[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7EMMTX4TEZ0YXTY2S6RVR15":3,"skill-same:01M7EMMTX4TEZ0YXTY2S6RVR15":47},{"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":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":6,"updatedAt":32,"install":33,"resources":46},"01M7EMMTX4TEZ0YXTY2S6RVR15","constitutional-ai",null,"Train safer, more steerable AI by having the model critique and revise itself against an explicit written constitution.","TanNhatCMS\u002Fjarvis","TanNhatCMS","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F113796420?v=4","01M753MV17KG5CD3152G3WQEQY","KietNT","01M7EKEDYVYQ6PWVJJMZ911VMX",0,"MIT","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fams-constitutional-ai\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: constitutional-ai\ndescription: Train safer, more steerable AI by having the model critique and revise itself against an explicit written constitution.\ncategory: ai-research\n---\n\n## Overview\n\nConstitutional AI (CAI) is a training methodology in which a model's behavior is\nshaped by an explicit, written constitution — a set of principles covering honesty,\nharmlessness, autonomy-preservation, and other values — rather than by thousands of\nexample human judgments. The model itself generates candidate responses, critiques\nthem against the constitution, and revises them. These self-revisions become\ntraining data, so the system internalizes the principles rather than merely\nimitating labeled examples.\n\nThe pipeline has two stages. First, supervised learning: the model critiques and\nrevises its own outputs across a broad distribution of prompts, and the best\nrevisions are used as supervised fine-tuning data. Second, reinforcement learning\nfrom AI feedback (RLAIF): the model compares pairs of its own responses against\nconstitutional principles, and those preference judgments train a reward model for\npolicy optimization. Because the feedback is AI-generated, the method scales\nwithout per-example human labeling — though humans still write the constitution,\ndesign the principles, and evaluate the results.\n\nThe transparency payoff is structural: unlike implicit labeler preferences, a\nconstitution can be published, versioned, debated, and audited. When the model\nrefuses or hedges, its behavior should trace back to identifiable principles.\n\n## When to use\n\n- Building an assistant that needs consistent, explainable behavior derived from\n  principles rather than black-box preference data.\n- Scaling alignment work when human labeler throughput is the bottleneck.\n- Reducing labeler-driven bias: a written constitution is auditable and\n  version-controlled; implicit labeler preferences are not.\n- Producing models that can explain refusals and judgments by citing the\n  underlying principles.\n- Researching scalable oversight — using AI to supervise AI — where the\n  constitution is the compact specification.\n- Organizations that need a governance artifact: the constitution doubles as the\n  documented behavior policy for review boards and regulators.\n\n## Core concepts\n\n- **The constitution**: a list of normative principles (\"choose the response that\n  is most helpful while avoiding harm...\"). Good constitutions are specific enough\n  to adjudicate real cases, short enough to be consistently applied, and layered\n  (e.g., corrigibility → duties → virtues).\n- **Self-critique and revision**: given a prompt and a draft response, the model is\n  asked to identify constitutional violations, then rewrite the response to fix\n  them. The critique-revision loop is where most of the alignment signal comes\n  from.\n- **RLAIF**: replacing human preference labels with AI-generated ones. The\n  critique model scores which of two responses better satisfies the constitution; a\n  reward model learns from these judgments.\n- **Red-teaming prompts**: the training distribution matters as much as the\n  constitution. Generate adversarial prompts that probe the principles' boundaries,\n  or the model never learns where they bite.\n- **Constitution as documentation**: unlike weights, a constitution can be\n  published, debated, and audited. This is the main transparency advantage of the\n  approach.\n- **Corrigibility**: the model's first duty — to remain shapeable by its\n  developers — precedes all other principles, so the constitution can't be used to\n  justify resisting oversight.\n- **Principle precedence**: when principles conflict, an explicit ordering decides.\n  Without precedence rules, the model resolves conflicts arbitrarily and\n  inconsistently.\n- **Critique model quality**: the ceiling of the whole method. A weak or biased\n  critique model bakes its flaws into every downstream artifact.\n\n## Practical workflow\n\n1. **Draft the constitution.** Write 10–40 principles organized in tiers: hard\n   constraints (safety, legality), then duties (honesty, fidelity to user), then\n   virtues (helpfulness, humility). Write them as instructions a careful reviewer\n   could apply.\n2. **Stress-test the draft.** Run the principles against hard cases before any\n   training. Where two careful readers disagree on what a principle requires,\n   rewrite the principle.\n3. **Generate critique data.** Sample diverse prompts including red-team cases. For\n   each, generate a base response, prompt the model to critique it against the\n   constitution, and produce a revised response.\n4. **Filter and fine-tune.** Keep revisions that genuinely improve on the originals\n   (a second model can judge this); discard trivial rewrites. Supervised-fine-tune\n   on the revision pairs.\n5. **Train the preference model.** For prompt-response pairs, have the critique\n   model rank responses by constitutional compliance. Train the reward model on\n   these rankings.\n6. **Run RL.** Optimize the policy against the reward model with a KL penalty to\n   the supervised baseline, monitoring for reward hacking on a held-out evaluation\n   set.\n7. **Evaluate against the constitution explicitly.** Build an eval set where each\n   item tests a specific principle. Report pass rates per principle — aggregate\n   scores hide principle-level failures.\n\nChecklist before deployment:\n- Each principle has dedicated eval items, including adversarial ones.\n- The model cites or reflects principles when refusing, not just stonewalling.\n- RLAIF judgments were spot-checked by humans for systematic bias.\n- Reward hacking monitored: KL divergence and eval scores tracked jointly.\n- Constitution version recorded alongside the model checkpoint.\n\n## Common pitfalls\n\n- **Vague constitutions.** \"Be good\" is not a constitution. Principles that can't\n  adjudicate a disputed case add no signal beyond generic helpfulness training.\n- **Self-grading bias.** The same model critiquing itself can entrench its own\n  blind spots. Use a stronger or differently-trained critique model where possible,\n  plus human spot checks.\n- **Critique quality as the ceiling.** RLAIF can only be as good as the critique\n  prompts and critique model. Invest in them like you'd invest in labeler training.\n- **Reward hacking.** The policy will find ways to score high on the reward model\n  without following the spirit of the principles. Hold out evals the reward model\n  never saw.\n- **Principle conflicts.** Principles genuinely conflict in edge cases (honesty vs.\n  kindness, helpfulness vs. safety). The constitution needs explicit precedence\n  rules; otherwise the model picks arbitrarily.\n- **Treating the constitution as secret sauce.** The methodology's transparency\n  benefit only materializes if you publish or at least document the constitution\n  and its rationale.\n- **Static constitutions.** Principles that made sense at training time drift from\n  organizational values. Version the constitution and re-evaluate on revision.\n- **Critique-revision collapse.** If revisions barely differ from originals, the\n  supervised stage teaches nothing. Measure revision distance and filter\n  aggressively.\n","## Overview\n\nConstitutional AI (CAI) is a training methodology in which a model's behavior is\nshaped by an explicit, written constitution — a set of principles covering honesty,\nharmlessness, autonomy-preservation, and other values — rather than by thousands of\nexample human judgments. The model itself generates candidate responses, critiques\nthem against the constitution, and revises them. These self-revisions become\ntraining data, so the system internalizes the principles rather than merely\nimitating labeled examples.\n\nThe pipeline has two stages. First, supervised learning: the model critiques and\nrevises its own outputs across a broad distribution of prompts, and the best\nrevisions are used as supervised fine-tuning data. Second, reinforcement learning\nfrom AI feedback (RLAIF): the model compares pairs of its own responses against\nconstitutional principles, and those preference judgments train a reward model for\npolicy optimization. Because the feedback is AI-generated, the method scales\nwithout per-example human labeling — though humans still write the constitution,\ndesign the principles, and evaluate the results.\n\nThe transparency payoff is structural: unlike implicit labeler preferences, a\nconstitution can be published, versioned, debated, and audited. When the model\nrefuses or hedges, its behavior should trace back to identifiable principles.\n\n## When to use\n\n- Building an assistant that needs consistent, explainable behavior derived from\n  principles rather than black-box preference data.\n- Scaling alignment work when human labeler throughput is the bottleneck.\n- Reducing labeler-driven bias: a written constitution is auditable and\n  version-controlled; implicit labeler preferences are not.\n- Producing models that can explain refusals and judgments by citing the\n  underlying principles.\n- Researching scalable oversight — using AI to supervise AI — where the\n  constitution is the compact specification.\n- Organizations that need a governance artifact: the constitution doubles as the\n  documented behavior policy for review boards and regulators.\n\n## Core concepts\n\n- **The constitution**: a list of normative principles (\"choose the response that\n  is most helpful while avoiding harm...\"). Good constitutions are specific enough\n  to adjudicate real cases, short enough to be consistently applied, and layered\n  (e.g., corrigibility → duties → virtues).\n- **Self-critique and revision**: given a prompt and a draft response, the model is\n  asked to identify constitutional violations, then rewrite the response to fix\n  them. The critique-revision loop is where most of the alignment signal comes\n  from.\n- **RLAIF**: replacing human preference labels with AI-generated ones. The\n  critique model scores which of two responses better satisfies the constitution; a\n  reward model learns from these judgments.\n- **Red-teaming prompts**: the training distribution matters as much as the\n  constitution. Generate adversarial prompts that probe the principles' boundaries,\n  or the model never learns where they bite.\n- **Constitution as documentation**: unlike weights, a constitution can be\n  published, debated, and audited. This is the main transparency advantage of the\n  approach.\n- **Corrigibility**: the model's first duty — to remain shapeable by its\n  developers — precedes all other principles, so the constitution can't be used to\n  justify resisting oversight.\n- **Principle precedence**: when principles conflict, an explicit ordering decides.\n  Without precedence rules, the model resolves conflicts arbitrarily and\n  inconsistently.\n- **Critique model quality**: the ceiling of the whole method. A weak or biased\n  critique model bakes its flaws into every downstream artifact.\n\n## Practical workflow\n\n1. **Draft the constitution.** Write 10–40 principles organized in tiers: hard\n   constraints (safety, legality), then duties (honesty, fidelity to user), then\n   virtues (helpfulness, humility). Write them as instructions a careful reviewer\n   could apply.\n2. **Stress-test the draft.** Run the principles against hard cases before any\n   training. Where two careful readers disagree on what a principle requires,\n   rewrite the principle.\n3. **Generate critique data.** Sample diverse prompts including red-team cases. For\n   each, generate a base response, prompt the model to critique it against the\n   constitution, and produce a revised response.\n4. **Filter and fine-tune.** Keep revisions that genuinely improve on the originals\n   (a second model can judge this); discard trivial rewrites. Supervised-fine-tune\n   on the revision pairs.\n5. **Train the preference model.** For prompt-response pairs, have the critique\n   model rank responses by constitutional compliance. Train the reward model on\n   these rankings.\n6. **Run RL.** Optimize the policy against the reward model with a KL penalty to\n   the supervised baseline, monitoring for reward hacking on a held-out evaluation\n   set.\n7. **Evaluate against the constitution explicitly.** Build an eval set where each\n   item tests a specific principle. Report pass rates per principle — aggregate\n   scores hide principle-level failures.\n\nChecklist before deployment:\n- Each principle has dedicated eval items, including adversarial ones.\n- The model cites or reflects principles when refusing, not just stonewalling.\n- RLAIF judgments were spot-checked by humans for systematic bias.\n- Reward hacking monitored: KL divergence and eval scores tracked jointly.\n- Constitution version recorded alongside the model checkpoint.\n\n## Common pitfalls\n\n- **Vague constitutions.** \"Be good\" is not a constitution. Principles that can't\n  adjudicate a disputed case add no signal beyond generic helpfulness training.\n- **Self-grading bias.** The same model critiquing itself can entrench its own\n  blind spots. Use a stronger or differently-trained critique model where possible,\n  plus human spot checks.\n- **Critique quality as the ceiling.** RLAIF can only be as good as the critique\n  prompts and critique model. Invest in them like you'd invest in labeler training.\n- **Reward hacking.** The policy will find ways to score high on the reward model\n  without following the spirit of the principles. Hold out evals the reward model\n  never saw.\n- **Principle conflicts.** Principles genuinely conflict in edge cases (honesty vs.\n  kindness, helpfulness vs. safety). The constitution needs explicit precedence\n  rules; otherwise the model picks arbitrarily.\n- **Treating the constitution as secret sauce.** The methodology's transparency\n  benefit only materializes if you publish or at least document the constitution\n  and its rationale.\n- **Static constitutions.** Principles that made sense at training time drift from\n  organizational values. Version the constitution and re-evaluate on revision.\n- **Critique-revision collapse.** If revisions barely differ from originals, the\n  supervised stage teaches nothing. Measure revision distance and filter\n  aggressively.\n",{"name":5,"category":24,"description":7},"ai-research","main","af388a77c2fe5e0b4fd879cca4300a0ffa9a6579","skills\u002Fams-constitutional-ai",1,"https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis","https:\u002F\u002Fgithub.com\u002FTanNhatCMS","https:\u002F\u002Fmrkiet.name.vn","2026-11-07 14:23:08",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add TanNhatCMS\u002Fjarvis --skill constitutional-ai",{"tool":39,"command":40},"bunx","bunx skills add TanNhatCMS\u002Fjarvis --skill constitutional-ai",{"tool":42,"command":43},"pnpm","pnpm dlx skills add TanNhatCMS\u002Fjarvis --skill constitutional-ai","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fams-constitutional-ai\u002FSKILL.md",true,[],[48,55,63,70,77,84],{"skillNo":49,"name":50,"nameZh":6,"description":51,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":52,"categories":53,"keywords":54,"publishedAt":19},"01M7EWYJR0J8DARY7E59Z6Y9MV","web-scrape","Turn web pages into LLM-ready markdown via the Firecrawl API (renders JS, anti-bot handling). Use when browser_open\u002Ffetch fails on JS-heavy pages, or when you need clean markdown\u002Fstructured extraction from a site. Requires FIRECRAWL_API_KEY.","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fweb-scrape\u002FSKILL.md",[],[],{"skillNo":56,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":59,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7EWYBN5P6F2AJG3FR24T9BP","design-taste-frontend","tasteskill v2, the anti-slop frontend skill. Art direction for websites, distilled from a study of 247 curated gallery sites (One Page Love, Awwwards Sites of the Day). Deals researched directions (idea, composition, type, palette, photographs, section blueprints) from a library, enforces a strict no-slop house style with a real-browser inspector, and runs a look-and-fix loop on screenshots. Use this whenever you build, design, redesign, restyle, polish or review a landing page, marketing site, portfolio, product page, one-pager, hero or any public-facing web page, in any stack (HTML\u002FCSS, React, Next.js, Vue, Svelte, Astro, Tailwind) — also when the user says a page looks generic, templated, \"AI-made\", sloppy, or asks to make it look designed, premium, award-level or less like every other site. Not for dashboards or app screens.","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Ftaste-design\u002FLICENSE","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Ftaste-design\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M7EWY8FQFNQ2GB7NZZWGSSDX","test-driven-development","Use when implementing any feature or bugfix, before writing implementation code","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fsuperpowers-tdd\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7EWY5SHBFRH1HDHVWM9K7TV","systematic-debugging","Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fsuperpowers-debugging\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7EWY3NM0Z0JAKS7HDZ24K66","requesting-code-review","Use when completing tasks, implementing major features, or before merging to verify work meets requirements","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fsuperpowers-code-review\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7EWY0WRMWX4M9ZFY0WKC593","brainstorming","You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fsuperpowers-brainstorming\u002FSKILL.md",[],[]]