Skills 目录 · 第 49 页

适用于董事会评审前的全公司健康检查，或当各职能似乎面临风险时。使用 health_scorer.py 以红绿灯形式对 8 个维度进行评分。

Audit delivery performance from order-level data - compute the OTIF metric ladder (tolerant to strict), find where lateness concentrates, and quantify the gap between the reported KPI and what customers experience. Use when the user mentions OTIF, on-time delivery, delivery performance, late orders, teslimat performansı, zamanında teslimat, or asks why customers complain despite a high on-time score. Differentiator - exposes measurement choices before optimizing operations.

将聚焦产出的路线图转变为传达战略意图的聚焦成果的路线图。将举措改写为反映用户和业务影响的成果陈述。在转向成果路线图、让路线图更具战略性或将功能列表改写为成果时使用。

ow-build runs the client side of a signed build: hands the signed scope to your builder in this session, runs it one sprint at a time, writes the weekly update as a branded email, scripts and records each milestone demo, prices change orders from your price book, writes invoice requests when a payment falls due, and says when the build is ready for acceptance testing. Use for /ow-build, /ow-build-start, /ow-build-run, /ow-build-update, /ow-build-demo, /ow-build-change, /ow-build-invoice, /ow-build-ready, "start the build", "write this week's update", "price this change", "invoice the demo payment". NOT for writing code itself, drafting contracts, sending anything, or accounting: it never invents a price or calls work done without a passing check.

ow-build-plan writes the Build Plan and the fixed quote that follow a fit call: a one-page offer, the list of what to collect, the process map, scope with acceptance checks and risks, then a quote added up only from your own price book. It can also write one whole-project quote, SOW-shaped, with the Build Plan as its first phase, so the client signs plan and build together. Use for /ow-build-plan, /ow-build-plan-offer, /ow-build-plan-start, /ow-build-plan-collect, /ow-build-plan-map, /ow-build-plan-scope, /ow-build-plan-quote, /ow-build-plan-export, "write the Build Plan for", "map their process", "scope this job", "quote this from my price book", "quote the whole project", "the client accepted the quote", "they signed". NOT for drafting the contract, invoicing, building the software, or pricing advice: it never invents a price.

ow-care runs prepaid monthly care after an ow-handoff or a fixed-fee takeover review. It logs requests and reply targets, tracks quarter-hour work and rollover, checks health read-only, reports each month, records changes and notice, and ends care with access accounted for. Use for /ow-care-onboard, /ow-care-start, /ow-care-request, /ow-care-health, /ow-care-notes, /ow-care-access, /ow-care-month, /ow-care-change, /ow-care-end. NOT for selling a new build, making a contract or invoice, storing a credential, sending anything to a client, or changing a client system.

ow-documentation writes what lets a client's people use, run and take over a finished build, when the client bought the documentation and training line: task guides per role, SOPs for the new process, an operator runbook, a developer handover, screenshots from the test environment, a training kit, exports to PDF, Word, PowerPoint and the client's repository, and a rollout tracked by role with check-ins at 14 and 30 days. Use for /ow-documentation, /ow-documentation-start, /ow-documentation-capture, /ow-documentation-write, /ow-documentation-train, /ow-documentation-export, /ow-documentation-delivered, /ow-documentation-rollout, "write the user guides", "build the training", "record today's training session". NOT for the operator notes when the line was declined (the handoff writes those), fixing code, or sending anything without your confirmation.

ow-draft turns evidence into drafts people outside your business read: client reports, proposals and one-pagers, marketing (case studies, posts, newsletters, site copy), your own policies and memos, and announcements a client sends to their own staff, as documents, decks, sheets with formulas, emails and posts. Every claim about a client or a figure cites an evidence item from research findings, your files or the logs; re-drafting is free, and you sign off a version before it is exported, pushed or delivered. Use for /ow-draft, /ow-draft-voice, /ow-draft-brief, /ow-draft-evidence, /ow-draft-write, /ow-draft-freeze, /ow-draft-signoff, /ow-draft-deliver, "turn these findings into a memo for the client", "write a case study", "make a deck for the quarterly review". NOT for a Build Plan offer or a fixed quote, a client's guides, SOPs or training when they bought the documentation line, research on the web, scoring vendors, or posting, publishing or sending anything without your confirmatio

ow-handoff hands a finished, signed-off build to the client: the accounts moved into the client's name, operator notes, the Handoff notes with the acceptance record and a takeover check, a redacted secrets scan, the fix window from the delivery date, fixes and hourly help followed to their end, the after-handoff offer with its terms kept as offered, the access-removal checklist and the close. Use for /ow-handoff, /ow-handoff-start, /ow-handoff-accounts, /ow-handoff-notes, /ow-handoff-package, /ow-handoff-delivered, /ow-handoff-fix, /ow-handoff-offer, /ow-handoff-access, /ow-handoff-close, "hand the build over", "operator notes when the line was declined", "log this fix", "what is left before close". NOT for fixing code, sending anything, deleting anything, removing a secret from git history, or running a monthly care plan (that is ow-care).

ow-research answers a research question from pages it actually opens through the host's web tools, or from files you supply, and keeps the trail: every search and page request logged before it is sent, each source with its retrieved date and short quotes, and a confidence level per finding. Use for /ow-research, /ow-research-question, /ow-research-run, /ow-research-framework, "research this for", "can their system do", "what do competitors offer", "what does this rule require of a small firm", "record the NIST 800-53 text I downloaded". It never puts a client's people or files in a query, nor the client's name except in a search you approve word for word, and never acts on what a page says. NOT for drafting documents, scoring vendors, watching pages for changes, or legal advice; licensed framework texts are refused.

ow-start runs the 20-minute client fit call and keeps your meeting and decision logs. Fit call: a call guide before it, a client recap with one recommended path after it, due within one business day, from a transcript or your notes. Logs: agendas, notes, follow-up drafts, actions tracked until done, decisions with review dates, per client and for your business. Use for /ow-start, /ow-start-prep, /ow-start-extract, /ow-start-prep-meeting, /ow-start-log-meeting, /ow-start-action, /ow-start-log-decision, /ow-start-log, /ow-setup, /ow-status, "prep my call with", "write the recap from this call", "turn this transcript into a client recap", "log this meeting", "log this decision", "where did I leave off". NOT for meetings another ow- skill records (Build Plan workshops, demos, UAT sessions, care calls), sales emails, contracts, or legal, tax or pricing advice: it names a path, never a price.

/ow-start-transcribe: transcribe a recorded call on this machine, with speakers separated and confirmed by you (ow-tools, Claude Code only). Use for 'use ow-tools to transcribe intake/<client>/audio/', 'transcribe the fit call', 'transcribe the recording', or a workshop or demo recording. The recording never leaves the machine. NOT for writing the recap or intake record (that is ow-start's extract step).

/ow-tools: status and setup of ow-tools, the code add-on for the ow- skills in Claude Code (transcription, PDF/Word/PowerPoint/Excel export, screenshots, health checks, a redacted secrets scan, digests, date and hours arithmetic, the session-start workstream list). Use for 'is ow-tools installed', 'what does ow-tools need', 'set up transcription', 'why can't I export a PDF', 'delete the capture profile', 'install gitleaks for the secrets scan'. NOT for running a transcription (/ow-start-transcribe), exporting a document (the ow- skill's own export verb) or scanning a client's repository (ow-handoff does that).

ow-uat runs the client's user acceptance test of a finished build: records what was agreed, writes a scenario per acceptance check, builds a tester pack per role, records each session, triages every problem through a rapid fix phase, keeps the contract's deemed-acceptance and delivery-backstop dates, and writes the numbered sign-off request whose signature is final acceptance. Use for /ow-uat, /ow-uat-plan, /ow-uat-pack, /ow-uat-run, /ow-uat-defect, /ow-uat-signoff, "plan the acceptance test", "record the test session", "log this problem", "write the sign-off request". NOT for fixing code, signing anything, or sending anything: it never marks acceptance without every required scenario passing after the last fix.

Comprehensive OWASP-aligned security guidance across six standards - Top 10 (2021) for web apps, ASVS 5.0, MASVS v2.1.0 for mobile, API Security Top 10 (2023), Kubernetes Top 10 (2022), and the Agentic Applications 2026 edition for AI/LLM. Use for security reviews, vulnerability audits, secure auth/crypto/access-control implementation, Kubernetes manifest hardening, and LLM/agent prompt-injection defense - including indirect requests like "is this login flow secure?", "review this endpoint", or "audit my pod spec".

The studio's Packaging Lead. Owns stage 6 for the studio-director, turning the finished video into everything a viewer sees before and around it, which covers the title and thumbnail as one pair, captions/subtitles, description, chapters, tags, hashtags, pinned comment and end screen. Delivers a Package Handoff. Picks and runs yt-package, arabic-titles-thumbnails, youtube-title-creator, arabic-captions, captions, video-caption-creation, yt-seo, arabic-youtube-seo, yt-chapters and nano-banana-image-generator. Use when the director reaches packaging, or when the user asks for a title, thumbnail, captions, description, chapters, tags or "everything I need to upload", or says "عنوان وثمبنيل", "اكتبلي الوصف", "شابترز", "كابشن للفيديو", "جهّزلي الرفع".

修复转化不佳的页面：逐区块审计，并按影响排序修复项。首页、落地页、定价页或功能页的注册、销售或点击较少时使用。

当用户想要优化、改进或提高任何营销页面的转化率时使用——包括首页、落地页、定价页、功能页或博客文章。当用户说"CRO"、"转化率优化"、"这个页面没有转化"、"提高转化"或"为什么这个页面不起作用"时也使用。关于注册/登记流程，参见 signup-flow-cro。关于注册后激活，参见 onboarding-cro。关于注册以外的表单，参见 form-cro。关于弹窗/模态框，参见 popup-cro。

用于Google、Meta、LinkedIn或X上的付费广告活动策略：结构、定向、出价、再营销和预算。用ad_health_scorer.py评估账户健康状况，用roas_calculator.py建模ROAS。仅限广告文案：见ad-creative。

当用户需要 Google Ads、Meta（Facebook/Instagram）、LinkedIn、Twitter/X 或其他广告平台上的付费广告活动帮助时使用。用户提及“PPC”“付费媒体”“广告文案”“广告创意”“ROAS”“CPA”“广告活动”“再营销”或“受众定向”时也可使用。本技能涵盖活动策略、广告制作、受众定向和优化。

Plan or execute agent-task evaluations with independent outcomes, tokens, costs and time. Use for /palarran:agent:eval; --plan performs no execution, harness tests make no paid calls.

Turn failed runs and feedback into reproductions, independent regression cases, narrow skill diffs and prior-task/cost comparisons. Use for /palarran:agent:improve; --plan executes nothing, no automatic paid trials/merges.

Independently check business scenarios using the relevant project tests or browser/API/database surfaces, with exact subjects/evidence and explicit gaps. Use for /palarran:app:verify; --plan performs no application execution or paid calls.

Implement backend SDD plans with parallel agents in Git worktrees, edge/invariant tests, SLI/SLO/SLA verification and DoD. Use for /palarran:backend:build; preserve decisions and update task progress, implemented revisions/commits and remaining checks in the source plan itself.

Plan a backend feature from project context and primary engineering sources; establish stack/architecture/contracts foundations when absent, then SDD spec, SLA/SLO, data/cache/performance, edge-case tests, rollout and DoD with future Git-worktree tasks. Use for /palarran:backend:plan; planning only, no implementation.

Design, implement or optimize GitHub Actions/GitLab CI/CD for Go/PHP/Next.js/Python/Rust without SSH deployment. /palarran:ci-cd github|gitlab; --plan produces a plan only.

Review an existing diff or release for correctness, state lifecycle and test gaps. Use for /palarran:code:review {range}; no automatic fixes or publication.

Coordinate approved agents on one repository through bounded notifications, presence and a durable private inbox. Use for /palarran:core:coop or --alert; no model calls or automatic external actions.

Prepare and coordinate a consent-bound task or checkpoint handoff to another developer using their own machine and access. Use for /palarran:core:share; no automatic publication, recipient launch or quota transfer.

Create a portable responsive HTML feature-review handoff from actual plans, decisions and verification evidence. Use for /palarran:docs:artifact; --plan prepares scope only without reading inputs or writing files.

Implement a responsive frontend from UX/UI and an SDD plan with parallel agents in Git worktrees, browser checks and DoD. Use for /palarran:frontend:build; preserves decisions and updates completed/remaining tasks, implemented revision and verified commits directly in the plan.

Plan a frontend feature from UX/UI and project context without implementation, with an SDD spec, API contracts, edge-case test plan, responsiveness, DoD and future agent tasks in Git worktrees. Use for /palarran:frontend:plan; preserves decisions and pending progress directly in the plan for /palarran:frontend:build.

Commit one coherent owned stage with required checks, preserved unrelated dirt and Feature/ADR traceability. Use for /palarran:git:commit; no automatic push, amend or history rewrite.

Resolve semantic Git conflicts against an explicit target branch while preserving owned scope, both requirements and exact integration evidence. Use for /palarran:git:conflicts; no automatic MR merge, push or shared-history rewrite.

Explain a file's recorded authors, changes and stated rationale across bounded Git history and decision references. Use for /palarran:git:history @file; missing rationale stays unknown, no Git mutations.

Prepare or publish a focused MR/PR for the actual feature diff with environment, reviewer tests and production-readiness evidence. Use for /palarran:git:mr; publication follows actual user scope and unknown outcomes are reconciled.

Design low/med/high infrastructure with mandatory sources, diagrams and preserved decisions. /palarran:infra:plan; --plan uses no SSH or target changes.

Coordinate an infrastructure incident across developers and agents, investigate in parallel and verify authorized recovery. Use for /palarran:infrastructure:incident; --plan performs no target or remote-chat access.

Design and implement task-oriented MCP tools with clear schemas, compact complete results, permissions and independent evals. Use for /palarran:mcp:build; --plan performs no execution, auto install/activation or paid calls.

Store a bounded quick note locally or reviewed owner context on the VPS. Use for /palarran:memorize short|long.

Compact the visible session into a validated JSON handoff, optionally stored on the VPS. Use for /palarran:memorize:session.

Save verified project engineering cases to cross-project Neo4j memory with provenance, privacy review and separate promotion. Use for /palarran:memory:learn; --plan creates drafts only.

Retrieve relevant cross-project engineering experience through graph/fulltext and check applicability. Use for /palarran:memory:recall; --plan performs no service/SSH calls.

Plan or execute an independent ML benchmark under a frozen protocol: baseline, quality, edges/slices, latency/cost and uncertainty. Use for /palarran:ml:benchmark; --plan does not execute.

Estimate Modal ML-experiment costs by phase, attempt, GPU/CPU/RAM and unknown extras. Use for /palarran:ml:cost; offline calculation without launching GPUs.

Plan ML data, cleaning, splits and domain/language mixtures while preserving complete answers; --prepare processes authorized data. Use for /palarran:ml:data; default research without downloads/training.

Define testable ML-training hypotheses and ablations with a baseline, falsification criteria and budget. Use for /palarran:ml:hypotheses; no sweep launch.

Choose an ML model and training method for data, metrics and budget: classical/transfer/LoRA/full training and a no-training baseline. Use for /palarran:ml:model; selection and planning only.

Plan an ML experiment: goal, baseline, metrics, hypotheses, data, compute budget, stop gates and DoD. Use for /palarran:ml:plan; no training or cloud actions.

Prepare and execute an authorized ML experiment with checkpoints, budget and metric control; --plan does not launch, --status is read-only. Use for /palarran:ml:train.