[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M756W1ZR0X5M2MNSP04P7M8Y":3,"skill-same:01M756W1ZR0X5M2MNSP04P7M8Y":53},{"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":29,"commitSha":30,"skillDirectory":31,"forksCount":32,"repositoryHtmlUrl":33,"repositoryHomepageUrl":34,"ownerHtmlUrl":35,"ownerBlogUrl":36,"ownerDescription":37,"updatedAt":38,"install":39,"resources":52},"01M756W1ZR0X5M2MNSP04P7M8Y","authoring-adept-loops",null,"Compose a loop — a scheduled system that discovers work, hands it to agents, verifies with an independent evaluator, persists state, and reschedules itself. Apply when the user wants automation that runs on a timer, a triage\u002Fbabysitter routine, or asks about `adept loop add`.","itaywol\u002Fadeptability","itaywol","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F6410154?v=4","01M756SKJABTDMYHWVMHBA0M6E","Itay Wolfish","01M756SKKJFSV073PS3KYR896D",9,"MIT","https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability\u002Fblob\u002Fmain\u002F.adeptability\u002Fstaging\u002F.claude\u002Fskills\u002Fauthoring-adept-loops\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: authoring-adept-loops\ndescription: Compose a loop — a scheduled system that discovers work, hands it to agents, verifies with an independent evaluator, persists state, and reschedules itself. Apply when the user wants automation that runs on a timer, a triage\u002Fbabysitter routine, or asks about `adept loop add`.\nallowed-tools: [Bash, Read, Edit, Write]\n---\n\n\n# Composing a loop with adept\n\nA loop is a system that prompts agents so a human doesn't have to: each turn it\n**discovers** work, **hands off** to an agent, **verifies** the result, **persists** state,\nand **reschedules**. Skip any move and the loop fails a predictable way — no verification is\nan agent nodding at itself; no persistence starts from scratch every morning; no schedule is\na script someone forgets to run.\n\nadept manages two of the six parts as canonical, synced resources — **skills** (discovery\nknowledge) and **agents** (the generator\u002Fevaluator pair). The rest (worktrees, connectors,\nmemory files, schedules) belong to the harness and repo; this skill tells you how to wire\nthem together.\n\n## Scaffold the composition\n\n```bash\nadept loop add ci-triage --workflow\n```\n\ncreates:\n\n- `skills\u002Fci-triage\u002F` — discovery skill (triage template: Read → Judge → Write → Hand off →\n  Stop). `activation: manual` on purpose: the *automation* invokes it by name; it must not\n  auto-fire in normal sessions.\n- `agents\u002Fci-triage-reviewer.md` — evaluator agent (adversarial template: assume broken,\n  act don't read, PASS\u002FREJECT verdict).\n- `.github\u002Fworkflows\u002Fadept-loop-ci-triage.yml` — cloud cron skeleton (only with\n  `--workflow`); fill in the harness invocation + auth secret.\n\nThen edit the scaffolds, `adept agent check ci-triage-reviewer`, `adept sync`.\n\n## The five moves, mapped\n\n| Move | Where it lives | adept's part |\n|---|---|---|\n| Discovery | triage **skill** — what to read, how to judge noise vs. actionable | `adept skill add \u003Cid> --template triage` |\n| Handoff | one git worktree per finding (`claude --worktree`, Codex background worktree) | skill's \"Hand off\" section emits the task lines |\n| Verification | evaluator **agent**, different instructions (ideally different model) than the maker | `adept agent add \u003Cid> --template evaluator` |\n| Persistence | `.\u002Fstate\u002F\u003Cid>.md` committed to the repo — the agent forgets, the repo does not | skill's \"Write\" section owns it |\n| Scheduling | cloud cron (machine off, ≥1h interval) or local `\u002Floop`\u002Fautomations (machine on, minute-level) | `--workflow` skeleton |\n\n## Non-negotiables (the parts that produce safety, not output)\n\n- **Independent evaluator.** An agent grading its own work praises it. Keep maker and\n  checker separate; tune the checker skeptical; make it verify by *acting* (run tests, click\n  the page), and let a stop condition (e.g. Claude Code `\u002Fgoal`) be judged by a fresh model.\n- **The Stop section.** The skill's boundary (\"Never merge. Never delete. Uncertain →\n  `.\u002Finbox\u002F` for a human\") is the one thing the loop cannot infer. Leave it out and the loop\n  merges with confidence it has not earned.\n- **Human review stays.** PRs open, never auto-merge. The pause is a permanent feature, not\n  scaffolding to remove once trusted.\n- **Cap before you ship.** Per-run budget, daily budget, max retries — set them before the\n  first unattended run, because one bug spinning idle overnight otherwise becomes a bill.\n- **Grow parallelism last.** Prove the evaluator catches real mistakes on one finding\n  end-to-end before letting five agents run at once.\n\n## Four silent costs (watch for them, tell the user)\n\nVerification debt (unreviewed output), comprehension rot (code nobody read), cognitive\nsurrender (nobody says \"this is wrong\" anymore), token blowout (retries all night). They\nreinforce each other; the guard for all four is a check that can say \"no\" plus a human who\nstill reads a sample daily.\n\nSee [[authoring-adept-skills]] for the discovery skill's craft, [[authoring-adept-agents]]\nfor evaluator design, and [[using-adept]] for the CLI. `adept loop add --help` is the source\nof truth for flags.\n","# Composing a loop with adept\n\nA loop is a system that prompts agents so a human doesn't have to: each turn it\n**discovers** work, **hands off** to an agent, **verifies** the result, **persists** state,\nand **reschedules**. Skip any move and the loop fails a predictable way — no verification is\nan agent nodding at itself; no persistence starts from scratch every morning; no schedule is\na script someone forgets to run.\n\nadept manages two of the six parts as canonical, synced resources — **skills** (discovery\nknowledge) and **agents** (the generator\u002Fevaluator pair). The rest (worktrees, connectors,\nmemory files, schedules) belong to the harness and repo; this skill tells you how to wire\nthem together.\n\n## Scaffold the composition\n\n```bash\nadept loop add ci-triage --workflow\n```\n\ncreates:\n\n- `skills\u002Fci-triage\u002F` — discovery skill (triage template: Read → Judge → Write → Hand off →\n  Stop). `activation: manual` on purpose: the *automation* invokes it by name; it must not\n  auto-fire in normal sessions.\n- `agents\u002Fci-triage-reviewer.md` — evaluator agent (adversarial template: assume broken,\n  act don't read, PASS\u002FREJECT verdict).\n- `.github\u002Fworkflows\u002Fadept-loop-ci-triage.yml` — cloud cron skeleton (only with\n  `--workflow`); fill in the harness invocation + auth secret.\n\nThen edit the scaffolds, `adept agent check ci-triage-reviewer`, `adept sync`.\n\n## The five moves, mapped\n\n| Move | Where it lives | adept's part |\n|---|---|---|\n| Discovery | triage **skill** — what to read, how to judge noise vs. actionable | `adept skill add \u003Cid> --template triage` |\n| Handoff | one git worktree per finding (`claude --worktree`, Codex background worktree) | skill's \"Hand off\" section emits the task lines |\n| Verification | evaluator **agent**, different instructions (ideally different model) than the maker | `adept agent add \u003Cid> --template evaluator` |\n| Persistence | `.\u002Fstate\u002F\u003Cid>.md` committed to the repo — the agent forgets, the repo does not | skill's \"Write\" section owns it |\n| Scheduling | cloud cron (machine off, ≥1h interval) or local `\u002Floop`\u002Fautomations (machine on, minute-level) | `--workflow` skeleton |\n\n## Non-negotiables (the parts that produce safety, not output)\n\n- **Independent evaluator.** An agent grading its own work praises it. Keep maker and\n  checker separate; tune the checker skeptical; make it verify by *acting* (run tests, click\n  the page), and let a stop condition (e.g. Claude Code `\u002Fgoal`) be judged by a fresh model.\n- **The Stop section.** The skill's boundary (\"Never merge. Never delete. Uncertain →\n  `.\u002Finbox\u002F` for a human\") is the one thing the loop cannot infer. Leave it out and the loop\n  merges with confidence it has not earned.\n- **Human review stays.** PRs open, never auto-merge. The pause is a permanent feature, not\n  scaffolding to remove once trusted.\n- **Cap before you ship.** Per-run budget, daily budget, max retries — set them before the\n  first unattended run, because one bug spinning idle overnight otherwise becomes a bill.\n- **Grow parallelism last.** Prove the evaluator catches real mistakes on one finding\n  end-to-end before letting five agents run at once.\n\n## Four silent costs (watch for them, tell the user)\n\nVerification debt (unreviewed output), comprehension rot (code nobody read), cognitive\nsurrender (nobody says \"this is wrong\" anymore), token blowout (retries all night). They\nreinforce each other; the guard for all four is a check that can say \"no\" plus a human who\nstill reads a sample daily.\n\nSee [[authoring-adept-skills]] for the discovery skill's craft, [[authoring-adept-agents]]\nfor evaluator design, and [[using-adept]] for the CLI. `adept loop add --help` is the source\nof truth for flags.\n",{"name":5,"description":7,"allowed-tools":24},[25,26,27,28],"Bash","Read","Edit","Write","main","c5b1b81b2e0ae71bd8afeeade28eaea6f8fd1bf5",".adeptability\u002Fstaging\u002F.claude\u002Fskills\u002Fauthoring-adept-loops",0,"https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability","https:\u002F\u002Fpkg.go.dev\u002Fgithub.com\u002Fitaywol\u002Fadeptability","https:\u002F\u002Fgithub.com\u002Fitaywol","https:\u002F\u002Fitaywol.com","Founding Engineer @ Echo","2026-07-25 15:28:37",{"source":8,"skillName":5,"commands":40,"prompt":50,"zipAvailable":51},[41,44,47],{"tool":42,"command":43},"npx","npx skills add itaywol\u002Fadeptability --skill authoring-adept-loops",{"tool":45,"command":46},"bunx","bunx skills add itaywol\u002Fadeptability --skill authoring-adept-loops",{"tool":48,"command":49},"pnpm","pnpm dlx skills add itaywol\u002Fadeptability --skill authoring-adept-loops","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability\u002Fblob\u002Fmain\u002F.adeptability\u002Fstaging\u002F.claude\u002Fskills\u002Fauthoring-adept-loops\u002FSKILL.md",true,[],[54,60,66,72,78,84],{"skillNo":55,"name":56,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":19},"01M756YHZ76TMVP0KK718K6KN0","using-adept","https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability\u002Fblob\u002Fmain\u002Fskills\u002Fusing-adept\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"publishedAt":19},"01M756YG9YGBG4783SFH8RE5PM","agents","https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability\u002Fblob\u002Fmain\u002Fskills\u002Fagents\u002FSKILL.md",[],[],{"skillNo":67,"name":68,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M756YD5MDX9ANB4QPVN34YQX","adept-writing-tests","https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability\u002Fblob\u002Fmain\u002Fskills\u002Fadept-writing-tests\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M756YB57DN31BKNT4GHM9N2S","adept-contributing","https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability\u002Fblob\u002Fmain\u002Fskills\u002Fadept-contributing\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":6,"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},"01M756Y9VM8W6E2HWMVYH0PH6H","adept-code-style","https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability\u002Fblob\u002Fmain\u002Fskills\u002Fadept-code-style\u002FSKILL.md",[],[],{"skillNo":85,"name":56,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":19},"01M756Y76FNSH8VWP77TX1A8Z1","Use the `adept` CLI to author an AI skill once and render it into Claude Code, Cursor, Codex, Copilot, OpenCode, and 45+ agents. Apply when installing\u002Fsyncing skills, editing SKILL.md, or touching .adeptability\u002F.","https:\u002F\u002Fgithub.com\u002Fitaywol\u002Fadeptability\u002Fblob\u002Fmain\u002Finternal\u002Fdefaultskills\u002Fassets\u002Fusing-adept\u002FSKILL.md",[],[]]