[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6KX1MJ7Y0CVJMYR8GHXZ7YA":3,"skill-same:01M6KX1MJ7Y0CVJMYR8GHXZ7YA":95},{"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":31,"publishedAt":52,"linkOnly":53,"markdownContent":54,"markdownBody":55,"frontMatter":56,"refName":58,"commitSha":59,"skillDirectory":60,"forksCount":61,"repositoryHtmlUrl":62,"repositoryHomepageUrl":6,"ownerHtmlUrl":63,"ownerBlogUrl":64,"ownerDescription":65,"updatedAt":66,"install":67,"resources":80},"01M6KX1MJ7Y0CVJMYR8GHXZ7YA","skill-writer",null,"Writes and reviews shared agent skills plus Claude Code Markdown and Codex TOML subagents for this plugin and bootstrapped projects. Covers choosing how prescriptive to be, the house writing style, whether a skill is user- or model-invoked and how skills call each other, writing the description field that controls when a skill fires, splitting content across reference files, the allowed-tools vs tools frontmatter distinction, and harness-kit's own conventions (shared scripts, dual plugin registration and host-specific script paths). Use when creating a new skill or agent, editing an existing one, writing a pointer line in AGENTS.md or CLAUDE.md, or deciding whether something should be a skill, an agent, a script, or nothing at all.","dougborg\u002Fharness-kit","dougborg","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F1261222?v=4","01M6KWS4PB51EYD4ZDJJSGNGB2","Doug Borg","01M6KWS4T8ZZW6SX5QZS9B5BN2",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fmeta\u002Fskill-writer\u002FSKILL.md",[18,23,26],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000018","写作与文案","SCENARIO","使用场景",{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":27,"categoryName":28,"taxonomyType":29,"taxonomyTypeDesc":30},"01JBSEEDCAT000000000000036","内容创作者","PROFESSION","适用职业",[32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51],"skill writer","writer","files","at","kit","reviews","scripts","line","host","shared","across","writing","edit","write","read","model","projects","markdown","how","code","2026-11-10 10:12:30",false,"---\nname: skill-writer\ndescription: >-\n  Writes and reviews shared agent skills plus Claude Code Markdown and Codex\n  TOML subagents for this plugin and bootstrapped projects. Covers choosing how\n  prescriptive to be, the house writing style, whether a skill is user- or\n  model-invoked and how skills call each other, writing the description field\n  that controls when a skill fires, splitting content across reference files,\n  the allowed-tools vs tools frontmatter distinction, and harness-kit's own\n  conventions (shared scripts, dual plugin registration and host-specific\n  script paths). Use when creating a new skill or agent, editing an existing\n  one, writing a pointer line in AGENTS.md or CLAUDE.md, or deciding whether\n  something should be a skill, an agent, a script, or nothing at all.\nallowed-tools: Glob, Grep, Read, Write, Edit\n---\n\n# \u002Fskill-writer — Authoring Skills and Agents\n\nClaude already knows the skill format. This skill exists for the judgment calls\nthe format does not make for you: how prescriptive to be, what earns a place in\na file that stays in context all session, and the conventions specific to this\nrepo.\n\n## First, decide whether to write one at all\n\nA skill is worth writing when the task recurs, and when Claude's default\nbehavior on it is wrong or under-informed. If default performance is already\ngood, a skill that restates the obvious makes output worse, not better — it\nspends context and over-constrains a capable model.\n\nPrefer the smallest thing that works:\n\n| Need | Reach for |\n| --- | --- |\n| A fixed sequence of commands | A script in `scripts\u002Fshared\u002F` |\n| Project facts Claude must always know | `CLAUDE.md` |\n| A repeatable workflow, or discipline the agent should reach for mid-task | A skill |\n| Wide reads\u002Fsearches whose output should stay out of the main context | An agent |\n| Something that must happen automatically, every time | A hook |\n\n## Match specificity to fragility\n\nThe central authoring decision is **degrees of freedom** — how much latitude to\nleave. Set it per-instruction, based on how fragile the operation is, not by\napplying one uniform level to a whole file.\n\n- **High freedom** — prose, heuristics, a stated goal. Use when several\n  approaches are valid and the right one depends on context. \"Summarize what\n  changed and why, in the reviewer's terms.\"\n- **Medium freedom** — a preferred pattern with room to vary: pseudocode, a\n  parameterized script, a worked example to adapt.\n- **Low freedom** — exact commands, explicit ordering, \"do not add flags to\n  this command.\" Use when the operation is fragile, when consistency across\n  runs matters, or when a specific sequence must hold.\n\nThe test: a narrow bridge with cliffs on both sides gets guardrails; an open\nfield gets a direction to walk in. Over-guardrailing an open field is the more\ncommon failure here — it reads as thorough and degrades results.\n\nCorollary: skills written for older models are often too prescriptive for\ncurrent ones. When editing an existing skill, actively look for instructions to\n**delete**, and check whether default behavior is already better than what the\nskill mandates.\n\n## Every line is a recurring cost\n\nOnce a skill loads, its content stays in context for the rest of the session.\nAfter auto-compaction, Claude Code re-attaches only the **first 5,000 tokens**\nof each invoked skill, inside a **25,000-token** combined budget filled\nmost-recent-first — an oversized skill gets truncated mid-file and pushes older\nskills out entirely.\n\nSo: state what to do, with the reason in a clause rather than a narrative.\nChallenge each line — does it justify its cost? Assume the reader is smart.\n\nThe real limits, and the only ones worth quoting:\n\n| Thing | Limit |\n| --- | --- |\n| SKILL.md body | Under 500 lines — split into sibling files past that |\n| `description` | 1,024 characters |\n| `description` + `when_to_use` combined | 1,536 characters in the skill listing |\n| Post-compaction re-attachment | First 5,000 tokens per skill, 25,000 total |\n\nThere is no prescribed section schema and no per-section token budget. Use the\nheadings the content actually needs.\n\n## Context pointers\n\nA **context pointer** is an always-loaded line that names material outside\nthe context and says when to reach for it: a skill's description, or an\n`AGENTS.md` \u002F `CLAUDE.md` line naming a doc. Its wording, not its target,\ndecides how reliably the agent reaches the material. When must-have material\nbehind a pointer gets missed, sharpen the pointer first; inline the material\nonly if that fails.\n\nA pointer states what the material is and lists its **branches**, the\ndistinct cases that should send the agent there. It costs tokens on every\nturn, so prune it harder than any body:\n\n- Front-load the leading word; that is where the triggering happens.\n- One trigger per branch. Synonyms for one branch are that branch written\n  twice.\n- Cut anything the target already says about itself.\n\n## Writing the body\n\nThe house style is short plain prose. These levers decide whether an agent\ntakes the same path through a skill every run.\n\n- **Give each step a completion criterion.** (A mode skill, which shapes\n  every response until switched off, has no steps; its exit phrase is its\n  criterion.) End every step on the condition\n  that tells the agent it is done (\"done when every captured value has a\n  destination\"). A vague bound (\"once you understand the code\") invites the\n  agent to finish early, pulled by the steps it can see ahead. Sharpen the\n  bound first; split later steps into another skill or subagent only when the\n  rush persists. Hiding later steps works only across a real context\n  boundary (a subagent, a handoff): an inline Skill call leaves them in view. A demanding criterion (\"every modified model accounted for\")\n  drives more legwork than \"produce a change list\".\n- **Use leading words.** A compact concept the model already knows (_tight_\n  loop, _tracer bullet_, _frontier_, _red_) anchors a whole region of behavior\n  in one token. Repeat the word, not a sentence restating it. Prefer an\n  existing word over a coined one: a coinage recruits no prior knowledge, so\n  you pay for its definition. Look for a triad spelled out at three sites, or\n  a sentence gesturing at one idea, and collapse it into the word.\n- **State the behavior you want.** A prohibition puts the forbidden behavior\n  into context and makes it more available. Write \"write one-line comments\",\n  not \"don't write long comments\". Keep a prohibition only as a hard guardrail\n  with no positive phrasing, and pair it with the target.\n- **Write in a normal register.** Current Claude and GPT models follow\n  instructions closely; CRITICAL, MUST, NEVER, and ALWAYS in capitals make them\n  overtrigger and overapply. Give the rule plus a one-clause reason instead.\n  Reach for stronger wording only as a targeted fix after testing shows the\n  agent skipping that specific rule.\n- **Delete no-ops.** Test each sentence: does it change behavior versus the\n  model's default? If not, delete the whole sentence. A word too weak to beat\n  the default (\"be thorough\") is also a no-op; the fix is a stronger leading\n  word (\"relentless\"), not more words.\n- **Let the environment be the source of truth.** A skill that restates\n  `package.json` scripts, config, or `--help` output is a cache that goes\n  stale. Write down what the agent cannot find by looking: the unwritten\n  convention, the reason behind a choice, the gotcha no config confesses.\n- **Prune sediment.** Stale layers settle because adding feels safe and\n  removing feels risky. Every edit is a chance to delete a line that no longer\n  bears on what the skill does.\n\n- **Shape what the user reads.** In a long workflow, each progress update\n  says where the run stands (\"step 3 of 7 done: CI green\"), and the final\n  report leads with the outcome, then ends on one concrete next action. The\n  user can't hold the run's state between messages; the report holds it for\n  them. The `adhd-mode` skill carries the full set of these rules.\n\nEach meaning lives in one place. Restating a rule in a second skill doubles its\nmaintenance and inflates its weight; point at the owner instead.\n\n## Structure: a default, not a contract\n\nA workflow skill usually wants: what this is for, the constraints that would\ncause real damage if violated, the happy path, then the exceptions. That\nordering is a good default because a reader who stops early still has the\nimportant parts.\n\nAdapt it freely. Named sections beat generic ones — `## Check the budget` tells\na skimmer more than `## STANDARD PATH`. A reference skill may be a table and\nnothing else. A skill with one instruction should be one paragraph.\n\nTwo things genuinely help on complex tasks and are worth including when they\napply:\n\n- **A copyable checklist** for multi-step work, so progress is trackable.\n- **A feedback loop** — run the validator, fix what it reports, repeat until\n  clean. Give Claude something to check its own work against.\n\n**Examples are not bloat.** When output quality depends on matching a shape —\ncommit messages, review comments, generated files — showing an input\u002Foutput\npair is the most token-efficient instruction available.\n\n## Frontmatter\n\n```yaml\n---\nname: skill-name\ndescription: >-\n  What it does and when to use it, third person, with the words a user would\n  actually say.\nallowed-tools: Read, Grep, Bash(git log*)\n---\n```\n\n### The description decides whether the skill ever fires\n\nIt is the only part Claude sees before deciding to load the skill, so it must\ncarry **what it does** and **when to use it**, written in third person, with\nspecific trigger terms a user would actually type. Vague descriptions produce\nskills that never fire; overly narrow ones fire only on exact phrasing.\n\n- Weak: \"Helps with commits.\"\n- Strong: \"Creates conventional commits with quality gates — runs validation,\n  stages, and writes the message. Use when committing changes, or when the user\n  asks to commit, stage, or write a commit message.\"\n\nAn optional `when_to_use` field can carry trigger phrasing separately; keep the\ntwo together under 1,536 characters.\n\n`paths:` scopes auto-activation to matching globs. Measured caveat: a skill\ncarrying `paths:` stops registering as a slash command, so the user loses\n`\u002Fname`. Only worth it for a skill nobody invokes by name.\n\nSkill-execution fields (`context: fork`, `effort:`, `model:`), the\n`allowed-tools` (skills) vs `tools` (agents) distinction, and when to reach for\na subagent at all live in [frontmatter-and-agents.md](frontmatter-and-agents.md).\nRead it before setting any of them: the wrong field name is silently ignored,\nand an agent with no `tools:` key inherits every tool.\n\n## Invocation and composition\n\nEvery skill is one of two kinds, decided by who can start it:\n\n- **Model-invoked** skills hold reusable discipline or reference: how to\n  review, how to write a commit, how to file an issue. The agent can reach for\n  them mid-task, and other skills can call them. The description is\n  model-facing: lead with the job, then the distinct situations that should\n  trigger it, one trigger per situation. It loads on every request, so it is a\n  permanent context cost paid for discoverability.\n- **User-invoked** skills are workflows the human times: grooming a backlog,\n  restructuring issues, generating a logo. Only the human typing `\u002Fname` can\n  start one; no agent and no other skill can. The description is a one-line\n  summary for a person browsing the menu, and costs no context.\n\nThe test is whether the agent could usefully reach for the skill on its own,\nor another skill must. If yes, keep it model-invoked; a skill that previews\nits side effects before acting (`issue-create`, `issue-close`) is safe to\nleave reachable. If it only ever fires by hand, make it user-invoked.\n\nA user-invoked skill moves its cost from context to the human, who becomes\nthe index that must remember it exists. That load is the price of keeping a\nperson in charge, so spend it where their timing matters. When user-invoked\nskills pile up past what anyone remembers, a router skill (`ask-harness`\nhere) names them and when to reach for each, so the human remembers one.\n\nSplit a model-invoked skill out of another only when it has its own trigger\nword you actually use, or another skill must reach it alone; the new\ndescription is a permanent context cost.\n\nA user-invoked skill may call model-invoked skills, never another user-invoked\none. When a step needs a user-invoked skill as a precondition, tell the user to\nrun it rather than calling it.\n\n**Gating an action must not gate the knowledge needed to design for it.** A\nuser-invoked skill is invisible while the agent designs, so guidance needed\nbefore the action (how to pick a print orientation, how to plan a migration)\nlives in a model-invoked skill or reference doc, and the user-invoked skill\nkeeps only the action.\n\n### Declaring it\n\n`agents\u002Fopenai.yaml` beside each `SKILL.md` is the single source of truth.\nUser-invoked skills set:\n\n```yaml\npolicy:\n  allow_implicit_invocation: false\n```\n\nCodex reads that policy directly. The Claude generator derives\n`disable-model-invocation: true` from it, so shared `SKILL.md` frontmatter\nstays free of Claude-only fields. A skill hidden from Codex only because it\nneeds Claude Code (such as `budget`) is listed in the generator's\n`claude_only` set instead. Every `openai.yaml` also carries\n`interface.display_name` and `interface.short_description` for the Codex\npicker. The Claude menu still shows a user-invoked skill;\n`user-invocable: false` is the separate field that hides it there.\n\n### Calling another skill\n\nWrite the call as an instruction to use the Skill tool: `Call the Skill tool\nwith \"commit\"`. Naming the tool fires it far more reliably than a bare `\u002Fcommit`\nin prose, and a bare name carries no host-specific slash syntax. One skill per\ncall: a step that needs two says `Call the Skill tool twice, for \"grilling\" and\n\"domain-modeling\"`. Shared reference lives in the skill that owns it; other\nskills reach it by calling that skill, not by linking into its folder.\n\nRouter prose that only lists skills for a human to pick from (`## Related`\nsections) is not a call and keeps `\u002Fname` labels.\n\n## Splitting across files\n\nSKILL.md is the navigation layer. When it outgrows ~500 lines, move depth into\nsibling files and link to them by name and purpose, so Claude can tell whether\na file is worth opening.\n\nDecide what moves out by **branch**: keep inline what every run needs, and\nput behind a pointer what only some runs reach. Reference left inline buries\nthe steps around it, so they get skipped at random. Push too much out and the\nagent misses material it needed. Whatever stays, **co-locate**: a concept's\ndefinition, rules, and caveats sit under one heading, not scattered through\nthe file. A file too long even when every line is live (**sprawl**) thins the\nagent's attention; split it by branch, or by sequence when later steps tempt\nthe agent to rush the current one.\n\n```text\nskills\u002F\u003Carea>\u002Fmy-skill\u002F\n  SKILL.md          Navigation + the path most runs take\n  reference.md      Full detail, loaded on demand\n  scripts\u002Frun.sh    Executable steps, not pasted into the body\n```\n\nTwo rules make this work:\n\n- **One level deep.** References from SKILL.md only. Nested references get\n  partially read (`head -100`) and yield incomplete information.\n- **Table of contents past 100 lines.** A partial read of a long reference\n  should still reveal its full scope.\n\nReference docs an agent needs live in one skill (the `harness-builder`\ncatalogs sit in the harness skill), and every skill that dispatches the agent\npasses their paths in its prompt. Reference two user-invoked skills both need\ncan live in neither, since neither can call the other: put it in a plain file\noutside the skill system that both point at.\n\n## harness-kit conventions\n\n- **Extract inline bash into a script.** Anything beyond a couple of lines\n  becomes a script — it gets ShellCheck coverage from `just check` and can be\n  tested. A script used by one skill lives in that skill's directory; anything\n  reused across skills is canonical in `scripts\u002Fshared\u002F`.\n- Address skill-local scripts as `\u003Cskill-dir>\u002Fname.sh`. Address scripts reused\n  by multiple skills as `\u003Cshared-scripts-dir>\u002Fname.sh`; they stay canonical in\n  `scripts\u002Fshared\u002F`. The Claude projection adds matching runtime paths to\n  `allowed-tools`. Codex plugin packaging omits skill symlinks, so do not use\n  symlinks as the distribution mechanism.\n- **Never put `eval` in the same command as an allowed script call.** Claude\n  Code cannot statically analyze a command containing `eval`, so no\n  `allowed-tools` rule matches and it prompts on every run. `var=$(script.sh)`\n  and `script.sh --flag arg` both match fine; `cmd=$(script.sh); eval \"$cmd\"`\n  does not. Split it into two calls.\n- Keep each shared script self-contained: the Claude projection copies only\n  the helpers a skill names, so a call from one helper to another breaks\n  there (`just check` catches a `$SCRIPT_DIR\u002F` call with nothing beside it).\n- Put each canonical skill in its topic-area folder,\n  `skills\u002F\u003Carea>\u002F\u003Cskill>\u002F` (`engineering`, `meta`, `orchestration`,\n  `project-management`, `thinking`, `writing`); names stay unique across areas\n  because the Claude projection is flat. Register canonical skills through\n  `.codex-plugin\u002Fplugin.json`, regenerate\n  `claude-skills\u002F`, and register the projection in\n  `.claude-plugin\u002Fplugin.json`. Claude agents use the Claude manifest; Codex\n  agents are project TOMLs installed by bootstrap.\n- **Run `just check`** before committing — plugin validation, hook schema\n  checks, ShellCheck, markdownlint, and whitespace hygiene.\n- **Skill directory name must match the frontmatter `name`.**\n\n## Evaluate before you elaborate\n\nBefore writing extensive instructions, write three concrete tasks the skill\nshould handle and run them **without** the skill. That baseline tells you what\nClaude already does well — which is exactly what the skill should not repeat.\nThen add only the instructions that move a failing case, and re-run.\n\nTest across model tiers when the skill will be used by more than one: what\nreads as over-explaining to Opus can be the necessary detail for Haiku.\n\n## Anti-patterns\n\n- Offering several options with no default. Pick one; mention alternatives only\n  if the choice is genuinely situational.\n- Time-sensitive content (\"as of the March release\", \"the new API\"). If old\n  behavior must be documented, put it in an `\u003Cdetails>` block labelled as\n  historical.\n- Inconsistent terminology — one name per concept, throughout.\n- Windows-style paths.\n- Restating general good practice Claude already follows.\n\n## Templates\n\nCopyable skill and agent skeletons live in [templates.md](templates.md).\n\n## Related\n\n- `\u002Fdocumentation-writer` — human-facing docs (README, guides, reference)\n- `\u002Fharness audit` — audits skills and agents in a project\n- The harness skill's `hooks-reference.md` — hooks.json schema, events, gotchas\n\n## Sources\n\n- [Skill authoring best practices](https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fagents-and-tools\u002Fagent-skills\u002Fbest-practices)\n- [Claude Code skills](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fskills)\n- [Prompting Claude Fable 5](https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fbuild-with-claude\u002Fprompt-engineering\u002Fprompting-claude-fable-5)\n- [Prompting best practices](https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fbuild-with-claude\u002Fprompt-engineering\u002Fclaude-prompting-best-practices) (dial back aggressive emphasis)\n- [skill-creator](https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fskill-creator\u002FSKILL.md)\n- [GPT-5 prompting guide](https:\u002F\u002Fdevelopers.openai.com\u002Fcookbook\u002Fexamples\u002Fgpt-5\u002Fgpt-5_prompting_guide)\n- [Agent Skills specification](https:\u002F\u002Fagentskills.io\u002Fspecification)\n- [mattpocock\u002Fskills `writing-for-agents`](https:\u002F\u002Fgithub.com\u002Fmattpocock\u002Fskills\u002Ftree\u002Fmain\u002Fskills\u002Fproductivity\u002Fwriting-for-agents) (MIT): context pointers, completion criteria, leading words, negation, no-ops, splitting by branch, and the invocation split; see `CREDITS.md`\n","# \u002Fskill-writer — Authoring Skills and Agents\n\nClaude already knows the skill format. This skill exists for the judgment calls\nthe format does not make for you: how prescriptive to be, what earns a place in\na file that stays in context all session, and the conventions specific to this\nrepo.\n\n## First, decide whether to write one at all\n\nA skill is worth writing when the task recurs, and when Claude's default\nbehavior on it is wrong or under-informed. If default performance is already\ngood, a skill that restates the obvious makes output worse, not better — it\nspends context and over-constrains a capable model.\n\nPrefer the smallest thing that works:\n\n| Need | Reach for |\n| --- | --- |\n| A fixed sequence of commands | A script in `scripts\u002Fshared\u002F` |\n| Project facts Claude must always know | `CLAUDE.md` |\n| A repeatable workflow, or discipline the agent should reach for mid-task | A skill |\n| Wide reads\u002Fsearches whose output should stay out of the main context | An agent |\n| Something that must happen automatically, every time | A hook |\n\n## Match specificity to fragility\n\nThe central authoring decision is **degrees of freedom** — how much latitude to\nleave. Set it per-instruction, based on how fragile the operation is, not by\napplying one uniform level to a whole file.\n\n- **High freedom** — prose, heuristics, a stated goal. Use when several\n  approaches are valid and the right one depends on context. \"Summarize what\n  changed and why, in the reviewer's terms.\"\n- **Medium freedom** — a preferred pattern with room to vary: pseudocode, a\n  parameterized script, a worked example to adapt.\n- **Low freedom** — exact commands, explicit ordering, \"do not add flags to\n  this command.\" Use when the operation is fragile, when consistency across\n  runs matters, or when a specific sequence must hold.\n\nThe test: a narrow bridge with cliffs on both sides gets guardrails; an open\nfield gets a direction to walk in. Over-guardrailing an open field is the more\ncommon failure here — it reads as thorough and degrades results.\n\nCorollary: skills written for older models are often too prescriptive for\ncurrent ones. When editing an existing skill, actively look for instructions to\n**delete**, and check whether default behavior is already better than what the\nskill mandates.\n\n## Every line is a recurring cost\n\nOnce a skill loads, its content stays in context for the rest of the session.\nAfter auto-compaction, Claude Code re-attaches only the **first 5,000 tokens**\nof each invoked skill, inside a **25,000-token** combined budget filled\nmost-recent-first — an oversized skill gets truncated mid-file and pushes older\nskills out entirely.\n\nSo: state what to do, with the reason in a clause rather than a narrative.\nChallenge each line — does it justify its cost? Assume the reader is smart.\n\nThe real limits, and the only ones worth quoting:\n\n| Thing | Limit |\n| --- | --- |\n| SKILL.md body | Under 500 lines — split into sibling files past that |\n| `description` | 1,024 characters |\n| `description` + `when_to_use` combined | 1,536 characters in the skill listing |\n| Post-compaction re-attachment | First 5,000 tokens per skill, 25,000 total |\n\nThere is no prescribed section schema and no per-section token budget. Use the\nheadings the content actually needs.\n\n## Context pointers\n\nA **context pointer** is an always-loaded line that names material outside\nthe context and says when to reach for it: a skill's description, or an\n`AGENTS.md` \u002F `CLAUDE.md` line naming a doc. Its wording, not its target,\ndecides how reliably the agent reaches the material. When must-have material\nbehind a pointer gets missed, sharpen the pointer first; inline the material\nonly if that fails.\n\nA pointer states what the material is and lists its **branches**, the\ndistinct cases that should send the agent there. It costs tokens on every\nturn, so prune it harder than any body:\n\n- Front-load the leading word; that is where the triggering happens.\n- One trigger per branch. Synonyms for one branch are that branch written\n  twice.\n- Cut anything the target already says about itself.\n\n## Writing the body\n\nThe house style is short plain prose. These levers decide whether an agent\ntakes the same path through a skill every run.\n\n- **Give each step a completion criterion.** (A mode skill, which shapes\n  every response until switched off, has no steps; its exit phrase is its\n  criterion.) End every step on the condition\n  that tells the agent it is done (\"done when every captured value has a\n  destination\"). A vague bound (\"once you understand the code\") invites the\n  agent to finish early, pulled by the steps it can see ahead. Sharpen the\n  bound first; split later steps into another skill or subagent only when the\n  rush persists. Hiding later steps works only across a real context\n  boundary (a subagent, a handoff): an inline Skill call leaves them in view. A demanding criterion (\"every modified model accounted for\")\n  drives more legwork than \"produce a change list\".\n- **Use leading words.** A compact concept the model already knows (_tight_\n  loop, _tracer bullet_, _frontier_, _red_) anchors a whole region of behavior\n  in one token. Repeat the word, not a sentence restating it. Prefer an\n  existing word over a coined one: a coinage recruits no prior knowledge, so\n  you pay for its definition. Look for a triad spelled out at three sites, or\n  a sentence gesturing at one idea, and collapse it into the word.\n- **State the behavior you want.** A prohibition puts the forbidden behavior\n  into context and makes it more available. Write \"write one-line comments\",\n  not \"don't write long comments\". Keep a prohibition only as a hard guardrail\n  with no positive phrasing, and pair it with the target.\n- **Write in a normal register.** Current Claude and GPT models follow\n  instructions closely; CRITICAL, MUST, NEVER, and ALWAYS in capitals make them\n  overtrigger and overapply. Give the rule plus a one-clause reason instead.\n  Reach for stronger wording only as a targeted fix after testing shows the\n  agent skipping that specific rule.\n- **Delete no-ops.** Test each sentence: does it change behavior versus the\n  model's default? If not, delete the whole sentence. A word too weak to beat\n  the default (\"be thorough\") is also a no-op; the fix is a stronger leading\n  word (\"relentless\"), not more words.\n- **Let the environment be the source of truth.** A skill that restates\n  `package.json` scripts, config, or `--help` output is a cache that goes\n  stale. Write down what the agent cannot find by looking: the unwritten\n  convention, the reason behind a choice, the gotcha no config confesses.\n- **Prune sediment.** Stale layers settle because adding feels safe and\n  removing feels risky. Every edit is a chance to delete a line that no longer\n  bears on what the skill does.\n\n- **Shape what the user reads.** In a long workflow, each progress update\n  says where the run stands (\"step 3 of 7 done: CI green\"), and the final\n  report leads with the outcome, then ends on one concrete next action. The\n  user can't hold the run's state between messages; the report holds it for\n  them. The `adhd-mode` skill carries the full set of these rules.\n\nEach meaning lives in one place. Restating a rule in a second skill doubles its\nmaintenance and inflates its weight; point at the owner instead.\n\n## Structure: a default, not a contract\n\nA workflow skill usually wants: what this is for, the constraints that would\ncause real damage if violated, the happy path, then the exceptions. That\nordering is a good default because a reader who stops early still has the\nimportant parts.\n\nAdapt it freely. Named sections beat generic ones — `## Check the budget` tells\na skimmer more than `## STANDARD PATH`. A reference skill may be a table and\nnothing else. A skill with one instruction should be one paragraph.\n\nTwo things genuinely help on complex tasks and are worth including when they\napply:\n\n- **A copyable checklist** for multi-step work, so progress is trackable.\n- **A feedback loop** — run the validator, fix what it reports, repeat until\n  clean. Give Claude something to check its own work against.\n\n**Examples are not bloat.** When output quality depends on matching a shape —\ncommit messages, review comments, generated files — showing an input\u002Foutput\npair is the most token-efficient instruction available.\n\n## Frontmatter\n\n```yaml\n---\nname: skill-name\ndescription: >-\n  What it does and when to use it, third person, with the words a user would\n  actually say.\nallowed-tools: Read, Grep, Bash(git log*)\n---\n```\n\n### The description decides whether the skill ever fires\n\nIt is the only part Claude sees before deciding to load the skill, so it must\ncarry **what it does** and **when to use it**, written in third person, with\nspecific trigger terms a user would actually type. Vague descriptions produce\nskills that never fire; overly narrow ones fire only on exact phrasing.\n\n- Weak: \"Helps with commits.\"\n- Strong: \"Creates conventional commits with quality gates — runs validation,\n  stages, and writes the message. Use when committing changes, or when the user\n  asks to commit, stage, or write a commit message.\"\n\nAn optional `when_to_use` field can carry trigger phrasing separately; keep the\ntwo together under 1,536 characters.\n\n`paths:` scopes auto-activation to matching globs. Measured caveat: a skill\ncarrying `paths:` stops registering as a slash command, so the user loses\n`\u002Fname`. Only worth it for a skill nobody invokes by name.\n\nSkill-execution fields (`context: fork`, `effort:`, `model:`), the\n`allowed-tools` (skills) vs `tools` (agents) distinction, and when to reach for\na subagent at all live in [frontmatter-and-agents.md](frontmatter-and-agents.md).\nRead it before setting any of them: the wrong field name is silently ignored,\nand an agent with no `tools:` key inherits every tool.\n\n## Invocation and composition\n\nEvery skill is one of two kinds, decided by who can start it:\n\n- **Model-invoked** skills hold reusable discipline or reference: how to\n  review, how to write a commit, how to file an issue. The agent can reach for\n  them mid-task, and other skills can call them. The description is\n  model-facing: lead with the job, then the distinct situations that should\n  trigger it, one trigger per situation. It loads on every request, so it is a\n  permanent context cost paid for discoverability.\n- **User-invoked** skills are workflows the human times: grooming a backlog,\n  restructuring issues, generating a logo. Only the human typing `\u002Fname` can\n  start one; no agent and no other skill can. The description is a one-line\n  summary for a person browsing the menu, and costs no context.\n\nThe test is whether the agent could usefully reach for the skill on its own,\nor another skill must. If yes, keep it model-invoked; a skill that previews\nits side effects before acting (`issue-create`, `issue-close`) is safe to\nleave reachable. If it only ever fires by hand, make it user-invoked.\n\nA user-invoked skill moves its cost from context to the human, who becomes\nthe index that must remember it exists. That load is the price of keeping a\nperson in charge, so spend it where their timing matters. When user-invoked\nskills pile up past what anyone remembers, a router skill (`ask-harness`\nhere) names them and when to reach for each, so the human remembers one.\n\nSplit a model-invoked skill out of another only when it has its own trigger\nword you actually use, or another skill must reach it alone; the new\ndescription is a permanent context cost.\n\nA user-invoked skill may call model-invoked skills, never another user-invoked\none. When a step needs a user-invoked skill as a precondition, tell the user to\nrun it rather than calling it.\n\n**Gating an action must not gate the knowledge needed to design for it.** A\nuser-invoked skill is invisible while the agent designs, so guidance needed\nbefore the action (how to pick a print orientation, how to plan a migration)\nlives in a model-invoked skill or reference doc, and the user-invoked skill\nkeeps only the action.\n\n### Declaring it\n\n`agents\u002Fopenai.yaml` beside each `SKILL.md` is the single source of truth.\nUser-invoked skills set:\n\n```yaml\npolicy:\n  allow_implicit_invocation: false\n```\n\nCodex reads that policy directly. The Claude generator derives\n`disable-model-invocation: true` from it, so shared `SKILL.md` frontmatter\nstays free of Claude-only fields. A skill hidden from Codex only because it\nneeds Claude Code (such as `budget`) is listed in the generator's\n`claude_only` set instead. Every `openai.yaml` also carries\n`interface.display_name` and `interface.short_description` for the Codex\npicker. The Claude menu still shows a user-invoked skill;\n`user-invocable: false` is the separate field that hides it there.\n\n### Calling another skill\n\nWrite the call as an instruction to use the Skill tool: `Call the Skill tool\nwith \"commit\"`. Naming the tool fires it far more reliably than a bare `\u002Fcommit`\nin prose, and a bare name carries no host-specific slash syntax. One skill per\ncall: a step that needs two says `Call the Skill tool twice, for \"grilling\" and\n\"domain-modeling\"`. Shared reference lives in the skill that owns it; other\nskills reach it by calling that skill, not by linking into its folder.\n\nRouter prose that only lists skills for a human to pick from (`## Related`\nsections) is not a call and keeps `\u002Fname` labels.\n\n## Splitting across files\n\nSKILL.md is the navigation layer. When it outgrows ~500 lines, move depth into\nsibling files and link to them by name and purpose, so Claude can tell whether\na file is worth opening.\n\nDecide what moves out by **branch**: keep inline what every run needs, and\nput behind a pointer what only some runs reach. Reference left inline buries\nthe steps around it, so they get skipped at random. Push too much out and the\nagent misses material it needed. Whatever stays, **co-locate**: a concept's\ndefinition, rules, and caveats sit under one heading, not scattered through\nthe file. A file too long even when every line is live (**sprawl**) thins the\nagent's attention; split it by branch, or by sequence when later steps tempt\nthe agent to rush the current one.\n\n```text\nskills\u002F\u003Carea>\u002Fmy-skill\u002F\n  SKILL.md          Navigation + the path most runs take\n  reference.md      Full detail, loaded on demand\n  scripts\u002Frun.sh    Executable steps, not pasted into the body\n```\n\nTwo rules make this work:\n\n- **One level deep.** References from SKILL.md only. Nested references get\n  partially read (`head -100`) and yield incomplete information.\n- **Table of contents past 100 lines.** A partial read of a long reference\n  should still reveal its full scope.\n\nReference docs an agent needs live in one skill (the `harness-builder`\ncatalogs sit in the harness skill), and every skill that dispatches the agent\npasses their paths in its prompt. Reference two user-invoked skills both need\ncan live in neither, since neither can call the other: put it in a plain file\noutside the skill system that both point at.\n\n## harness-kit conventions\n\n- **Extract inline bash into a script.** Anything beyond a couple of lines\n  becomes a script — it gets ShellCheck coverage from `just check` and can be\n  tested. A script used by one skill lives in that skill's directory; anything\n  reused across skills is canonical in `scripts\u002Fshared\u002F`.\n- Address skill-local scripts as `\u003Cskill-dir>\u002Fname.sh`. Address scripts reused\n  by multiple skills as `\u003Cshared-scripts-dir>\u002Fname.sh`; they stay canonical in\n  `scripts\u002Fshared\u002F`. The Claude projection adds matching runtime paths to\n  `allowed-tools`. Codex plugin packaging omits skill symlinks, so do not use\n  symlinks as the distribution mechanism.\n- **Never put `eval` in the same command as an allowed script call.** Claude\n  Code cannot statically analyze a command containing `eval`, so no\n  `allowed-tools` rule matches and it prompts on every run. `var=$(script.sh)`\n  and `script.sh --flag arg` both match fine; `cmd=$(script.sh); eval \"$cmd\"`\n  does not. Split it into two calls.\n- Keep each shared script self-contained: the Claude projection copies only\n  the helpers a skill names, so a call from one helper to another breaks\n  there (`just check` catches a `$SCRIPT_DIR\u002F` call with nothing beside it).\n- Put each canonical skill in its topic-area folder,\n  `skills\u002F\u003Carea>\u002F\u003Cskill>\u002F` (`engineering`, `meta`, `orchestration`,\n  `project-management`, `thinking`, `writing`); names stay unique across areas\n  because the Claude projection is flat. Register canonical skills through\n  `.codex-plugin\u002Fplugin.json`, regenerate\n  `claude-skills\u002F`, and register the projection in\n  `.claude-plugin\u002Fplugin.json`. Claude agents use the Claude manifest; Codex\n  agents are project TOMLs installed by bootstrap.\n- **Run `just check`** before committing — plugin validation, hook schema\n  checks, ShellCheck, markdownlint, and whitespace hygiene.\n- **Skill directory name must match the frontmatter `name`.**\n\n## Evaluate before you elaborate\n\nBefore writing extensive instructions, write three concrete tasks the skill\nshould handle and run them **without** the skill. That baseline tells you what\nClaude already does well — which is exactly what the skill should not repeat.\nThen add only the instructions that move a failing case, and re-run.\n\nTest across model tiers when the skill will be used by more than one: what\nreads as over-explaining to Opus can be the necessary detail for Haiku.\n\n## Anti-patterns\n\n- Offering several options with no default. Pick one; mention alternatives only\n  if the choice is genuinely situational.\n- Time-sensitive content (\"as of the March release\", \"the new API\"). If old\n  behavior must be documented, put it in an `\u003Cdetails>` block labelled as\n  historical.\n- Inconsistent terminology — one name per concept, throughout.\n- Windows-style paths.\n- Restating general good practice Claude already follows.\n\n## Templates\n\nCopyable skill and agent skeletons live in [templates.md](templates.md).\n\n## Related\n\n- `\u002Fdocumentation-writer` — human-facing docs (README, guides, reference)\n- `\u002Fharness audit` — audits skills and agents in a project\n- The harness skill's `hooks-reference.md` — hooks.json schema, events, gotchas\n\n## Sources\n\n- [Skill authoring best practices](https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fagents-and-tools\u002Fagent-skills\u002Fbest-practices)\n- [Claude Code skills](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fskills)\n- [Prompting Claude Fable 5](https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fbuild-with-claude\u002Fprompt-engineering\u002Fprompting-claude-fable-5)\n- [Prompting best practices](https:\u002F\u002Fplatform.claude.com\u002Fdocs\u002Fen\u002Fbuild-with-claude\u002Fprompt-engineering\u002Fclaude-prompting-best-practices) (dial back aggressive emphasis)\n- [skill-creator](https:\u002F\u002Fgithub.com\u002Fanthropics\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fskill-creator\u002FSKILL.md)\n- [GPT-5 prompting guide](https:\u002F\u002Fdevelopers.openai.com\u002Fcookbook\u002Fexamples\u002Fgpt-5\u002Fgpt-5_prompting_guide)\n- [Agent Skills specification](https:\u002F\u002Fagentskills.io\u002Fspecification)\n- [mattpocock\u002Fskills `writing-for-agents`](https:\u002F\u002Fgithub.com\u002Fmattpocock\u002Fskills\u002Ftree\u002Fmain\u002Fskills\u002Fproductivity\u002Fwriting-for-agents) (MIT): context pointers, completion criteria, leading words, negation, no-ops, splitting by branch, and the invocation split; see `CREDITS.md`\n",{"name":5,"description":7,"allowed-tools":57},"Glob, Grep, Read, Write, Edit","main","023b01bfb4a86fbf1e38a749f1044cce3e5da67d","skills\u002Fmeta\u002Fskill-writer",1,"https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit","https:\u002F\u002Fgithub.com\u002Fdougborg","http:\u002F\u002Fdougborg.org","Platforms and Infrastructure Software Engineer","2026-11-13 07:43:37",{"source":8,"skillName":5,"commands":68,"prompt":78,"zipAvailable":79},[69,72,75],{"tool":70,"command":71},"npx","npx skills add dougborg\u002Fharness-kit --skill skill-writer",{"tool":73,"command":74},"bunx","bunx skills add dougborg\u002Fharness-kit --skill skill-writer",{"tool":76,"command":77},"pnpm","pnpm dlx skills add dougborg\u002Fharness-kit --skill skill-writer","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fmeta\u002Fskill-writer\u002FSKILL.md",true,[81,87,91],{"path":82,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":85,"sourceHtmlUrl":86},"agents\u002Fopenai.yaml","REFERENCE","参考文档",120,"https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fmeta\u002Fskill-writer\u002Fagents\u002Fopenai.yaml",{"path":88,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":89,"sourceHtmlUrl":90},"frontmatter-and-agents.md",4053,"https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fmeta\u002Fskill-writer\u002Ffrontmatter-and-agents.md",{"path":92,"resourceType":83,"resourceTypeDesc":84,"sizeBytes":93,"sourceHtmlUrl":94},"templates.md",1341,"https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fmeta\u002Fskill-writer\u002Ftemplates.md",[96,113,130,155,178,195],{"skillNo":97,"name":98,"nameZh":6,"description":99,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":100,"categories":101,"keywords":105,"publishedAt":52},"01M6KX4J65S5FZKMSN5DQNPSZM","teach","Teach yourself a topic over several sessions, with lessons, references, and a record of what you've learned.","https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fwriting\u002Fteach\u002FSKILL.md",[102],{"categoryNo":103,"categoryName":104,"taxonomyType":29,"taxonomyTypeDesc":30},"01JBSEEDCAT000000000000039","教师与教育",[98,106,107,108,109,110,111,112],"lessons","you","with","learned","record","references","sessions",{"skillNo":114,"name":115,"nameZh":6,"description":116,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":117,"categories":118,"keywords":122,"publishedAt":52},"01M6KX4FSSG7MHF00JDT5QT0BQ","svg-logo-designer","Generate SVG logos with multiple concepts, layouts, and color variations.","https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fwriting\u002Fsvg-logo-designer\u002FSKILL.md",[119],{"categoryNo":120,"categoryName":121,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000016","图像与多媒体",[123,124,125,126,127,128,108,45,46,129],"logo","designer","svg logo designer","SVG","color","concepts","generate",{"skillNo":131,"name":132,"nameZh":6,"description":133,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":134,"categories":135,"keywords":138,"publishedAt":52},"01M6KX4BD3J9MZPPSH4MENH7FP","documentation-writer","Writes and revises human-facing documentation — READMEs, contributor and setup guides, reference pages, architecture notes, CLAUDE.md, and changelog entries. Covers picking the right document type for the question being asked, structuring for skimming, writing examples that actually run, and keeping docs from going stale. Use when creating or rewriting docs, when a README has drifted from the code, or when deciding where a piece of information belongs. For SKILL.md and agent files, use \u002Fskill-writer instead.","https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fwriting\u002Fdocumentation-writer\u002FSKILL.md",[136,137],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":27,"categoryName":28,"taxonomyType":29,"taxonomyTypeDesc":30},[33,139,140,141,142,143,144,34,145,146,147,148,51,149,150,151,152,153,154,43],"documentation writer","documentation","entries","changelog","architecture","document","use","question","README","notes","examples","pages","for","information","guides","type",{"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":52},"01M6KX47ZJ4ZWHDN3MY8SYAWZZ","adhd-mode","Shapes every response for a reader with ADHD: the next action first, numbered steps, progress restated each turn, tangents held back, concrete time estimates, wins made visible. Stays on for the rest of the session until the reader says \"stop adhd mode\". On Claude Code, the harness-kit:adhd output style does the same without invoking this skill.","https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fwriting\u002Fadhd-mode\u002FSKILL.md",[],[162,163,164,165,151,166,36,108,167,168,169,170,171,172,173,174,175,176,177,51],"adhd","mode","adhd mode","next","action","output","style","progress","time","back","stop","session","reader","first","REST","harness",{"skillNo":179,"name":180,"nameZh":6,"description":181,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":182,"categories":183,"keywords":184,"publishedAt":52},"01M6KX463Z6E18EPDW873GDHKB","wait-what","That last message didn't land: re-explain it with the missing context.","https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fthinking\u002Fwait-what\u002FSKILL.md",[],[185,186,187,108,188,189,190,191,192,193,194],"wait","wait what","re","it","missing","message","context","explain","land","last",{"skillNo":196,"name":197,"nameZh":6,"description":198,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":199,"categories":200,"keywords":201,"publishedAt":52},"01M6KX43NE1NB0Y2HECY515730","to-questionnaire","Turn a decision you can't make alone into a questionnaire for the one person who can answer it, filled in async or together in a meeting.","https:\u002F\u002Fgithub.com\u002Fdougborg\u002Fharness-kit\u002Fblob\u002Fmain\u002Fskills\u002Fthinking\u002Fto-questionnaire\u002FSKILL.md",[],[202,203,151,204,205,206,207,107,208,188,209,210],"to questionnaire","questionnaire","person","can","answer","meeting","make","async","decision"]