[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7DNENKBJQ0TH9ST67WY9BH9":3,"skill-same:01M7DNENKBJQ0TH9ST67WY9BH9":48},{"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":32,"updatedAt":33,"install":34,"resources":47},"01M7DNENKBJQ0TH9ST67WY9BH9","conduct",null,"Walk this project's phases with the conductor pattern — read where it stands, brief a subagent on the next phase, verify the named proof yourself, record what changed course, move the status, commit the phase whole to main, and go on until a step needs a person. Use for \"\u002Fconduct\", \"conduct the next phase\", \"do phase 3\", \"where do we stand\", or \"keep going\".","dalmaer\u002Fplates","dalmaer","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F4770?v=4","01M7DNDRF9ZW4QHZAKEX3WWX99","Dion Almaer","01M7DNDRGD9NZADJEA2J02ET39",4,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fimprove\u002Funhealthy\u002F.agents\u002Fskills\u002Fconduct\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: conduct\ndescription: Walk this project's phases with the conductor pattern — read where it stands, brief a subagent on the next phase, verify the named proof yourself, record what changed course, move the status, commit the phase whole to main, and go on until a step needs a person. Use for \"\u002Fconduct\", \"conduct the next phase\", \"do phase 3\", \"where do we stand\", or \"keep going\".\nargument-hint: \"[status | \u003Cphase number> | one]\"\n---\n\n\u003C!-- Source: dglazkov\u002Fisocan .claude\u002Fskills\u002Fconduct\u002FSKILL.md @ 7227f325 (Apache-2.0).\n     Adapted by plates for one-file-per-phase projects (docs\u002Fphases\u002FNN-*.md).\n     Plates manages this file: change it upstream in plates, or send the change home with `plates lessons`. -->\n\n# conduct: one phase, verified, recorded, committed\n\nThe phases in `docs\u002Fphases\u002F` are the contract. Each one is a file. Its front\nmatter owns its status, and its sections say **Done when**, **Acceptance**,\n**Proof**, **Deliberately open** and **Next action**. `docs\u002Fgoals.json` says\nwhat the phases are for. `docs\u002Fdesign.md` (or the project's equivalent) is the\nargument they cite.\n\nThe session that runs this skill is the **conductor**. It does not build. It\nbriefs a subagent that builds, then verifies the proof the phase named up\nfront, never taking the subagent's word for it. Only then does it write the\nrecord and commit. Then it goes on to the next phase.\n\n**The default is unattended.** The conductor stops for a person, and a person\nis needed for exactly three things:\n\n- a token or a login;\n- money or a cloud resource (a ⚑ step);\n- a hand the phase names (a real device, a decision assigned to someone by name).\n\nEverything else is the conductor's to decide, record and commit: a wrong\ndesign, a proof that can't run as written, two docs that disagree.\n\n`AGENTS.md` applies to everyone, conductor and subagent alike. Invoking\n`\u002Fconduct` is the \"land this\": the conductor commits and pushes each phase to\n`main` without asking again.\n\n## Arguments\n\n- `\u002Fconduct`: orient, then conduct phase after phase until a step needs a\n  person or nothing is left. Commit and push after each phase. The user reads\n  progress from the commits.\n- `\u002Fconduct status`: orient and report. Change nothing.\n- `\u002Fconduct \u003CN>`: conduct that phase and stop. If a phase it depends on isn't\n  built, conduct that one first and say so.\n- `\u002Fconduct one`: the next phase only, then stop.\n\n## 0. Orient\n\n```sh\nnpm run next                          # the next phase whose dependencies are built, and its next action\nnode scripts\u002Froadmap.mjs --json       # everything, if you need more\n```\n\nThen read, in this order:\n\n1. the phase file;\n2. the design sections it cites;\n3. the **Trajectory** of the phases it depends on (what the design didn't know);\n4. `docs\u002Flessons.md` rows that share its shape.\n\nNote the wall clock. The commit records what a phase cost.\n\nKeep the conductor's own context lean, because it outlives every builder:\n\n- Send long output to a file and read the lines that matter (`grep -E \"fail|not ok|# pass\"`).\n- Don't re-read what a builder already summarised unless verification needs it.\n- On a long run, keep a short handoff file in the scratchpad: current commit,\n  owned paths, decisions, open failures.\n\n**Three gates before any code:**\n\n- **The docs agree.** If the phase's Proof, its Acceptance and the design\n  disagree, fix the docs first, as their own commit. Building on a\n  contradiction produces something that satisfies one document and not the other.\n- **Its dependencies are built.** `depends:` says which phases. Conduct an unbuilt one first.\n- **⚑ steps are asked, not done.** List them to the user with the price, and\n  get a yes for each. Before stopping to ask, do every part that doesn't need\n  the answer, so the ask is the only thing left.\n\n## 1. Brief\n\nWrite the brief to a scratchpad file, so it can be reread, reused if the\nsubagent must be restarted, and quoted in the commit.\n\n**Point, don't paste.** The subagent can read the repo, and `AGENTS.md` is\nalready in its context. Name the phase file, the design sections and the\nlessons by path and heading. In words, the brief carries only what the docs\n*don't* say:\n\n- decisions taken in this session;\n- which files are whose;\n- who else is working in the tree;\n- where to stop.\n\nAim for under 600 words.\n\n```\n# \u003Cproject> phase \u003CN>: \u003Ctitle>\n\nContract: docs\u002Fphases\u002F\u003CNN-name>.md. Design: \u003Cdoc> \u003Csections>.\nTrajectory that binds you: \u003Cone line each, only ones that change what you build>.\nDecided here, not in the docs: \u003C…>\n\n## Who else is in the tree   (another builder's paths, and which red checks are theirs)\n## What you own\nFiles under \u003Cpaths>. Not docs\u002Fphases\u002F, docs\u002FROADMAP.md, docs\u002Flessons.md or the\nevidence: the conductor writes the record.\n\n## Where you stop\n- At each ⚑ step without a yes: build up to it, report.\n- When the proof would need a facade (a fixture that can't fail, a shim that\n  answers the test and nothing else): stop and say so. That's a finding.\n- When the design turns out wrong: stop, say what you found and what you'd\n  change. The conductor changes the design.\n- A bug beside the phase: report it, don't fix it.\n\n## How you test\nThe test files you touched or wrote, by path (`node --test tests\u002F\u003Cfile>`),\nas often as you like. Do NOT run the whole check (`npm run check`): the\nconductor runs it once on the integrated tree. Never run a build beside\nanother build.\n\n## What you return (under 600 words)\n1. What was built: the files, and one paragraph on how it works.\n2. The proof: exact commands from the repo root, each with its exit code and\n   the one line that says so. Not \"tests pass\", and not the whole output.\n3. What you couldn't do, why, and where you stopped.\n4. Candidate trajectory lines: dated, one claim each, about forty words, only\n   for things that change what comes after.\n5. Anything a later phase should know that the docs don't say.\n```\n\nSpawn the subagent with the Agent tool (`general-purpose`), in the background.\n\nParallel builders belong inside one phase, split by file ownership. Never run\ntwo phases of one project at once: two builders in one checkout see each\nother's half-done files. Name each builder's paths in the other's brief, and\nsay which red checks to expect from the neighbour.\n\nA subagent's report isn't shown to the user; relay what matters.\n\n## 2. Verify\n\nThe builder's report is a map. The phase's **Proof** is the territory. Run it\nas written, from the repo root, and read exit codes, not tails.\n\n- `git pull --ff-only` **first**.\n- Then mark the phase's new files intent-to-add: `git add -N \u003Cnew files>`.\n  Anything that measures with `git ls-files` can't see an untracked file\n  locally, and it will fail the same check in CI after the commit (lesson 10).\n  Never stage before an autostash pull: the pull hands staged changes back unstaged.\n- `git status --short`: only the phase's files changed, no leftovers, no\n  `.env`, nothing under `docs\u002Fphases\u002F`.\n- **The whole check, once**, on the integrated tree: `npm run check`. Start it\n  detached and do the reading below while it runs. If something fails, re-run\n  just that, alone, before calling it a flake.\n- The named Proof, command by command.\n- The walk, when the phase has one: a real browser, a real device, a real\n  `gh` call. A phase with a walk isn't built until the walk is walked. If a\n  person must walk it, write the steps into the evidence file.\n- Open each new test and ask two things. Could it fail? Does it exercise the\n  thing, or a stand-in for it? A test that asserts what the code returns,\n  rather than what the phase requires, proves only that the code agrees with\n  itself.\n- Read the diff for a facade, and for real data carried into fixtures.\n  Fixtures are synthetic: \"Acme\", always.\n\n**Work that fails goes back down.** Send it to the same subagent with\n`SendMessage`, so it keeps its context. Include the exact command, the exact\noutput, and the line of the Proof it violates. Don't fix code yourself; the\nconductor edits documents.\n\n## 3. Record\n\nWhen the proof holds, and only then, write the record, all in one change:\n\n- **The evidence file**, `docs\u002Fevidence\u002F\u003Cdate>-\u003Cslug>.md`, from\n  `docs\u002Ftemplates\u002Fevidence.md`: what was run, with exit codes; what was\n  checked by hand; what was not. Never write expectations as observations.\n- **The phase's front matter**: `status`, `since` (today), and a `note`\n  saying in one line why it stands there. Add the evidence path. Check the\n  Acceptance boxes that the proof actually showed.\n- **Trajectory** in the phase file: only what changes the course. One line\n  per claim: `- **YYYY-MM-DD** — Claim. Evidence.` A phase that went as\n  planned writes `*Nothing — the phase went as planned.*`\n- **Deliberately open**: settle the questions that were settled, in place,\n  dated, saying what settled them.\n- **Next action** for the phase, if it isn't built.\n- **Lessons**: a new row only if a bug turned out to have a shape.\n- **Other phases** this one changed the facts for: `grep -rn` the term across\n  `docs\u002F` and fix each mention.\n- `npm run roadmap`, then `npm run roadmap:check`.\n\n## 4. Commit\n\nOne commit per phase, with the phase whole: code, tests, record.\n\n- **Title:** `phase \u003CN>: \u003Ctitle>`, or `\u003Carea>: \u003Cwhat>` for something smaller.\n- **Body:** the argument, in prose a reader who wasn't there can follow. What\n  was built, what the proof showed, what changed course, and what it cost in\n  wall time.\n\nLand on `main`, no pull request:\n\n1. Integrate `origin\u002Fmain` before the final check, and regenerate the roadmap\n   if it conflicts.\n2. Stage an **explicit list** of the phase's files, taken from the builder's\n   report and `git diff --stat`. Never `git add -A`: a builder still writing in\n   the tree creates files between your read and your stage.\n3. Read `git show --stat HEAD` before pushing.\n4. Push, and follow CI on that exact commit while preparing the next brief.\n\nWrite a short report between phases (the phase, its new status, the proof and\nwhat it printed, what changed course). Don't wait for a reply.\n\n## Stopping\n\nThe conductor stops for a person, and for nothing else. Record the stop first,\nso a later session can pick up from the docs alone:\n\n- Set status to `partial`, with a note naming exactly what waits on whom.\n- Put the ask in Next action.\n- Commit, push.\n\nThen look past it: if a later phase needs none of that, conduct it. The final\nreport lists every ask in one place, each in one sentence with its price.\n\nTwo things look like stops and aren't. Decide them, record them, and go on:\n\n- **The design is wrong.** Change the design doc as its own commit, with the\n  reason, then re-brief. A phase can be superseded; it can't be quietly redefined.\n- **The proof can't run as written.** Fix the Proof, say why in the commit,\n  continue. Never mark a phase built against a proof other than the one it named.\n\nIf in doubt, ask: is the missing thing a credential, money or a hand? If not,\nit's yours.\n\n## Things that have gone wrong before\n\n- A subagent said the suite passed. It had run the suite piped into `tail`,\n  and six tests had failed. Read the exit code.\n- Builders ran the whole suite 53 times in one session, and the conductor's\n  run repeated all of it. Builders test by file.\n- A brief pasted the phase, the design and the house rules. The conductor read\n  them once and the subagent twice. Point, don't paste.\n- A script shipped without its library, because `git add -A` ran while a\n  builder was still writing. Stage an explicit list; read `git show --stat HEAD`.\n- An autostash pull handed staged files back unstaged, and the commit missed\n  them. Pull first, then mark files intent-to-add.\n- A subagent fixed a bug beside the phase, and the fix was wrong, because the\n  fix was a design call. Report it; don't fix it unasked.\n- Trajectory became a work log, and the one line that redrew the map was\n  buried. Trajectory holds course changes only.\n- A real project's names ended up in a fixture and then in a shipped example.\n  Fixtures are synthetic.\n- The conductor proved a drift guard by editing a file, then undid it with\n  `git checkout`, which also threw away the builder's uncommitted change to\n  that file. Probe a temp copy, never the tree a builder has written.\n- The shell's cwd was reset between commands. Use absolute paths in every\n  command you give a subagent.\n","\u003C!-- Source: dglazkov\u002Fisocan .claude\u002Fskills\u002Fconduct\u002FSKILL.md @ 7227f325 (Apache-2.0).\n     Adapted by plates for one-file-per-phase projects (docs\u002Fphases\u002FNN-*.md).\n     Plates manages this file: change it upstream in plates, or send the change home with `plates lessons`. -->\n\n# conduct: one phase, verified, recorded, committed\n\nThe phases in `docs\u002Fphases\u002F` are the contract. Each one is a file. Its front\nmatter owns its status, and its sections say **Done when**, **Acceptance**,\n**Proof**, **Deliberately open** and **Next action**. `docs\u002Fgoals.json` says\nwhat the phases are for. `docs\u002Fdesign.md` (or the project's equivalent) is the\nargument they cite.\n\nThe session that runs this skill is the **conductor**. It does not build. It\nbriefs a subagent that builds, then verifies the proof the phase named up\nfront, never taking the subagent's word for it. Only then does it write the\nrecord and commit. Then it goes on to the next phase.\n\n**The default is unattended.** The conductor stops for a person, and a person\nis needed for exactly three things:\n\n- a token or a login;\n- money or a cloud resource (a ⚑ step);\n- a hand the phase names (a real device, a decision assigned to someone by name).\n\nEverything else is the conductor's to decide, record and commit: a wrong\ndesign, a proof that can't run as written, two docs that disagree.\n\n`AGENTS.md` applies to everyone, conductor and subagent alike. Invoking\n`\u002Fconduct` is the \"land this\": the conductor commits and pushes each phase to\n`main` without asking again.\n\n## Arguments\n\n- `\u002Fconduct`: orient, then conduct phase after phase until a step needs a\n  person or nothing is left. Commit and push after each phase. The user reads\n  progress from the commits.\n- `\u002Fconduct status`: orient and report. Change nothing.\n- `\u002Fconduct \u003CN>`: conduct that phase and stop. If a phase it depends on isn't\n  built, conduct that one first and say so.\n- `\u002Fconduct one`: the next phase only, then stop.\n\n## 0. Orient\n\n```sh\nnpm run next                          # the next phase whose dependencies are built, and its next action\nnode scripts\u002Froadmap.mjs --json       # everything, if you need more\n```\n\nThen read, in this order:\n\n1. the phase file;\n2. the design sections it cites;\n3. the **Trajectory** of the phases it depends on (what the design didn't know);\n4. `docs\u002Flessons.md` rows that share its shape.\n\nNote the wall clock. The commit records what a phase cost.\n\nKeep the conductor's own context lean, because it outlives every builder:\n\n- Send long output to a file and read the lines that matter (`grep -E \"fail|not ok|# pass\"`).\n- Don't re-read what a builder already summarised unless verification needs it.\n- On a long run, keep a short handoff file in the scratchpad: current commit,\n  owned paths, decisions, open failures.\n\n**Three gates before any code:**\n\n- **The docs agree.** If the phase's Proof, its Acceptance and the design\n  disagree, fix the docs first, as their own commit. Building on a\n  contradiction produces something that satisfies one document and not the other.\n- **Its dependencies are built.** `depends:` says which phases. Conduct an unbuilt one first.\n- **⚑ steps are asked, not done.** List them to the user with the price, and\n  get a yes for each. Before stopping to ask, do every part that doesn't need\n  the answer, so the ask is the only thing left.\n\n## 1. Brief\n\nWrite the brief to a scratchpad file, so it can be reread, reused if the\nsubagent must be restarted, and quoted in the commit.\n\n**Point, don't paste.** The subagent can read the repo, and `AGENTS.md` is\nalready in its context. Name the phase file, the design sections and the\nlessons by path and heading. In words, the brief carries only what the docs\n*don't* say:\n\n- decisions taken in this session;\n- which files are whose;\n- who else is working in the tree;\n- where to stop.\n\nAim for under 600 words.\n\n```\n# \u003Cproject> phase \u003CN>: \u003Ctitle>\n\nContract: docs\u002Fphases\u002F\u003CNN-name>.md. Design: \u003Cdoc> \u003Csections>.\nTrajectory that binds you: \u003Cone line each, only ones that change what you build>.\nDecided here, not in the docs: \u003C…>\n\n## Who else is in the tree   (another builder's paths, and which red checks are theirs)\n## What you own\nFiles under \u003Cpaths>. Not docs\u002Fphases\u002F, docs\u002FROADMAP.md, docs\u002Flessons.md or the\nevidence: the conductor writes the record.\n\n## Where you stop\n- At each ⚑ step without a yes: build up to it, report.\n- When the proof would need a facade (a fixture that can't fail, a shim that\n  answers the test and nothing else): stop and say so. That's a finding.\n- When the design turns out wrong: stop, say what you found and what you'd\n  change. The conductor changes the design.\n- A bug beside the phase: report it, don't fix it.\n\n## How you test\nThe test files you touched or wrote, by path (`node --test tests\u002F\u003Cfile>`),\nas often as you like. Do NOT run the whole check (`npm run check`): the\nconductor runs it once on the integrated tree. Never run a build beside\nanother build.\n\n## What you return (under 600 words)\n1. What was built: the files, and one paragraph on how it works.\n2. The proof: exact commands from the repo root, each with its exit code and\n   the one line that says so. Not \"tests pass\", and not the whole output.\n3. What you couldn't do, why, and where you stopped.\n4. Candidate trajectory lines: dated, one claim each, about forty words, only\n   for things that change what comes after.\n5. Anything a later phase should know that the docs don't say.\n```\n\nSpawn the subagent with the Agent tool (`general-purpose`), in the background.\n\nParallel builders belong inside one phase, split by file ownership. Never run\ntwo phases of one project at once: two builders in one checkout see each\nother's half-done files. Name each builder's paths in the other's brief, and\nsay which red checks to expect from the neighbour.\n\nA subagent's report isn't shown to the user; relay what matters.\n\n## 2. Verify\n\nThe builder's report is a map. The phase's **Proof** is the territory. Run it\nas written, from the repo root, and read exit codes, not tails.\n\n- `git pull --ff-only` **first**.\n- Then mark the phase's new files intent-to-add: `git add -N \u003Cnew files>`.\n  Anything that measures with `git ls-files` can't see an untracked file\n  locally, and it will fail the same check in CI after the commit (lesson 10).\n  Never stage before an autostash pull: the pull hands staged changes back unstaged.\n- `git status --short`: only the phase's files changed, no leftovers, no\n  `.env`, nothing under `docs\u002Fphases\u002F`.\n- **The whole check, once**, on the integrated tree: `npm run check`. Start it\n  detached and do the reading below while it runs. If something fails, re-run\n  just that, alone, before calling it a flake.\n- The named Proof, command by command.\n- The walk, when the phase has one: a real browser, a real device, a real\n  `gh` call. A phase with a walk isn't built until the walk is walked. If a\n  person must walk it, write the steps into the evidence file.\n- Open each new test and ask two things. Could it fail? Does it exercise the\n  thing, or a stand-in for it? A test that asserts what the code returns,\n  rather than what the phase requires, proves only that the code agrees with\n  itself.\n- Read the diff for a facade, and for real data carried into fixtures.\n  Fixtures are synthetic: \"Acme\", always.\n\n**Work that fails goes back down.** Send it to the same subagent with\n`SendMessage`, so it keeps its context. Include the exact command, the exact\noutput, and the line of the Proof it violates. Don't fix code yourself; the\nconductor edits documents.\n\n## 3. Record\n\nWhen the proof holds, and only then, write the record, all in one change:\n\n- **The evidence file**, `docs\u002Fevidence\u002F\u003Cdate>-\u003Cslug>.md`, from\n  `docs\u002Ftemplates\u002Fevidence.md`: what was run, with exit codes; what was\n  checked by hand; what was not. Never write expectations as observations.\n- **The phase's front matter**: `status`, `since` (today), and a `note`\n  saying in one line why it stands there. Add the evidence path. Check the\n  Acceptance boxes that the proof actually showed.\n- **Trajectory** in the phase file: only what changes the course. One line\n  per claim: `- **YYYY-MM-DD** — Claim. Evidence.` A phase that went as\n  planned writes `*Nothing — the phase went as planned.*`\n- **Deliberately open**: settle the questions that were settled, in place,\n  dated, saying what settled them.\n- **Next action** for the phase, if it isn't built.\n- **Lessons**: a new row only if a bug turned out to have a shape.\n- **Other phases** this one changed the facts for: `grep -rn` the term across\n  `docs\u002F` and fix each mention.\n- `npm run roadmap`, then `npm run roadmap:check`.\n\n## 4. Commit\n\nOne commit per phase, with the phase whole: code, tests, record.\n\n- **Title:** `phase \u003CN>: \u003Ctitle>`, or `\u003Carea>: \u003Cwhat>` for something smaller.\n- **Body:** the argument, in prose a reader who wasn't there can follow. What\n  was built, what the proof showed, what changed course, and what it cost in\n  wall time.\n\nLand on `main`, no pull request:\n\n1. Integrate `origin\u002Fmain` before the final check, and regenerate the roadmap\n   if it conflicts.\n2. Stage an **explicit list** of the phase's files, taken from the builder's\n   report and `git diff --stat`. Never `git add -A`: a builder still writing in\n   the tree creates files between your read and your stage.\n3. Read `git show --stat HEAD` before pushing.\n4. Push, and follow CI on that exact commit while preparing the next brief.\n\nWrite a short report between phases (the phase, its new status, the proof and\nwhat it printed, what changed course). Don't wait for a reply.\n\n## Stopping\n\nThe conductor stops for a person, and for nothing else. Record the stop first,\nso a later session can pick up from the docs alone:\n\n- Set status to `partial`, with a note naming exactly what waits on whom.\n- Put the ask in Next action.\n- Commit, push.\n\nThen look past it: if a later phase needs none of that, conduct it. The final\nreport lists every ask in one place, each in one sentence with its price.\n\nTwo things look like stops and aren't. Decide them, record them, and go on:\n\n- **The design is wrong.** Change the design doc as its own commit, with the\n  reason, then re-brief. A phase can be superseded; it can't be quietly redefined.\n- **The proof can't run as written.** Fix the Proof, say why in the commit,\n  continue. Never mark a phase built against a proof other than the one it named.\n\nIf in doubt, ask: is the missing thing a credential, money or a hand? If not,\nit's yours.\n\n## Things that have gone wrong before\n\n- A subagent said the suite passed. It had run the suite piped into `tail`,\n  and six tests had failed. Read the exit code.\n- Builders ran the whole suite 53 times in one session, and the conductor's\n  run repeated all of it. Builders test by file.\n- A brief pasted the phase, the design and the house rules. The conductor read\n  them once and the subagent twice. Point, don't paste.\n- A script shipped without its library, because `git add -A` ran while a\n  builder was still writing. Stage an explicit list; read `git show --stat HEAD`.\n- An autostash pull handed staged files back unstaged, and the commit missed\n  them. Pull first, then mark files intent-to-add.\n- A subagent fixed a bug beside the phase, and the fix was wrong, because the\n  fix was a design call. Report it; don't fix it unasked.\n- Trajectory became a work log, and the one line that redrew the map was\n  buried. Trajectory holds course changes only.\n- A real project's names ended up in a fixture and then in a shipped example.\n  Fixtures are synthetic.\n- The conductor proved a drift guard by editing a file, then undid it with\n  `git checkout`, which also threw away the builder's uncommitted change to\n  that file. Probe a temp copy, never the tree a builder has written.\n- The shell's cwd was reset between commands. Use absolute paths in every\n  command you give a subagent.\n",{"name":5,"description":7,"argument-hint":24},"[status | \u003Cphase number> | one]","main","6d2869a5bf10b3c2b966834c76784bbd07070478","tests\u002Ffixtures\u002Fimprove\u002Funhealthy\u002F.agents\u002Fskills\u002Fconduct",0,"https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates","https:\u002F\u002Fgithub.com\u002Fdalmaer","http:\u002F\u002Falmaer.com\u002F","Nudging in the grey, surprisingly older than @bgalbs, and democratizing software creation with AI @Google.","2026-11-15 06:37:38",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":46},[36,39,42],{"tool":37,"command":38},"npx","npx skills add dalmaer\u002Fplates --skill conduct",{"tool":40,"command":41},"bunx","bunx skills add dalmaer\u002Fplates --skill conduct",{"tool":43,"command":44},"pnpm","pnpm dlx skills add dalmaer\u002Fplates --skill conduct","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fimprove\u002Funhealthy\u002F.agents\u002Fskills\u002Fconduct\u002FSKILL.md",true,[],[49,56,63,69,76,81],{"skillNo":50,"name":51,"nameZh":6,"description":52,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":53,"categories":54,"keywords":55,"publishedAt":19},"01M7DNEKKNX8ANMQEE97AG4R5A","acme-conduct","Acme Canvas's own conductor (the original plates adapted).","https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fancestors\u002Facme-canvas\u002F.claude\u002Fskills\u002Fconduct\u002FSKILL.md",[],[],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7DNED6AD5FV5QG9XW7A7H1B","spin","Walk this project's phases with the conductor pattern — read where it stands, brief a subagent on the next phase, verify the named proof yourself, record what changed course, move the status, land independently safe increments on main throughout the phase, and go on until a step needs a person. Use for \"\u002Fspin\", \"conduct the next phase\", \"do phase 3\", \"where do we stand\", or \"keep going\"; and for \"\u002Fgoodnight\", when the owner asks to shape tonight's plan.","https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates\u002Fblob\u002Fmain\u002Fpractices\u002Fconduct\u002Ffiles\u002F.agents\u002Fskills\u002Fspin\u002FSKILL.md",[],[],{"skillNo":64,"name":5,"nameZh":6,"description":65,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"publishedAt":19},"01M7DNEA8D2GE16ARW4P1BZ8Z2","Compatibility alias for \u002Fspin; walk the same project phases with the same arguments, proof, claims, reviews, permissions and stopping rules.","https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates\u002Fblob\u002Fmain\u002Fpractices\u002Fconduct\u002Ffiles\u002F.agents\u002Fskills\u002Fconduct\u002FSKILL.md",[],[],{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":19},"01M7DNE76HNVDJGG3HFTT0VSPK","plates","Start here in any repo with a .plates\u002Fplates.json, or when asked about its phases, the roadmap, lessons, the night shift, evidence, or plates itself — it says what plates is and where its instructions live.","https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates\u002Fblob\u002Fmain\u002Fpractices\u002Fagents-md\u002Ffiles\u002F.agents\u002Fskills\u002Fplates\u002FSKILL.md",[],[],{"skillNo":77,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M7DNE0C2WV8DPVJA1E0MEYF1","https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fspin\u002FSKILL.md",[],[],{"skillNo":82,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M7DNDWQJG6D80NK9FC5PV8YA","https:\u002F\u002Fgithub.com\u002Fdalmaer\u002Fplates\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fplates\u002FSKILL.md",[],[]]