← Back to Training

Full program · KENSHIN Bootcamp

6 weeks to have
your operational agent.

15 modules organized in 6 blocks. Short theory blocks, build-along exercises, curated external references for going deeper. At the end you have a real Kenshin agent, in production, solving a problem in your business.

Weeks

6

Total hours

60h

Cohort

20

Guarantee

14days

Reserve your seat →
01. Fundamentals02. Agent design03. Tools and real connections04. Automation by functional area05. Production — voice, security, deployment06. Final project

Week 1 · Fundamentals

Understand the box before opening it.

How a model thinks, what Claude is inside, how to design prompts that work in production. Without fundamentals, everything after is fragile magic.

01

How an LLM thinks (and why Claude)

🎯 Understand what a language model does under the hood and why Claude is different (long context, native tool use, alignment).

2h theory + 1h practice

📚 Theory

  • Tokens, context window and why they matter (100k vs 200k vs 1M)
  • Temperature, top-p, seed — when to touch each one
  • Real differences between Claude, GPT-4/5, Gemini for enterprise tasks
  • Cost per token and how to estimate it before building

🛠 Guided exercise

Run the same 5 prompts on Claude, GPT and Gemini. Document differences in a comparison.md file.

02

Professional prompt engineering

🎯 Write system prompts that hold up in production. Distinguish the Claude style (XML tags) from generic prompting and apply chain-of-thought.

2h theory + 2h practice

📚 Theory

  • System prompt vs user prompt vs prefill — when each one
  • XML tags: Anthropic's preferred technique for structured inputs
  • Few-shot examples: when they work and when they get in the way
  • Chain-of-thought and extended thinking (Claude 3.7+)
  • How to make a prompt reproducible — versioning and evaluation

🛠 Guided exercise

Write the system prompt of a sales agent for your own business, with at least 3 few-shot examples and XML tags. Test it on 10 different cases.

Week 2 · Agent design

From chatbot to digital operator.

An operating agent needs identity, memory and rules. Here you build the foundations that make your agent recognizable and consistent — not a generic that sounds like everyone.

03

9-layer architecture (Kenshin framework)

🎯 Learn the 9-layer framework that structures any serious agent: Identity, Memory, Context, Tools, Channels, Voice, Routing, UI, Security.

3h theory + 3h practice

📚 Theory

  • Why a chatbot is not an agent — the operational difference
  • The 9 layers, one by one
  • Base files: SOUL.md, IDENTITY.md, USER.md, AGENTS.md, OPERATING_AGREEMENT.md
  • Autonomy levels: green (acts alone), yellow (confirms), red (blocked)

🛠 Guided exercise

Create your 5 base files for your own agent. Explicitly define what it can do alone, what it confirms, and what it never executes.

04

Layered memory

🎯 Distinguish short (per-conversation), curated persistent and automatic daily memory. Implement all three.

2h theory + 3h practice

📚 Theory

  • Why one memory isn't enough — problems with infinite context
  • Summary strategies: extractive vs abstractive vs hybrid
  • MEMORY.md as human-curated persistent memory
  • memory/YYYY-MM-DD.md as automatic journal
  • Vector stores (Pinecone, Supabase pgvector) — when needed and when overkill

🛠 Guided exercise

Implement the 3-memory system in your agent. After 3 conversations, verify it remembers what matters and forgets the noise.

Week 3 · Tools and real connections

Give your agent hands.

An agent without tools is a chatbot with memory. You connect to real APIs, use MCP (Anthropic's standard for agents) and give your agent power to execute actions in the world.

05

Function calling and tool use

🎯 Learn how Claude decides when to use a tool, define clean JSON schemas and chain multiple calls.

2h theory + 3h practice

📚 Theory

  • How tool use works in Claude — the model → tool → result → model loop
  • Effective JSON schemas: what fields, what descriptions, what examples
  • Multi-tool orchestration: parallel vs sequential
  • Tool error handling — the model must be able to recover
  • Parallel tool use (Claude 3.5+)

🛠 Guided exercise

Give your agent 3 tools: search_web, read_file, send_email. Execute task: 'search 3 news items about X, summarize them into a file and send it to me by email'.

06

MCP — Model Context Protocol

🎯 Understand MCP, Anthropic's open standard for connecting agents to tools. Install and use 3 existing MCP servers.

3h theory + 3h practice

📚 Theory

  • What MCP is and why it replaces custom integrations
  • Architecture: client ↔ MCP server
  • Ecosystem of existing servers (filesystem, github, slack, notion, brave-search…)
  • How authorizations and permissions work
  • When to build your own MCP server vs use an existing one

🛠 Guided exercise

Install 3 MCP servers (filesystem, brave-search, github) in Claude Desktop. Make your agent search the web, read local files and create a GitHub issue.

07

Connect real services (Google, HubSpot, Notion…)

🎯 OAuth authentication, secrets management, webhooks. Connect your agent to the tools where your team already works.

2h theory + 4h practice

📚 Theory

  • OAuth 2.0 explained without the confusing part
  • Secrets: never in code, never in logs — where they do go
  • Webhooks vs polling — when each one
  • Rate limits and how to respect them
  • Debugging failed integrations — useful logs

🛠 Guided exercise

Connect your agent to Google Calendar (create event), Notion (search page) and HubSpot (read contacts). All with real auth, no hardcoded tokens.

Week 4 · Automation by functional area

The 4 agents that move any business.

Now apply what you learned to real areas: sales, customer service, content and operations. Each area has its own logic, data and KPIs. You build an agent per domain, not one all-powerful agent.

08

SALES — Hunter archetype (狩人)

🎯 Build an agent that prospects, qualifies and books meetings. Without burning your inbox or your LinkedIn account.

3h theory + 4h practice

📚 Theory

  • ICP (Ideal Customer Profile) — how to define it precisely
  • BANT and modern qualification frameworks
  • Multichannel cadences — email + LinkedIn + WhatsApp
  • Real personalization vs glorified merge tag
  • Metrics: open rate, reply rate, meeting rate, positive ratio

🛠 Guided exercise

Agent that receives a LinkedIn profile (URL), analyzes the lead with AI and generates a personalized first-contact message. Then: automatic follow-up at 3 and 7 days.

09

CUSTOMER CARE — Guardian archetype (守護者)

🎯 Build an agent that resolves 80% of queries 24/7, with clean handoff to a human when needed.

3h theory + 4h practice

📚 Theory

  • RAG (Retrieval Augmented Generation) — the essential technique
  • Build your knowledge base (FAQ, docs, past tickets)
  • Chunking and embeddings — the art of slicing information
  • Human handoff: when, how, with what context
  • Metrics: CSAT, response time, resolution rate without human

🛠 Guided exercise

Chatbot for your own website connected to your knowledge base (Notion or Markdown). Automatic email handoff when the user says 'I want to talk to someone'.

10

CONTENT — Herald archetype (伝令)

🎯 Editorial agent that plans, writes, schedules and monitors social content. In your voice, not generic.

2h theory + 4h practice

📚 Theory

  • Voice guide — how to capture your tone so AI can replicate it
  • Hook frameworks — Alex Hormozi, Justin Welsh, Nicolas Cole
  • Content calendar: cadence, formats, evergreen vs timely topics
  • Analysis of what content converts (not likes — leads)

🛠 Guided exercise

Agent that generates a 30-day LinkedIn editorial calendar based on your voice guide + schedules the posts.

11

OPERATIONS, HR, FINANCE

🎯 Internal automations: weekly reports, anomaly alerts, invoice and CV processing.

3h theory + 4h practice

📚 Theory

  • Reports generated from multiple sources (Sheets, ERP, CRM)
  • Threshold-based alerts — send only what matters
  • Document processing: invoices, CVs, contracts
  • HR and Finance compliance — where NOT to put AI

🛠 Guided exercise

Agent that every Monday generates a KPI report for your business consolidating data from 3 different sources and sends it to your email.

Week 5 · Production — voice, security, deployment

From toy to system that survives real users.

What separates a demo from a product: voice that survives imperfect transcription, permissions that prevent disasters, logs that let you audit, budgets that don't ruin you.

12

Operational voice — complete pipeline

🎯 Build a voice agent that works: STT + normalization + intent + response + TTS + barge-in.

3h theory + 4h practice

📚 Theory

  • Pipeline: audio → Whisper (STT) → error normalization → intent detection → agent → TTS
  • Expected vocabulary: how to teach your business's proper nouns to STT
  • Latency — target < 1.5s audio to audio
  • Barge-in: letting the user interrupt the voice
  • ElevenLabs vs OpenAI TTS vs Azure — when each one

🛠 Guided exercise

Complete voice-to-agent: you speak into the mic, your agent executes an action (search calendar, create task) and responds by voice.

13

Security, permissions, budgets

🎯 Apply the green/yellow/red system, per-channel budgets, auditable logs. So your agent doesn't delete, spend or publish without permission.

2h theory + 3h practice

📚 Theory

  • OWASP LLM Top 10 — the 10 risks you must know
  • Prompt injection and how to protect yourself
  • Sandboxing destructive tools
  • Cost caps per channel and per user
  • Structured logs: what to save, what NEVER to save (PII)

🛠 Guided exercise

Add to your agent: (a) tool allowlist per channel, (b) €2/day cost cap, (c) logs to Supabase, (d) confirmation on any destructive action.

14

Deployment and monitoring

🎯 Take your agent to production with real hosting, database, logging and observability.

2h theory + 3h practice

📚 Theory

  • Vercel + Supabase — Kenshin's default stack
  • LangSmith / Braintrust / Langfuse — observability platform comparison
  • Critical metrics: latency, cost, error rate, human fallback rate
  • Rolling deployments and feature flags
  • Alerting — when to wake you up at 3 AM

🛠 Guided exercise

Deploy your agent to Vercel + Supabase, with logging to Langfuse and email alert if daily cost exceeds €5.

Week 6 · Final project

Your own agent, running, for your business.

Apply everything to a real project: your operational Kenshin agent, solving a real daily problem. Group presentation and structured feedback.

15

Final project — your Kenshin in production

🎯 Deliver a complete operational agent that solves a real problem in your business, with the 9-layer architecture applied, at least 3 tools connected and deployed to production.

1 week

📚 Theory

  • How to choose the right problem (focus, not perimeter)
  • Define success metrics BEFORE building
  • Present your agent: live demo + what you learned + what didn't work

🛠 Guided exercise

Build, deploy and present your agent. Reviewer: Ana + community. Structured feedback.

Limited cohort

Start building
your Kenshin.

Maximum 20 students to keep live sessions and final-project feedback high-quality.

Reserve your seat →

14-day full refund