技能目录
这里列出的技能来自检查入库的活动清单，并在构建时读取各自的结构化 frontmatter。目录帮助选择方法，不改变技能本身。
01 · principle-cultivation
Section titled “01 · principle-cultivation”Steward a project’s fallible principle lineage: preserve source-bound research, turn only gate-passing inquiries into candidates, convene selective P-ID reviews, clarify interpretations, trial human-nominated alternates, and prepare human-only Sequence decisions. Use for reusable-principle research, candidate proposals, principle reviews, interpretation drift, alternate trials, or maintaining a principle-centered skill collection.
02 · context-engineering
Section titled “02 · context-engineering”Design, audit, and verify how decision-relevant project information reaches an agent through the actual runtime. Use when an agent misses a governing boundary, receives irrelevant context, cannot discover an instruction or skill, or when deciding what should be always available, activated for a task, or retrieved on demand. Do not use to author domain knowledge, organize its source artifacts, write a one-off prompt, configure vendor hooks, or build an agent runtime.
03 · improve-agent-workflow
Section titled “03 · improve-agent-workflow”Diagnose and improve how coding agents work in an existing project: project skills, AGENTS instructions, prompt and context delivery, tools or CLI surfaces, hooks, verification, handoffs, and agent-facing workflow. Use when an agent repeatedly misses scope or instructions, the owning agent-facing surface is unclear, several surfaces interact, agent experience is costly or confusing, or a user asks Codex to improve how agents work in a company repository. Start from observed project evidence, change the smallest owning surface, and verify behavior through the ordinary agent entry path. If the issue is already localized to a skill’s trigger, prompt, context layering, or behavior evaluation and a dedicated skill-engineering method is available, use that method directly. Do not use for ordinary product features, generic CI or business workflows, one-off prompt wording, or a process redesign with no observed agent-action gap.
04 · skill-engineering
Section titled “04 · skill-engineering”Design, rewrite, review, and test agent skills that improve a repeated agent action or judgment. Use when creating a skill, rewriting a SKILL.md, improving a known skill trigger or action layer, reducing skill-context bloat, or checking whether a skill instruction actually changes behavior. Triggers include “create a skill”, “rewrite this skill”, “review this SKILL.md”, “improve this project skill”, “skill design”, and “test this skill prompt”. If an agent-work failure has not yet been localized to a skill, let an agent-workflow diagnostic method identify the owning surface first.
05 · disciplined-development
Section titled “05 · disciplined-development”Lightweight behavioral discipline for any development task — ground decisions in evidence, choose the smallest valid change, test before claiming done, and analyze the specific situation before applying a pattern. Triggers on “be more disciplined”, “professional habits”, “stop overengineering”, “check your work”, “don’t guess”, “bad habits”, “test strategy”, “meaningful tests”, or “avoid test bloat”; also when an agent shows overconfidence, untested claims, or repetitive overengineering across tasks.
06 · practice-cycle
Section titled “06 · practice-cycle”Move a non-trivial task through practice, observation, reflection, and the next smallest test without forcing a generic plan or feedback artifact. Use when a phase ends, an implementation or investigation must inform its next move, a prior attempt failed or produced ambiguous evidence, work is being resumed, or the user asks to turn results into the next bounded action. Do not use as a mandatory preflight for routine one-step work or when a more specific methodology already owns the live judgment.
07 · form-guidance
Section titled “07 · form-guidance”Decide the smallest truthful form for a repeated problem before designing its implementation: no new artifact, local instruction, durable decision record, skill, runtime, projection, or bounded campaign. Use when asking “should this be a skill?”, “what form should this take?”, “how should this capability exist?”, “is this another agent/runtime?”, “should we add a command or document?”, or when a generic workflow would hide an authority, inheritance, or recurrence decision. Do not use for ordinary implementation after the form is already accepted.
08 · naming-and-articulation
Section titled “08 · naming-and-articulation”Choose and place a project, product, or concept name with its smallest useful definition and explanation. Use when naming a project or shared concept, defining terminology, clarifying status words, deciding where documentation belongs, or asking “what should we call this?” / “how should this be explained?” / “命名” / “术语” / “定义” / “说明”. Do not use for branding, ordinary copyediting, one-off prose, or a Sequence change.
09 · work-estimation
Section titled “09 · work-estimation”Estimate the necessary work before converting it into agent tokens, time, or money. Use when comparing implementation options, setting an agent budget, choosing estimate granularity or error tolerance, planning a discovery probe, or asking “how much work is this really?” / “预算怎么估？” / “工作量评估” / “误差容忍度”. Do not use to invent a token price, approve spend, or replace strategic choice, execution, or calibration evidence.
10 · model-evaluation
Section titled “10 · model-evaluation”Build or revise an evidence-linked capability profile for a model execution setup by running repeated, representative real tasks under matched conditions. Use when comparing models, providers, plans, coding harnesses, or prompt/tool profiles for task allocation; when asking “which model is good enough for this work?”, “evaluate this model”, “compare model capability”, “模型评测/能力画像/模型适合什么任务”, or whether a characterized setup may have degraded. Do not use for provider setup, public leaderboard summaries, one-off response review, automatic routing, or a degradation verdict without an accepted baseline.
11 · task-shaping
Section titled “11 · task-shaping”Judge a consequential task against a conservative Flash-class reference execution profile, then choose direct execution, guarded verification, or a transformation of oversized, branchy, long-context, weakly observable, or completeness-sensitive work into locally verifiable units that can be reconstructed without losing global constraints. Use when deciding one agent versus bounded Cells, an agent loop varies or misses obligations, a verifier skips claims despite seeing evidence, a task needs semantic decomposition before orchestration, a harness needs a dependable task primitive, or asking “can an agent reliably do this?”, “how should this task be split?” / “任务怎么拆才能稳定” / “稳定能力包络”. Do not use for ordinary one-step planning, proving a model capability without repeated evidence, choosing provider/concurrency/budget, executing a prepared queue, or replacing a domain skill that already owns the semantic partition.
12 · systems-engineering
Section titled “12 · systems-engineering”Design or revise a sufficiently reliable whole from fallible human, Agent, software, and organizational parts under concrete constraints and accepted residual risk. Use when a workflow keeps failing despite locally reasonable tasks, when deciding where verification, redundancy, retries, rollback, budget margin, or human acceptance belong, when a Swarm or multi-stage Agent process needs an end-to-end reliability model, when component quality is being confused with system quality, or when asking “how can unreliable agents form a reliable system?”, “where should the feedback loop close?”, “怎么把会犯错的 Agent 组成可靠系统?” / “如何从流程上保证近似可靠?”. Do not use for an ordinary bounded task, generic planning, proving model capability, converting work into token/time/cost estimates, or operating an already prepared queue.
13 · strategic-advisory
Section titled “13 · strategic-advisory”Prepare an evidence-backed strategic recommendation when a project phase has settled, material conditions changed, or a human asks for short-, medium-, and long-horizon direction. Synthesizes verified phase evidence into a proposed Strategy Case with a principal contradiction, strongest preservation case, rolling horizons, mission candidates, and replan signals. It advises only: humans approve commitments, and Work Cells or other owners execute them. Do not use for ordinary task planning, implementation sequencing, or automatic portfolio management.
14 · artifact-organization
Section titled “14 · artifact-organization”Audit whether a project’s artifact layout still expresses its accepted design, then apply one smallest safe organization transition when a material gap exists. Use for “organize the repo”, “组织架构”, “根据地”, stale project structure, source/projection confusion, or campaign residue. Not for ordinary file cleanup, system-module design, Sequence change, or skill authoring.
15 · structural-refactoring
Section titled “15 · structural-refactoring”Analyze, plan, execute, and verify behavior-preserving structural code changes. Use when asked to refactor code across meaningful boundaries: split a large file or module, extract responsibilities, untangle dependencies, reorganize a package, migrate callers, or assess the impact of a consequential refactor. Do not use for routine local cleanup, feature work that intentionally changes behavior, project artifact organization, or an unsettled architecture choice.
16 · visual-design
Section titled “16 · visual-design”Design, redesign, or review real interfaces, documents, illustrations, and related visual systems from their content, audience, intended action, and existing project direction. Use when creating UI, improving visual hierarchy or coherence, reviewing a rendered artifact, or when “the page works but feels generic/inconsistent”. Also use to establish a continuing project’s visual direction from selected references when none exists, or to turn an ambiguous style cue into provisional guidance for later design. Also use when an accepted style or direction must be applied consistently to real code without reopening aesthetic selection. Triggers include “design this interface”, “redesign this page”, “review the UI”, “make these pages feel unified”, “the direction is right but the components feel generic”, “polish/refine the interface details”, “align this logo or illustration with the product”, “apply the selected style”, “art direction”, and “审美/视觉设计/组件细节/插画或品牌视觉统一/确立风格/按现有风格落实”. Do not use for a purely functional code fix, naming task, or automatic aesthetic acceptance. Never impose a portable fixed style.
17 · code-review
Section titled “17 · code-review”Review a proposed code change, staged diff, commit, patch, or pull request for decision-changing defects before acceptance or merge. Use when asked “review this code”, “review the diff/PR”, “can this merge”, “find regressions”, or when a consequential implementation needs an impact-aware independent review. Do not use to implement the change, process already-filed PR comments, settle an unaccepted architecture, review visual design, or orchestrate reviewers.
18 · project-cognition
Section titled “18 · project-cognition”Build, verify, or selectively refresh a source-linked working model of a substantial project when later agents or tasks need to reuse its architecture, responsibilities, authority, state, causal, evidence, or change relations. Use for “initialize project cognition”, “map this project for future work”, “建立项目认知”, “更新项目认知”, “what existing cognition is invalidated by this change?”, or when repeated tasks keep reconstructing the same project model. Do not use for a one-off code review, ordinary repository orientation, context-delivery configuration, or as a mandatory project setup step.
19 · agent-environment
Section titled “19 · agent-environment”Use this Skill, not a named task Skill, when that Skill or capability is the object of user-level setup, installation, migration, or configuration. An installed or runtime-discovered copy is target evidence, not a desired source. If the human has not supplied the selected capability’s source, return NEEDS_INPUT before reading it, inventory, lookup, planning, or action; never inspect secret values. Use for “set up my agents on this machine”, “migrate my AI coding setup”, “新设备配置 Codex/Cursor/Claude Code”, “update my global agent workflow”, or reconstructing selected skills, harness guidance, plugins, MCP, hooks, permissions, runtimes, and authentication prerequisites. Setup does not imply marketplace discovery, CLI/provider installation, or full-toolbox adoption. Do not use for project-local agent workflow design, fleet policy, sessions/caches, or unrelated dotfiles.