Coding & Development

Master Anthropic's prompt engineering techniques to generate new prompts or improve existing ones using best practices for Claude AI models.

Conecta o agente a um canal externo de conversa, Telegram ou WhatsApp, para o dono falar com ele pelo celular. Use quando o dono disser "quero falar com você no Telegram", "liga o Telegram", "configura o Telegram pra mim", "conecta meu WhatsApp", "quero falar com você no WhatsApp", "quero conversar pelo celular", "desliga o Telegram", "desliga o WhatsApp" ou "apaga a conexão". Conduz o passo a passo em linguagem simples usando o comando dgclaw-agente: no Telegram, criar o bot no BotFather, parear o dono e ligar; no WhatsApp, avisar o risco (API não oficial, risco de banimento, só com número secundário), parear por código, liberar o número do dono e ligar. Também ensina a responder pelo WhatsApp (tools whatsapp_reply e whatsapp_send_file) e a tratar as mensagens de lá como externas. Não funciona no agente Admin, que não tem canal externo.

Marca, lista e desmarca compromissos de verdade do agente NA MAQUINA onde ele mora (timers do systemd, via persona-cron-mgr.py), pra ele agir sozinho na hora certa mesmo com ninguem falando com ele. Use SEMPRE que o dono pedir pra agendar, marcar ou repetir um trabalho DO AGENTE num horario ("todo dia as 9h me manda o resumo", "cria um cron", "cria um agendamento", "roda isso de hora em hora", "quais crons voce tem", "cancela aquele agendamento"). Vale tambem quando a transcricao de audio trouxer CHROME, CROME, CRONO, CROM ou CRONI num pedido de horario, que e quase sempre a palavra CRON mal ouvida. Pedido para AVISAR o dono ("me lembra...", "me avisa as 15h") nao e rotina: use a skill lembretes, que insiste ate ele confirmar. NUNCA use o agendamento da sessao do Claude Code pra isso: aquilo morre com a sessao, isto fica na maquina.

Transforma um processo que o agente acabou de fazer com o dono numa skill dele (pasta .claude/skills/<nome>/SKILL.md no workspace), pra fazer igual da próxima vez, e melhora essa skill quando o dono corrige o resultado. Use quando terminar com o dono um processo de 3 passos ou mais que dá pra repetir (montar um relatório, preparar um roteiro, conferir números numa fonte, organizar um material), quando o dono disser "aprende isso", "guarda esse jeito de fazer", "faz uma skill disso", "da próxima vez faz igual", "transforma isso num processo", "/dgclaw:aprender", ou quando ele corrigir alguma coisa feita por uma skill que você mesmo escreveu. Também serve pra reflexão da noite decidir se vale sugerir uma skill. Nunca cria skill sem o sim do dono.

Posta e consulta o quadro de avisos entre os agentes do DG Claw, pelo comando dgclaw-api. Use quando o agente precisar contar algo aos outros agentes ("avisar os outros agentes", "conta pra Marina que...", "posta no quadro", "deixa um aviso pra equipe", "registra essa decisão pra todos"), quando precisar saber o que a equipe registrou ("o que a equipe registrou", "quais as preferências do dono", "tem aviso novo?"), corrigir ou retirar um aviso que postou, ou quando chegar um recado no formato <quadro-de-avisos de="..." id="..." tipo="...">. Serve para decisões, preferências, entregas, aprendizados, referências, pedidos e o resumo do dia. Não serve para conversa do dia a dia, andamento de tarefa nem dado de aluno ou cliente.

Plan a sustainable X routine, including a daily 30 to 45 minute hand-posting rhythm, a weekly content rotation, posting time, and a weekly review. Does not schedule or publish.

Revise X drafts so they sound like the user, removing hype, generic AI phrasing, bait asks, and banned punctuation while keeping meaning and authentic vocabulary. Does not score AI authorship.

Coordinate project requests, focused work, shared context, and delivery from one chat.

Build a reply list and draft replies and quote posts on other people's X posts, including bigger accounts and peers, to grow a network. Does not like, follow, post, or DM.

This skill should be used when the user asks to spawn a new agent (or agent-shell). It creates an isolated git worktree for the new agent, runs any repository setup, and launches the agent.

Open a pull/merge request and stay with it until it merges: trigger the reviewers, watch for their verdict, work each round of findings, push, re-trigger, and merge when it is clean. Handles AI reviewers (CodeRabbit) and human maintainers, on GitHub (gh) and GitLab (glab). Use when asked to 'open a PR and babysit it', 'ship this and see it through', 'get this PR merged', 'watch the PR until it lands', or '/babysit-pr'. Also the review-and-merge half of ship-feature-pipeline.

Work one round of review comments on a pull/merge request: fetch every unresolved thread, verify each finding against the current code, fix what is still valid, reply with a justification to what is not, then commit, push and answer every thread. Use when a PR has review comments to deal with - from an AI reviewer (CodeRabbit and friends) or a human - and when asked to 'address the review', 'fix the review comments', 'respond to CodeRabbit', or 'handle the PR feedback'. The babysit-pr skill calls this for each of its rounds.

Turn an owner-supplied service request and catalogue into a traceable local quote draft, record attributed clarification, edit its wording, reopen saved work, and export an internal review packet or selected buyer-wording draft. Use for request-to-quote drafting, duplicate intake, interrupted work, or recovering a saved quote. Requires Node.js 20 or newer. Does not send quotes, connect accounts, approve prices, or make binding commitments.

Install, update, or prepare Sanad for normal use or source development on macOS, Windows, or Linux. Use when an AI agent must inspect a machine for an existing Sanad installation or checkout, explain the release-user and developer-source paths, install the Client and local Agent, install an Agent on a remote server, quickly create a Google Cloud Free Tier server and pair Sanad, update packaged components, or safely update and run a development checkout.

Build and operate principal-owned AI income systems that perform accepted paid work. Use when discovering an agent service opportunity, validating demand, creating an Outcome Pack, implementing a vertical AI system, defining agent authority, evaluating delivery, exposing a Skill/MCP/A2A service, or proving net economics. Active services belong here; affiliate and owned income assets belong to Agentic Passive Income; capital, crypto, lending, staking, trading, and yield experiments belong to DPI.

The operating brain for building income systems with AI agents. Use when planning, building, or scaling an affiliate/content/product income network — deciding what to build next, where money actually comes from, how to make it compound, and how the system improves itself. Composes affiliate-audit. Trigger phrases: build an income system, make money with agents, scale my content network, what should I build next, passive income with AI, monetize this site, agentic income.

Map which content mentions paying AI tools but lacks affiliate links, and which programs to join first. Use when monetizing AI-tool comparison content, choosing affiliate programs, or auditing a passive-income site's link coverage.

Map which content mentions paying AI tools but lacks affiliate links, and which programs to join first. Use when monetizing AI-tool comparison content, choosing affiliate programs, or auditing a passive-income site's link coverage.

Freeze a Product Document as a dated PDF from its latest saved Word version, review outstanding feedback, and attach the verified snapshot to a specified Jira issue when requested. A freeze records a revision; it does not grant approval.

Update text in an existing shared OneDrive or SharePoint Word document while preserving its file identity, revision history and embedded content. Use for scoped edits from a revised draft, including documents with active collaborators.

智能错题分析 Skill。读取用户上传的错题文件（图片/PDF/TXT/Markdown），结构化提取题目、答案与解题过程，识别知识点并分析错因，生成知识依赖图、题型地图、7/14 天个性化复习计划，最终导出 Markdown 与 PDF 报告。当用户要求分析错题、整理薄弱知识点、制定复习计划、生成错题报告时使用本 Skill。

>- Audit guardrail: treat missing evidence as a finding, never as a pass. Use when reviewing agent outputs, security scans, compliance checklists, or any report that claims something is "fine" - force the report to distinguish PASS from UNKNOWN before you act on it.

>- Verify an agent-payable API charges what it advertises: fetch the 402 challenge, compare against the published OpenAPI/pricing page, and flag drift. Use when publishing per-call API prices, before repricing, or when a buyer disputes a charge. One check, five minutes, works on any x402-style API.

Build and maintain Next.js web applications using a consistent src folder structure, domain controllers, Zod API contracts, composed components, and explicit client/server boundaries. Use for project architecture and application coding, while honoring the user's chosen libraries and providers.