Agentic Coding Cheat Sheet

The basic model

An LLM is the brain. A chatbot is the conversation. An agent is a junior teammate with tools: it can read files, make plans, change code, use the browser, and check results.

The key shift

Vibe coding says: "Make something that roughly fits." Agentic coding means: give context, sharpen the goal, limit scope, review the work, iterate.

Remember

You do not need the perfect prompt. You need a small safe workspace, clear boundaries, and a good feedback loop.

Project

A focused folder with exactly the files the agent may edit.

Skill

Reusable working rules: style, process, quality criteria, examples.

Tool

The ability to act: write files, open browsers, edit spreadsheets, operate apps.

Concept	Simple explanation	Analogy
LLM	The thinking engine	Brain
Chatbot	The conversation layer	Eyes / ears / mouth
Agent	The worker	Junior teammate
Project	The workspace	Office / desk
Skill	A repeatable workflow	Handbook / checklist
Tool	What the worker can use	Apps
Permission	Safety boundary	Key / locks

Your first safe agent setup

The agentic workflow in 8 steps

1Goal: Describe the concrete value, not only the technology.
2Scope: One small function first, no "build everything".
3Context: Provide files, examples, audience, rules, and limits.
4Plan: Ask the agent to explain briefly what it wants to change.
5Action: Allow small, reviewable changes.
6Check: Open the browser, test data, read errors, inspect the result.
7Feedback: Say exactly what feels wrong and what should stay.
8Save: Keep the result and start the next mini-iteration.

Practical starting rules

How to start

Create a test folder, not your entire desktop.
Use only example or demo data.
Write down one real, small task from your day-to-day work.
Ask the agent to read the files and summarize the structure first.
Then ask it for a small plan.
Approve changes step by step.

For small apps

Use simple files: index.html, styles.css, app.js.
No frameworks, no build system, no npm unless necessary.
Need browser storage? Use IndexedDB, not cloud persistence.
The app should work by double-clicking it in the browser.
Usability first, then visual polish, then convenience features.

Copy-paste prompts to start

The best prompt formula

Role: "Work like a calm senior frontend developer."

Goal: "Build a small local helper for ..."

Context: "Read the project files first and summarize the structure."

Boundaries: "No frameworks, no servers, no external databases."

Output: "Change the files directly and briefly explain the important points afterward."

Checks: "Verify that the page works directly in the browser."

Starter prompts

Understand the project

Read the files in this project. Summarize the structure, the main entry points, and obvious risks. Do not change anything yet.

Build a small app

Build a small local browser app for [workflow]. Use only HTML, CSS, and JavaScript. If data is stored, use IndexedDB. Keep the first step deliberately small.

Improve an existing app

Improve this app for [audience]. Keep the style, change only what is needed, then check usability, empty states, and mobile layout.

Fix a bug

The app behaves like this: [problem]. Reproduce the issue mentally from the code, find the most likely cause, fix it, and tell me how to verify it.

Prompt tip

If a result misses the mark, do not only say "wrong". Show an example of what "right" should look like.

Context and mini-app recipes

The context pack before every task

1. User

Who uses the result, and when?

2. Workflow

What happens before and after?

3. Data

Which fields, examples, rules, and edge cases matter?

4. Output

What should be visible, copyable, or exportable at the end?

5. Limits

What must the agent not do?

6. Quality

How will you know it is good enough?

Mini-app recipes

Tracker

For tasks, campaigns, leads, approvals, experiments, deadlines. Fields, status, filters, notes, export.

Checker

For lists, exports, bookings, texts, assets, QA rules. Import, rules, warnings, result list.

Generator

For briefings, checklists, emails, variants, handovers. Form, templates, preview, copy.

Mini scope

If you cannot explain the task in one sentence, it is probably still too large. Make it smaller.

Use cases and guardrails

Good first use cases

1. Marketing

Content calendar, campaign tracker, competitor notes, UTM checklist.

2. Design

Feedback board, asset checklist, responsive QA list, handoff tracker.

3. Sales

Lead briefings, call notes, follow-up tracker, quote library.

4. Operations

Meeting actions, incident postmortems, process checklists, handover board.

5. Finance & Tax

Export checker, duplicate review, category audit, receipt status.

6. Product

Feature request collection, interview notes, experiment memory, roadmap helpers.

Guardrails for safe work

Do this

Work with demo data.
Approve small changes individually.
Stop before deleting or overwriting.
Inspect the result yourself in the browser.
Test empty, wrong, and very long inputs.
Ask the agent to name risks and assumptions.

Avoid this

Passwords, API keys, or customer data in prompts.
Giving the agent your whole computer as workspace.
"Finish everything" without review checkpoints.
Automatic sending, deleting, or publishing without review.
Production data as the first test.
Clicking errors away instead of reading them.

What you should always check

The quick review checklist

Does the app open directly in the browser without installation?Can the main workflow be tested in under 60 seconds?Are labels, buttons, and empty states clear?Does it work on a narrow screen?Is data stored locally and not sent elsewhere silently?Are errors understandable enough to continue?Can you explain which files changed?Is the next improvement deliberately small?

When progress stalls

When the agent gets stuck

Ask it to summarize the current state in 5 sentences.
Ask for the smallest next change.
Give a concrete example of the correct output.
Ask for one hypothesis, not random trial and error.

The best closing prompt

Summarize what changed, how I test it, which assumptions you made, and which two next improvements would create the most value.

Start small. Check often. Give context. Keep responsibility.

An agent is not a magic wand. It is a very fast junior teammate with a toolbox. Good guidance makes it useful.

Choose the right entry point into the workshop series.

The formats build on each other: from short hands-on prototypes to concrete internal apps and production-ready applications.

Impulse Workshop

Entry point, half day

The compact entry point creates a shared view of agentic coding and already leads to short hands-on coding prototypes based on your first use cases.

Shared baseline understanding across the team
Relevant use cases from your daily work
Short coding prototypes for selected tasks
Clear recommendation for the next workshop step

from€990net

incl. up to 4 participants

Discuss the entry point

Hands-on Workshop

Implementation, 1 day

Building on the prioritized use cases, the team works hands-on with agent tools and builds concrete apps, automations, or workflows for daily work.

Work on real tasks from your company
Concrete apps, automations, or internal tools
Review and quality logic for daily work
Reusable templates for further use cases

from€2,990net

incl. up to 8 participants

Plan a hands-on workshop

Pilot Sprint

Pilot, 2 to 4 weeks

The pilot takes one selected app or automation to a production-ready application with preparation, implementation, reviews, tests, deployment, and rollout guidance.

Focused business use case with a clear goal
Production-ready app, automation, or workflow
Tests, review, deployment, and operating logic
Rollout recommendation for further teams

On request

individually scoped

Plan a pilot sprint