Module 6.3 · Ready-to-use skills

🧬 Anatomy of a reusable skill

Every well-crafted skill has the same parts. Memorize this skeleton and you'll create new skills in minutes.

SKILL: {{nome_da_skill}}            # v1.0
PAPEL:    Você é {{especialista_em_X}}.
OBJETIVO: {{o_que_a_skill_entrega}}.

ENTRADA (variáveis):
  - {{var_1}} : descrição
  - {{var_2}} : descrição

REGRAS:
  - {{regra_1}}
  - {{regra_2}}
  - Quando não tiver certeza: {{fallback}}.

FORMATO DE SAÍDA:
  {{schema_ou_estrutura}}

EXEMPLO (few-shot):
  entrada -> {{exemplo_in}}
  saída   -> {{exemplo_out}}

ENTRADA REAL:
"""
{{entrada}}
"""

When to use

Whenever you’re turning a standalone prompt into something reusable and versionable. It’s the template for the next four skills.

Customization points

ROLE e RULES — domain-specific
FORMAT — output schema
EXAMPLE — 1–2 client cases

Role

Who the AI is

Rules

Limits and fallback

Format

Predictable output

Few-shot

Guide example

📨 Skill: classify and route emails/messages

Reads a message, classifies it into fixed categories, and decides where to send it.

SKILL: Classificar & Rotear            # v1.0
PAPEL: Você é um classificador de mensagens de atendimento.

REGRAS:
  - Categorias permitidas: {{categorias}}   // ex: pedido, dúvida, reclamação, spam
  - Escolha SOMENTE uma categoria da lista.
  - Dê uma confiança de 0 a 1.
  - Se confiança < 0.6, destino = "humano".

FORMATO DE SAÍDA (apenas JSON):
  { "categoria": string, "confianca": number,
    "destino": string, "resumo": string }

Mensagem:
"""
{{mensagem}}
"""

When to use

Triage for inboxes, WhatsApp, and tickets. Put it at the start of any workflow that needs to decide "where this goes."

Customization points

{{categorias}} — the client’s
Confidence threshold (0.6) based on error tolerance
Category → destination map in the flow

📄 Skill: summarize and extract data from documents

Takes a long document and returns a summary + structured fields, citing the source.

SKILL: Resumir & Extrair Documento     # v1.0
PAPEL: Você extrai dados de documentos com precisão.

REGRAS:
  - Use SOMENTE o que está no documento. Não infira.
  - Para cada campo extraído, cite o trecho de origem.
  - Campo ausente = null.

FORMATO DE SAÍDA (apenas JSON):
  {
    "resumo": string,                  // 3 a 5 linhas
    "campos": { {{campos_alvo}} },     // ex: "valor", "vencimento", "partes"
    "citacoes": { "campo": "trecho..." }
  }

Documento:
"""
{{documento}}
"""

When to use

Contracts, invoices, proposals, long emails. For large documents, split them into chunks and process them in parts.

Customization points

{{campos_alvo}} — what to extract
Summary length
Toggle citations

Source only

Anti-hallucination

Citation

Source excerpt

Chunking

Large docs

Fields

Structured output

📈 Skill: generate a report from a spreadsheet

Takes tabular data and generates a narrative report with highlights and recommendations.

SKILL: Relatório de Planilha            # v1.0
PAPEL: Você é um analista que vira dados em insight executivo.

DADOS (CSV/tabela):
"""
{{dados}}
"""
Período: {{periodo}}   ·   Métricas-chave: {{metricas}}

REGRAS:
  - Baseie-se SÓ nos números fornecidos.
  - Destaque variações relevantes (subiu/caiu) com o número.
  - Tom executivo, direto. Sem jargão.

FORMATO DE SAÍDA (markdown):
  ## Resumo (2 linhas)
  ## Destaques (3 bullets com números)
  ## Pontos de atenção (1-2 bullets)
  ## Recomendações (2 ações concretas)

When to use

Automatic weekly/monthly reports from a sales, finance, or traffic spreadsheet. A deliverable the client immediately sees the value in.

Customization points

{{metricas}} — what matters
Report sections
Tone (executive / technical)

Data only

No made-up numbers

Variations

Went up/down with no.

Fixed sections

Consistent output

Actions

Recommendations

🤖 Skill: answer FAQs using a knowledge base (RAG)

Answers questions using only the supplied knowledge base—the RAG pattern that prevents hallucinations.

SKILL: FAQ com base (RAG)               # v1.0
PAPEL: Você é o atendente de {{empresa}}. Responde com base na fonte.

BASE DE CONHECIMENTO (trechos recuperados):
"""
{{contexto_recuperado}}
"""

REGRAS:
  - Responda SOMENTE com o que está na base acima.
  - Se a resposta não estiver na base, diga:
    "Não tenho essa informação, vou te encaminhar para um atendente."
  - Cite o trecho usado entre colchetes [fonte].
  - Tom: {{tom}}   // ex: cordial e objetivo

Pergunta do cliente:
"""
{{pergunta}}
"""

✓ What makes it work

✓"Use only the database" + an "I don’t know" phrase prevents fabrication
✓Retrieving good passages first (search) is half the work
✓Citing the source builds trust and makes auditing easier

✗ Where it breaks

✗Poorly retrieved knowledge base = confidently wrong answer
✗Without the “I don’t know” phrase, AI makes things up
✗Too much context dilutes the answer

💡 Practical tip

RAG quality depends more on the recovery (which snippets you inject) than on the prompt. Start with a small, well-organized foundation before scaling.

Foundation

Retrieved context

Source only

Without hallucinating

"I don’t know"

Routes to a human

Citation

[source] of the excerpt

🔁 How to adapt and version your skills

Reuse is what scales an automation business. You never start from scratch—you adapt a template and version it.

Adaptation workflow (step by step)

Copy the closest skill and rename it for the client: faq-paocia-v1.

Change only the customization points (role, rules, categories, format). Don’t change the skeleton.

Run the regression tests (same inputs, check the output). Only then release the version.

Version header (paste at the top of the skill)

# skill: faq-paocia
# v1.0  2026-06-02  base inicial (5 categorias)
# v1.1  2026-06-09  +categoria "entrega", limiar 0.6 -> 0.65
# v2.0  2026-07-01  schema de saída mudou: +campo "prioridade"
#
# CUSTOMIZAÇÕES deste cliente:
#   - categorias = [pedido, dúvida, entrega, reclamação]
#   - tom = cordial e curto
# TESTES DE REGRESSÃO: ver casos/faq-paocia.csv (8 casos)

When to use

With every new client or rule change. Versioning prevents “I fixed it here and broke it there” and preserves a history of decisions.

Customization points

Naming convention (skill-cliente-vX)
Where to keep the changelog
Regression test suite

Copy

Never from scratch

Customize

The right points only

Changelog

v1 → v2

Regression

Test before deploying

🎓 Module Summary

✓

Skill = packaged prompt - input, rules, format, and examples in one reusable piece.

✓

Four ready-to-use, high-ROI skills - classify, extract from documents, generate reports and FAQs with RAG.

✓

Adapt using the customization points - keep the skeleton, change only what's necessary.

✓

Version and test for regressions - it's what makes reuse safe and the business scalable.

You completed Track 6 — and the program!

With ready-to-use exercises, prompts, and skills, you have everything you need to apply the full AI automation cycle in any business. Go back to the beginning to review any track.

← Previous module Complete training →