Slides: March 2026 Agentic Hamburg

Most agents are built to be generally helpful. That's not the same as being useful for you.

This kit helps you configure an existing agent into a personal work companion: one that reflects how you actually work, covers the gaps you don't naturally fill, and knows when to slow down instead of just producing output.

CLAUDE.md              ← the rules (what your agent always does)
operator_profile.yaml  ← who you are and how you work
modes.md               ← task-specific behaviour you can switch on
drift_log.md           ← where you record when it goes wrong
NOTES.md               ← session memory (optional — see guides/05-notes.md)

questionnaire.md       ← 9 questions to calibrate your profile

guides/
  00-setup.md          ← get this working on your machine
  01-questionnaire.md  ← fill in the questionnaire
  02-edit-config.md    ← load your profile + add a watchout
  03-writing-style.md  ← generate your writing style
  04-drift-log.md      ← keep the config honest over time
  05-notes.md          ← session memory (bonus)
  06-modes.md          ← task-specific behaviour (bonus)

Follow the guides in order:

1. Get set up — clone, copy, or paste into your agent
2. Fill the questionnaire — 9 questions, first instinct
3. Load your profile — paste your results, add a watchout
4. Generate your writing style — let the agent observe, then describe
5. Drift log — how to keep the config honest over time

Bonus (when you're ready):

• Session memory — NOTES.md
• Modes

It helps to know upfront what's automatic and what's on you.

Your agent does this automatically:

• Stops before producing detailed output and asks clarifying questions (Alignment Gate)
• Detects which mode fits your session and names it if you didn't
• Flags when a new mode pattern is emerging and offers to draft it
• Offers to update your writing_style when it needs to compact a long conversation
• Nudges you at the end of each session: "anything that went wrong?"
• Maintains NOTES.md if you've activated that rule (see guides/05-notes.md)

You do this:

• Answer the questionnaire and load your profile (once)
• Choose your trigger word and update it in CLAUDE.md + operator_profile.yaml
• Write drift log entries (or ask your agent to draft them)
• Generate your writing style after a real working session
• Add new modes when a repeating pattern emerges
• Update the config when something keeps going wrong

Your agent won't just start producing output when you give it a task. It will:

1. Restate the goal in one sentence
2. Ask 1–3 clarifying questions
3. Propose a direction and flag what's still unclear

Then it stops and waits.

When you're ready for it to proceed, write your trigger word (default: Engage). That's what moves it from alignment to output.

This feels slow at first. It isn't. It's the thing that stops you getting a long, detailed response to the wrong question.

• Claude Code — run claude inside this folder
• Gemini CLI — use GEMINI.md (same structure, different filename)
• Claude Cowork — open this folder as your workspace, no setup needed
• Any agent with a system prompt — paste CLAUDE.md contents manually

You'll leave today with a starting point — not a finished product. The more you use it, the more you'll notice what's wrong. That's what drift_log.md is for.

When something annoys you enough to notice: log it. Update the config. The relationship calibrates over time.