Public pages are available in English. Some course lessons are still being translated and may fall back to German.
agenticbuilders
Free handout

Learn to build AI apps: a curated starting point.

Here is the thread to follow when you want to build with AI agents: understand first, plan small steps, verify the results, and guide the next round cleanly. The bottleneck is no longer only code, but context, clear assignments, and real verification.

Updated: 2026-06-08 · free to share

The four guiding principles

These four principles survive every tool change. Internalize them, and you can learn any tool.

  1. Learn the loop, not the buttons.

    Features expire. The loop remains: collect context, plan, act, verify, repeat. Invest in the foundation, not feature tours.

  2. Verification is the key skill.

    Give the agent a check it can run itself, and ask for evidence instead of blind trust. Small, verifiable diffs beat big ones.

  3. Vibe engineering, not blind vibe coding.

    Unreviewed code is fine for prototypes and risky in production. Responsibility always stays with the human. That is the credibility anchor.

  4. Increase delegation deliberately.

    AI touches a lot of work today, but only a small part is fully and safely delegable. Every technique has to answer this question: does it move me toward safer delegation?

The learning path

Four levels, each with learning goals and the best resources to get there. Claude Code is the thread because its ecosystem is especially dense. The concepts also apply to other tools.

  1. Level 0: Understand why everything is changing

    Before touching a tool: the mindset. Three sources for the conceptual frame.

  2. Level 1: Foundations

    A coding agent runs on your machine. You understand the loop, prompt specifically, read diffs, steer a session, and commit as your safety net.

  3. Level 2: The productive work loop

    You work in the Explore -> Plan -> Code -> Commit pattern, know spec-first work and the verification ladder, and maintain context hygiene.

  4. Level 3: Scaling, automation & cross-tool literacy

    Context engineering as a discipline, headless mode for CI, hooks, and MCP. You know the tool landscape and can run the loop in Cursor, Codex, or Aider as well.

Common failure patterns

Recognizing and fixing the most common pitfalls is a skill of its own.

Everything in one session

One session collects five tasks, context overflows, and quality drops.

Remedy: /clear between tasks; one session equals one task.

Correction loop

Fourth, fifth, sixth correction prompt, and it still does not improve.

Remedy: After about two failed attempts, restart instead of wrestling further.

Overloaded CLAUDE.md

The memory file is 300 lines, and the agent ignores most of it.

Remedy: Keep it short and high-signal. Only keep what truly applies.

Trust-then-verify gap

The output looks plausible and gets accepted without checks.

Remedy: Ask for evidence, run a check, and assign a reviewer subagent.

Endless exploration

The agent keeps reading and never acts.

Remedy: Scope the task, force a plan, and point to files and patterns.

Proof of method

Numbers from the engine room

What you read here does not come from a theory course, but from real operation. These figures come from the orchestrated workflow behind this site and point to the public tool behind it.

Updated: 2026-06

Orchestrated sessions
645
Documented work sessions across 17 real projects.
Documented learnings
1,680
Around 80 percent of them with high confidence.
Agent runs
7,700+
98.8 percent completed cleanly.
Parallel agents
up to 34
Coordinated in one session, typically 12 per deep session.

Measured with the open-source plugin session-orchestrator (MIT), the tool behind these sessions.

Want to understand how the system behind this works? Course: Multi-Agent Orchestration in Practice (coming soon, €49)

This list is the map.

The handout is the first guided step on it: no video course, no registration beyond email. A PDF you work through in an hour, with a real method loop you can apply to your own project right away.

Honestly: the depth around the ten use cases and reliability topics lives in the €29 guide. But the entry door is free, and it is enough for the first clean build.

Get the free handout

Curated link reference

Deliberately curated instead of a link dump. The full depth lives in the guide.

32 of 32 resources

Canonical primary sourcesBest Practices for Claude Code (opens in a new tab)

Canonical patterns: context management, plan mode, parallel sessions.

Source in EnglishCanonical primary sourcesEffective context engineering for AI agents (opens in a new tab)

Fill and manage the context window efficiently.

Source in EnglishCanonical primary sourcesBuilding effective agents (opens in a new tab)

Foundational article on composable agent patterns.

Source in EnglishCanonical primary sourcesAnthropic Academy: alle Kurse (opens in a new tab)

Around 17 free courses plus certificates, email required.

Source in EnglishfreeCanonical primary sourcesClaude Code Quickstart (opens in a new tab)

From installation to the first commit.

Source in EnglishCanonical primary sourcesMemory / CLAUDE.md (opens in a new tab)

Project instructions and auto-memory.

Source in EnglishMindset & methodKarpathy: Software Is Changing (Again) (opens in a new tab)

LLMs as a new kind of computer: Software 3.0, autonomy slider, jagged intelligence.

Source in EnglishMindset & methodWillison: Not all AI-assisted programming is vibe coding (opens in a new tab)

Terminology cleanup: vibe coding means no review before deploy.

Source in EnglishMindset & methodWillison: Vibe engineering (opens in a new tab)

The responsible version: agent use with clear accountability.

Source in EnglishMindset & methodBöckeler / Fowler: The role of developer skills in agentic coding (opens in a new tab)

Why experienced developers still matter.

Source in EnglishMindset & methodPhilipp Schmid: The New Skill in AI is Context Engineering (opens in a new tab)

A widely cited foundational definition of context engineering.

Source in EnglishPractitioner workflows & spec-drivenHarper Reed: My LLM codegen workflow atm (opens in a new tab)

The canonical brainstorm -> spec -> plan -> execute workflow.

Source in EnglishPractitioner workflows & spec-drivengithub/spec-kit (opens in a new tab)

Open-source toolkit for spec-driven development: spec -> plan -> tasks -> implement.

Source in EnglishPractitioner workflows & spec-drivenThorsten Ball: How to Build an Agent (opens in a new tab)

A coding agent as LLM + loop + tokens, shown in under 400 lines of Go.

Source in EnglishPractitioner workflows & spec-drivenArmin Ronacher: Agentic Coding Recommendations (opens in a new tab)

Practice recommendations: simple code, fast tools, and logging as the key to autonomy.

Source in EnglishTool landscapeClaude Code (opens in a new tab)

Anthropic's terminal and IDE agent; dense ecosystem with skills, subagents, hooks, and MCP.

Source in EnglishTool landscapeOpenAI Codex CLI (opens in a new tab)

Open-source Rust terminal agent plus asynchronous cloud tasks in isolated containers.

Source in EnglishTool landscapeCursor (opens in a new tab)

AI-native editor with agents; up to 8 parallel agents via Git worktrees.

Source in EnglishTool landscapeAider (opens in a new tab)

Often-cited open-source terminal pair programmer, deeply integrated with Git and model-agnostic.

Source in EnglishTool landscapeGoogle Gemini CLI (opens in a new tab)

Open-source Apache 2.0, 1M context, Google Search grounding, free with a Google account.

Source in EnglishTool landscapeJetBrains Research: Which AI Coding Tools Do Developers Actually Use? (opens in a new tab)

Survey data from more than 10,000 developers. Neutral comparison.

Source in EnglishCoursesClaude Code 101: Anthropic Academy (opens in a new tab)

Recommended starting point for beginners.

Source in EnglishfreeCoursesClaude Code: DeepLearning.AI × Anthropic (opens in a new tab)

Short course with Elie Schoppik, beginner to intermediate.

Source in EnglishfreeCoursesClaude Code: Frontend Masters (opens in a new tab)

By Anthropic DX engineer Lydia Hallie, advanced.

Source in EnglishfreeCoursesAgent Skills with Anthropic: DeepLearning.AI (opens in a new tab)

Build reusable skills, advanced.

Source in EnglishfreeCommunities, reports & standardsDiscord: agenticbuilders.at (opens in a new tab)

Our community: ask questions, show progress, and get honest feedback.

Source in EnglishCommunities, reports & standardsr/ClaudeAI (Reddit) (opens in a new tab)

Active Claude Code discussion: comparisons, prompts, community tooling.

Source in EnglishCommunities, reports & standardsAnthropic / Claude Discord (opens in a new tab)

Official community, direct exchange, and feedback channels.

Source in EnglishCommunities, reports & standards2026 Agentic Coding Trends Report: Anthropic (opens in a new tab)

Eight developments changing software engineering through coding agents, with case studies.

Source in EnglishCommunities, reports & standardsModel Context Protocol (MCP) (opens in a new tab)

Open standard, often described as USB-C for AI, for connecting tools and data to LLM apps.

Source in EnglishCommunities, reports & standardsAGENTS.md (opens in a new tab)

Machine-readable Markdown manifest in the repo root, cross-tool and Linux Foundation backed.

Source in EnglishCommunities, reports & standardsllms.txt (opens in a new tab)

Proposal for /llms.txt as a clean Markdown table of contents for LLMs.

Source in English

Glossary

Open 16 terms
Agentic Coding
Working with an LLM that acts in a loop on its own, reading, editing, running, and verifying under human supervision.
Agentic loop
Collect context -> plan -> act -> verify -> repeat.
Vibe Coding
Generating code and running it without reviewing it. Fine for prototypes, risky in production.
Vibe Engineering
Responsible agent use: the human remains accountable for the outcome.
Context Engineering
The discipline of giving the model the right tokens at the right time, not simply as many as possible.
Context Rot
Quality decay when the context becomes overloaded or noisy.
Delegation Gap
The gap between AI touching the work and AI being fully and safely delegable.
Spec-driven Development
Write a specification as an executable artifact first, then plan, then implement.
Plan Mode
A mode where the agent prepares a plan first, and you approve it before changes are made.
Subagent
A specialized helper agent with its own context window, used for research or isolated verification.
Skill (SKILL.md)
On-demand bundle of domain knowledge or workflow instructions that the agent loads when needed.
MCP
Model Context Protocol: an open standard for connecting tools and data to agents.
Hook
A shell command that fires deterministically on a lifecycle event, such as lint after every edit.
Headless Mode
Agent without an interactive UI (claude -p), for CI, pre-commit, and scripting.
Jagged Intelligence
LLMs can be superhuman at some tasks and unreliable at seemingly simple ones, a jagged capability profile.
Autonomy Slider
How much leash you give an agent per task, from tightly controlled to broadly autonomous.

Closing note

This collection is deliberately neutral and free to share. The tool landscape moves quickly, and links are checked from time to time. If you want the first guided step: get the free handout.