petfish.ai

胖鱼PEtFiSh常驻伙伴：感知需求与能力缺口（Tier1领域映射+Tier2意图检测）、 查询已装pack状态、自动检查更新、推荐安装/升级并提供/petfish命令入口。Use when users ask /petfish, /petfish upgrade, /petfish uninstall, "what skills do I have", "what else can you do", "check for updates", "uninstall pack", need deploy/course/ppt/testdocs/petfish/calibrate/ context/research capabilities, or need cross-marketplace skill/MCP discovery and skill ecosystem governance.

Search/discover skills and MCP servers across PEtFiSh, PEtFiSh Market (community), Glama, Smithery, SkillKit, anthropics/skills, and GitHub. Use for /petfish search, “find a skill for…”, “search marketplace”, “is there a skill that…”, “MCP server for…”, “discover tools for…”, or when local capabilities are missing. Returns ranked cross-source results plus install/config guidance.

Publish validated skill packs to PEtFiSh Market. Generates registry JSON entries from pack-manifest.json for optional packs. Trigger on 'publish skill', 'publish pack', 'release to market', '发布到市场'. Runs after quality-gate PASS. Outputs registry JSON files ready for commit to petfish-market. Supports --generate-index to regenerate index.json and --push to auto-commit and push to petfish-market via gh CLI.

Rewrite, polish, humanize, de-AI, or formalize Chinese or English technical, academic, business, course, proposal, patent, and email content into Petfish's structured, evidence-based, engineering-oriented style. Supports custom style profiles, de-ai-detector reports, enhanced Chinese de-AI rules, and optional taste enhancement. Triggers: 用我的语言习惯表达, 说人话, 润色, 去AI味, 按我的风格写, 改得更像人写的, 论文润色, 学术写作, abstract rewrite, rewrite my paper.

Extract personal writing style from samples to create a style profile. Analyzes sentence patterns, vocabulary preferences, argumentation structure, punctuation habits, tone markers, and AI-distinguishability signals across Chinese and English. Use when "提炼我的写作风格", "extract my style", "create style profile", "analyze my writing", "风格画像", "个性化风格提取", or when setting up petfish-style-rewriter for the first time.

Detect AI writing patterns in Chinese or English text. Use when the user asks to 检测AI味 / 检测AI痕迹 / 去AI检测 / AI写作检测, detect AI writing, check for AI patterns, is this AI-generated, AI slop check, or detect_ai. Produces a structured quantitative + qualitative detection report; does not rewrite.

五人顾问团多视角对抗式判断。用 Council 分析方案评估、战略判断、产品定位、技术路线、研究设计、课程设计、Presentation主线、逻辑审查、写作结构取舍、是否值得做、如何向客户/评审/老板表达等复杂判断。显式触发：用Council分析、五人顾问团审查、多视角评估、对抗式审查、不要迎合我。比 fish-calibrate 更深：5个logical subagents + Arbiter删除弱观点，输出可执行结论。/ Use Council for multi-perspective adversarial review of strategy, product positioning, tech roadmap, research design, course design, business analysis, logic checks, and how-to-communicate decisions. Triggers: "Council analysis", "five-advisor review", "multi-perspective evaluation", "adversarial review".

For 评审, 评价, 批判, review, critique, feedback, judgment, decision, evaluation, calibration, sycophancy, 迎合, 校准, 方案评估, code review, 可行性分析, architecture evaluation, proposal critique, strategic judgment, design review, or asks whether an idea/approach/understanding is right. It reduces sycophancy by neutralizing leading prompts, defining rubrics before conclusions, contrasting support vs opposition, separating conclusion from confidence, and improving reasoning quality in judgment-heavy tasks.

Production-grade skill authoring. Create, improve, refactor, or extract OpenCode/Claude skills with precise activation, executable workflows, evals, templates, and quality gates. Trigger on "create a skill", "generate skill", "new skill", "write a skill", "improve skill", "add evals", "refactor skill boundaries", "extract workflow into skill", "skill for X". Produces SKILL.md, references, assets, scripts, evals with toolchain handoff to lint/audit/gate/optimize.

topic_detect is high risk, users ask to 整理/切换/合并/归档话题 or 清空上下文, or mention topic governance/上下文污染/继承隔离/session resume. It routes continue/fork/switch/merge/archive/reset/bridge, applies context policy, builds context packages, logs decisions, and manages session boundaries via context-state MCP.

Convert PDF/DOCX/XLSX/HTML/PPTX/EPUB to Markdown for reading, review, and extraction. Use when user needs to read documents, extract text from PDF, convert DOCX to markdown, extract spreadsheet content, or any non-Markdown document needs structured reading. Trigger for 读取文档, read document, PDF转markdown, DOCX内容, convert to markdown, 文档内容提取.

Generate test cases/test matrix for the current repo: API/CLI/UI/SDK/service, smoke/regression/acceptance/negative/boundary tests, traceability mapping, coverage gap补齐, or automation scaffolds (pytest/Playwright/contract). Trigger for 根据项目/模块/设计文档生成测试用例；not for generic testing theory.

Generate grounded usage docs from the current repo: README, Quick Start, configuration, usage, API/CLI/SDK docs, troubleshooting, FAQ, deployment/operator guides, or to整理零散设计说明 into user-facing doc sets. Trigger for 根据代码与设计文档生成交付级文档；not for generic polishing/translation.

系列文档风格一致性治理。当用户要求跨文档统一风格、术语、命名和排版时激活。Use when the user is writing a series of books, chapters, long-form articles, course materials, whitepapers, or Markdown documents and wants to keep style, terminology, naming conventions, section structure, and Markdown layout consistent. Triggers on: 系列风格, 跨文档一致性, 风格画像, 术语漂移, 排版漂移, 风格归一化, style profile, terminology drift, layout drift, style audit, series consistency, reference style, naming rules. Extract a style profile from a reference file or the first document. Audit, report, and optionally rewrite targets to match the baseline while preserving facts and author intent.

Skill trust/governance requests: skill trust, skill安全, 治理, 可信度, trust scan, risk score, redline check, pre-publish trust verification. It wraps trustskills CLI via scripts/trust_scan.py for single-skill or root scans, manifest generation/verification, custom policy YAML, and returns governance level (allow/allow_with_ask/sandbox_required/manual_review_required/deny) with JSON-ready output.

对已部署/升级/回滚后的服务做功能验证：health/readiness、核心API smoke test、页面可访问性、端口监听、日志与依赖(DB/Redis/MQ/proxy)核验。Trigger for验收、交接、巡检、故障修复后复验；用于证明“服务可用”而非仅“进程已启动”。

Read, normalize, compare, extract, or convert reference materials in PDF, Markdown, DOC, DOCX, images, slides, or mixed formats into course inputs.

Read/inspect/summarize/audit/compare PPT/PPTX, extract slide inventory (titles, structure, notes, comments, media, links), review templates/layout/placeholders/sensitive info, or produce rewrite briefs and per-slide action plans for ppt-writer. Trigger for 读取/总结/审阅/对比课件、提案、汇报材料 and visual QA requests.

Initialize/scaffold/bootstrap AI-agent workspaces, generate AGENTS.md/README/.opencode/docs/tasks/qa/mcp templates, run profile setup (minimal/course/code/ops/security/research/writing/skills-package/comprehensive), configure uv dev env, and provide post-init wizard + pack install guidance. Enforces safe no-overwrite init with explicit confirmation for risky operations.

PEtFiSh skill发布门禁：串行执行skill-lint + skill-security-auditor + 元数据校验， 输出门禁报告与PASS/CONDITIONAL/FAIL决策。Use for “publish skill”, “can this skill be released”, “run quality gate”, “check before publish”, pre-merge checks, and recursive batch gating of skill roots before registry release.

Analyze GitHub/local repos to mine reusable workflows for PEtFiSh skills. Trigger on “analyze this repo for skills”, “mine skills from…”, “what skills can we extract”, “skillize this repo”. Scans docs + executable entrypoints, filters one-off/unsafe/non-automatable ideas, and outputs candidate skills, boundaries, required tools, risks, and priority ranking.

Analyze/optimize SKILL.md frontmatter descriptions for trigger precision. Trigger on “optimize description”, “improve trigger”, “fix skill description”, “description too broad”, “skill not triggering”. Checks length, trigger phrase density, specificity score, activation boundaries, sibling overlap/collision, then outputs actionable rewrite suggestions and a replacement description.

Lint/check/validate skill quality and trigger reliability. Trigger on “lint skill”, “check skill”, “validate skill”, “is this skill valid”, pre-publish QA, or debugging load/match issues. Runs lint_skill.py on single skill or recursive roots, reports ERROR/WARN/INFO with rule IDs, score/counts, JSON output, and optional dry-run/apply fix workflow.

Audit skills for security risks. Trigger on “audit skill security”, “check skill safety”, “security review”, “is this skill safe”. Scans SKILL.md, scripts/, references/, MCP/tool scope for prompt injection, secret access, dangerous commands, remote execution, excessive permissions, unsafe network/ file operations; returns 0.0-1.0 risk score, severity findings, pass/fail, and remediation guidance.

Evaluate whether a skill triggers correctly using positive should_trigger and negative should_not_trigger query sets. Trigger on “evaluate triggers”, “test skill trigger”, “trigger accuracy”, “false positive rate”, “is my skill triggering correctly”. Reports pass rate, false positive/negative rates, per-query decisions, and sibling cross-trigger conflicts when requested.

追踪并分析skill使用：记录激活事件、会话覆盖、helpful/not_helpful反馈， 生成usage report识别高价值与低活跃skill并给出推荐优化。Use for “usage stats”, “which skills are popular”, “skill analytics”, “track usage”, project-skill affinity analysis, and local governance insights via .opencode/skill-usage.json.

结构化反思与经验沉淀。Use when 反思, reflect, what went wrong, lessons learned, 复盘, 经验总结, 失败分析, root cause analysis, why did this fail, 返工原因, rework analysis, postmortem, 教训, takeaway, or when user corrects agent output, or when same operation fails 2+ times consecutively. Turns one-off corrections into reusable prevention rules and project knowledge assets via L1 instant reflection, L2 task debrief, and L3 guidance file generation.

Create, revise, expand, compress, or review course chapter content, including explanations, examples, transitions, key takeaways, and teaching flow.

Drive course projects end to end — plans, outlines, content, labs, learner/instructor materials, QA, QC, and release decisions. Also triggers for broad course development or curriculum requests.

Create, reorganize, normalize, or audit a course project directory tree, including naming rules, artifact placement, archive cleanup, and scaffold generation.

Create, modify, review, or operationalize course labs, exercises, demos, or hands-on projects, including objectives, environment, steps, validation, troubleshooting, and answer keys.

Reusable course-development methods, review heuristics, historical conventions, playbooks, or packaged guidance extracted from prior course discussions and project practice.

Create, modify, or review a course outline, syllabus, chapter tree, hour allocation, module progression, learning objectives, or prerequisite map.

Structured course QA: completeness checks, consistency review, pedagogical review, lab readiness checks, artifact coverage, issue logging, or release readiness assessment.

Turn QA findings into concrete quality control actions, remediation plans, closure tracking, and release or re-review reports.

Create, revise, or review a course development plan, including milestones, owners, dependencies, deliverables, risks, acceptance gates, and change control.

Course-related diagrams in draw.io form, including architecture diagrams, module maps, timelines, workflows, role flows, lab topologies, or slide-ready visual structures.

Instructor-only assets such as teaching notes, speaking points, timing guidance, answer keys, discussion prompts, demo cues, or delivery risk reminders.

Learner-facing course assets such as handouts, reading packs, worksheets, pre-class guides, post-class exercises, glossaries, or concise recap notes.

Polished Markdown artifacts for course plans, outlines, lesson notes, lab guides, learner handouts, instructor guides, QA/QC records, or course reports; also converts PDF/DOCX/images/notes into structured Markdown with normalized terms, explicit assumptions, and verification placeholders.