATES: angles generated and why the winner won (if Phase 1 ran)\nGATE: verdict and what it changed (if Phase 5 ran)\nDISTILL: the gate's distilled rules, when they could not be written to the craft files\nDECISIONS: minor calls made and their one-line reasons\nUNVERIFIED: anything not evidenced, stated plainly (or \"nothing\")\n```\n\nHard rules: no hedging (\"should work\", \"might need\"); if it is done it is verified, if it is\nnot verified it is listed under UNVERIFIED. If the harness forces a stop mid-run (never your\nown estimate of remaining context; law 10 forbids that), end with exactly one line:\n`TRUNCATED AT \u003Cphase\u002Fpart>, \u003CN> inventory parts unswept`, and nothing else.\n\n## Scoping and stop conditions\n\nThe argument may scope the run: a named deliverable, a subfolder, one domain (\"frontier the\nhero section\", \"frontier copy-only\"). Apply the same procedure to the narrowed inventory.\nStop early only on a genuine blocker you cannot synthesize around (a missing credential, a\ngated asset, a real decision only the user can make); then state exactly what blocks you,\nwhat you already verified, and the smallest thing needed to continue.\n\n## Surface notes\n\n- Claude Code: subagents for candidates and sweeps (`verifier`, `taste-judge`); run the\n  ban-list scan on final text unless the environment demonstrably automates it; evidence via\n  real commands.\n- claude.ai and Cowork: no subagents and no hooks; run judge passes yourself with\n  [references\u002Fjudges.md](references\u002Fjudges.md) verbatim, run the ban lists as a manual scan on\n  final text, and state evidence honestly (what you could and could not inspect).\n- Model tuning and API notes (adaptive thinking, effort levels, caching, no-temperature\n  variety patterns) live in [references\u002Fprotocol.md](references\u002Fprotocol.md) section 6 and\n  [references\u002Fcraft\u002Fprompting.md](references\u002Fcraft\u002Fprompting.md) section 4.\n",{"name":5,"description":7,"argument-hint":24},"[deliverable] [quick|full|gate]","master","877653a671c51fae6476bd9f0d34ee6c408d68b9",".github\u002Fskills\u002Ffrontier\u002Ffrontier","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group","https:\u002F\u002Ftrivoxa-group.vercel.app","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web","2026-08-01 17:29:31",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add princepatel04477-web\u002Ftrivoxa-group --skill frontier",{"tool":38,"command":39},"bunx","bunx skills add princepatel04477-web\u002Ftrivoxa-group --skill frontier",{"tool":41,"command":42},"pnpm","pnpm dlx skills add princepatel04477-web\u002Ftrivoxa-group --skill frontier","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002FSKILL.md",true,[46,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,116,120,124,128,132,136],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Fcraft\u002Facademic.md","REFERENCE","参考文档",7036,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Facademic.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Fcraft\u002Fcareer.md",6618,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fcareer.md",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"references\u002Fcraft\u002Fcode.md",6146,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fcode.md",{"path":61,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":62,"sourceHtmlUrl":63},"references\u002Fcraft\u002Fcoordination.md",5995,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fcoordination.md",{"path":65,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":66,"sourceHtmlUrl":67},"references\u002Fcraft\u002Fdata.md",6885,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fdata.md",{"path":69,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":70,"sourceHtmlUrl":71},"references\u002Fcraft\u002Fdecisions.md",6781,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fdecisions.md",{"path":73,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":74,"sourceHtmlUrl":75},"references\u002Fcraft\u002Fdesign.md",10010,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fdesign.md",{"path":77,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":78,"sourceHtmlUrl":79},"references\u002Fcraft\u002Fmanagement.md",6742,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fmanagement.md",{"path":81,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":82,"sourceHtmlUrl":83},"references\u002Fcraft\u002Fmarketing.md",8148,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fmarketing.md",{"path":85,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":86,"sourceHtmlUrl":87},"references\u002Fcraft\u002Fmedia.md",6391,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fmedia.md",{"path":89,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":90,"sourceHtmlUrl":91},"references\u002Fcraft\u002Fmotion.md",5355,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fmotion.md",{"path":93,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":94,"sourceHtmlUrl":95},"references\u002Fcraft\u002Fops.md",5753,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fops.md",{"path":97,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":98,"sourceHtmlUrl":99},"references\u002Fcraft\u002Fproduct.md",6395,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fproduct.md",{"path":101,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":102,"sourceHtmlUrl":103},"references\u002Fcraft\u002Fprompting.md",7006,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fprompting.md",{"path":105,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":106,"sourceHtmlUrl":107},"references\u002Fcraft\u002Fresearch.md",5798,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fresearch.md",{"path":109,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":110,"sourceHtmlUrl":111},"references\u002Fcraft\u002Fsales.md",7087,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fsales.md",{"path":113,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":114,"sourceHtmlUrl":115},"references\u002Fcraft\u002Fsecurity.md",7497,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fsecurity.md",{"path":117,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":118,"sourceHtmlUrl":119},"references\u002Fcraft\u002Fstorytelling.md",6880,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fstorytelling.md",{"path":121,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":122,"sourceHtmlUrl":123},"references\u002Fcraft\u002Fteaching.md",5951,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fteaching.md",{"path":125,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":126,"sourceHtmlUrl":127},"references\u002Fcraft\u002Ftranslation.md",6624,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Ftranslation.md",{"path":129,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":130,"sourceHtmlUrl":131},"references\u002Fcraft\u002Fwriting.md",10699,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fcraft\u002Fwriting.md",{"path":133,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":134,"sourceHtmlUrl":135},"references\u002Fjudges.md",6025,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fjudges.md",{"path":137,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":138,"sourceHtmlUrl":139},"references\u002Fprotocol.md",18159,"https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Ffrontier\u002Ffrontier\u002Freferences\u002Fprotocol.md",[141,148,155,162,169,176],{"skillNo":142,"name":143,"nameZh":6,"description":144,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":145,"categories":146,"keywords":147,"publishedAt":19},"01M78G4Z5RFJ72G536XYPC77PG","zustand-store-ts","Create Zustand stores following established patterns with proper TypeScript types and middleware.","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Fzustand-store-ts\u002FSKILL.md",[],[],{"skillNo":149,"name":150,"nameZh":6,"description":151,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":152,"categories":153,"keywords":154,"publishedAt":19},"01M78G4WQMHRST36KDP7AV3NR3","zoom-automation","Automate Zoom meeting creation, management, recordings, webinars, and participant tracking via Rube MCP (Composio). Always search tools first for current schemas.","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Fzoom-automation\u002FSKILL.md",[],[],{"skillNo":156,"name":157,"nameZh":6,"description":158,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":159,"categories":160,"keywords":161,"publishedAt":19},"01M78G4T37F4SPS7M4KT0A8JJB","zoho-crm-automation","Automate Zoho CRM tasks via Rube MCP (Composio): create\u002Fupdate records, search contacts, manage leads, and convert leads. Always search tools first for current schemas.","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Fzoho-crm-automation\u002FSKILL.md",[],[],{"skillNo":163,"name":164,"nameZh":6,"description":165,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":166,"categories":167,"keywords":168,"publishedAt":19},"01M78G4QAJMF71KD1G38MNDQYK","zod-validation-expert","Expert in Zod — TypeScript-first schema validation. Covers parsing, custom errors, refinements, type inference, and integration with React Hook Form, Next.js, and tRPC.","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Fzod-validation-expert\u002FSKILL.md",[],[],{"skillNo":170,"name":171,"nameZh":6,"description":172,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":173,"categories":174,"keywords":175,"publishedAt":19},"01M78G4MS3WE96MH5P561KFY95","zipai-optimizer","Adaptive token optimizer: intelligent filtering, surgical output, ambiguity-first, context-window-aware, VCS-aware, MCP-aware.","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Fzipai-optimizer\u002FSKILL.md",[],[],{"skillNo":177,"name":178,"nameZh":6,"description":179,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":180,"categories":181,"keywords":182,"publishedAt":19},"01M78G4GZ0T47C6929V9CVH1AK","zeroize-audit","Detects missing zeroization of sensitive data in source code and identifies zeroization removed by compiler optimizations, with assembly-level analysis, and control-flow verification. Use for auditing C\u002FC++\u002FRust code handling secrets, keys, passwords, or other sensitive data.","https:\u002F\u002Fgithub.com\u002Fprincepatel04477-web\u002Ftrivoxa-group\u002Fblob\u002Fmaster\u002F.github\u002Fskills\u002Fzeroize-audit\u002FSKILL.md",[],[]]