AI Coding Agent Guide: Context, Instructions and Review

Learn an AI Coding Agent Workflow

The FindUtils AI Agent Starter Guide uses simulated terminals and editors to teach coding-agent workflows. Start with project context, permissions, a bounded task, and verification. The exercises do not run a real coding agent. Check the selected product’s current documentation before applying commands or configuration files.

This guide walks through the most impactful techniques for each tool, so you know exactly what to learn first.

Why Most Developers Underuse AI Coding Tools

AI coding assistants have evolved far beyond simple autocomplete. Modern tools like Claude Code can autonomously read your entire project, run commands, manage git workflows, and spawn parallel agents. Yet most developers interact with them the same way they did with early Copilot: wait for a suggestion and press Tab.

• Project instructions give an agent persistent context. Use the instruction file that the selected agent supports.
• Reasoning controls vary by agent and model. Check the current documentation before you use a mode name.
• Slash commands handle common workflows in seconds (/review, /commit, /tests), but are invisible unless you know to type /
• Multi-agent workflows let the AI parallelize research and coding, but are only documented in advanced sections of official docs

The gap between basic usage and power-user productivity is enormous. The interactive guide at findutils.com closes that gap through hands-on practice rather than documentation reading.

Claude Code: The 10 Tips That Matter Most

Claude Code is Anthropic's agentic coding CLI that runs in your terminal. It reads files, writes code, runs commands, and manages git workflows autonomously. Here are the techniques that deliver the biggest productivity gains.

Step 1: Set Up CLAUDE.md for Persistent Context

Create a CLAUDE.md file in your project root. Claude Code reads this at the start of every conversation. Include your tech stack, coding conventions, build commands, and explicit rules the AI should always follow.

# Project Instructions

## Tech Stack
React 19 + TypeScript + Tailwind CSS 4

## Rules
- Use pnpm not npm for all package commands
- Max 300 lines per component file
- All API calls go through src/lib/api-client.ts
- Use Zod for all input validation

The AI Agent Starter Guide includes an interactive editor exercise where you practice writing CLAUDE.md files with real project conventions.

Step 2: Check the Available Reasoning Controls

Reasoning controls depend on the selected model and product version. Use the current Claude Code workflow documentation to check supported controls.

Describe the problem and acceptance criteria directly. For example: "Review the token refresh flow. Identify expiry, retry, and session-revocation cases. Propose a change before editing." A keyword does not guarantee a correct design.

Step 3: Check Commands in Your Installed Version

Open the product's help or command menu. Confirm the command behavior before using it. A command that rewinds code can remove work; review its scope first. The interactive exercise is a demonstration, not authorization to run that action in a real repository.

Step 4: Use Plan Mode Before Coding

For changes with unclear scope or significant effects, inspect the problem and agree on the intended behavior first. Claude explores the codebase, considers tradeoffs, and presents a detailed implementation plan for your approval before writing any code. This prevents wasted effort from going down the wrong path.

Step 5: Master Context Management

Long conversations with irrelevant context degrade AI performance. Use /clear between unrelated tasks. Use /compact with specific focus instructions when context gets long. Use @ file mentions to point Claude at specific files instead of letting it search broadly.

GitHub Copilot: Supply Project Context

Copilot features differ by client and mode. Use the supported instruction file for your environment. GitHub documents repository-wide instructions in .github/copilot-instructions.md and additional instruction scopes in its repository instruction guide.

Document the project commands, coding rules, and verification requirements. Keep credentials out of these files. Check whether the selected client applies the instruction scope before relying on it.

For a multi-file change, identify the relevant files and expected behavior. Review the complete diff, including files outside the initially selected context.

Cursor: Apply Rules at the Correct Scope

Cursor supports project rules under .cursor/rules/. Use the current rule documentation to select the file format and activation scope.

A rule should describe a real project constraint. For example, name the shared API client rather than asking the agent to invent another one. Use the editor's current context controls to select the files needed for the task.

Check the generated diff before accepting changes. Editor context is useful, but it does not prove that the agent inspected every affected call site.

Universal AI Prompting Techniques

These techniques work across every AI coding tool. They are the highest-leverage skills because they transfer regardless of which tool you choose.

Be Specific About Files and Functions

"Fix the login bug" produces generic solutions. "The login form in LoginForm.tsx is not validating email format. Fix it using the isValidEmail function from utils/validators.ts" provides more specific context. Review the result against the requirement. Always reference specific files, functions, and expected behavior.

Use the Writer/Reviewer Pattern

An independent review can provide another perspective. Give the reviewer the requirements, diff, and verification results. A new session does not guarantee that it finds defects.

Break Large Tasks Into Focused Prompts

Separate unrelated changes when that makes review easier. Keep the task’s required implementation and verification together. Do not omit a necessary check just to make the prompt shorter.

Have the AI Interview You

For complex features with unclear requirements, ask the AI to interview you before coding. It asks targeted questions about scope, constraints, and edge cases that surface hidden complexity before any code is written.

Choose a Tool by Your Workflow

Try a small task in the interactive guide. Then verify the same workflow with the actual product documentation. Do not compare current products using old prices or fixed context-window figures.

Common Mistakes When Using AI Coding Agents

Mistake 1: Kitchen Sink Sessions

Mixing unrelated tasks in one conversation pollutes the context. The AI's performance degrades as irrelevant information accumulates. Start fresh sessions for each distinct task.

Mistake 2: No Project Configuration File

Without CLAUDE.md, .cursorrules, or copilot-instructions.md, you repeat the same instructions every conversation. Spend 10 minutes creating a config file once, and save hours of repetition.

Mistake 3: Not Verifying AI Output

AI-generated code often looks correct but misses edge cases. Always provide verification criteria in your prompt: "Write validateEmail AND run the tests after implementing." Give the AI a way to self-check.

Mistake 4: Over-Specifying Instructions

Keep instructions relevant and resolve contradictions. Put detailed references in clearly linked files when the agent supports them. There is no universal line count that guarantees an agent reads or follows a rule.

Mistake 5: Correcting the Same Mistake Repeatedly

After two failed corrections, start fresh with a better prompt. The conversation context is polluted with failed attempts that bias future output.

Tools Used in This Guide

• AI Agent Starter Guide -- Interactive learning playground with interactive lessons for supported coding tools
• AI Model Picker -- Find the best AI model for your coding workflow based on speed, accuracy, cost, and context window
• Claude Code Usage Analyzer -- Track and optimize your Claude Code token usage and costs

FAQ

Q1: Is the AI Agent Starter Guide free? A: Yes. The findutils.com AI Agent Starter Guide is available without signup required. The interactive exercises and quizzes are accessible without creating an account. Processing happens entirely in your browser.

Q2: Do I need an AI coding tool installed to use the guide? A: No. The guide uses simulated terminals and editors so you can learn prompting techniques, configuration patterns, and workflow tips without installing anything. When ready, apply what you learned to your actual AI tool.

Q3: What is ultrathink in Claude Code? A: Reasoning behavior depends on the current model and Claude Code version. Check the official documentation rather than assuming a keyword enables a particular mode.

Q4: Which AI coding tool should beginners start with? A: Choose the interface you already use and verify its permissions and supported workflow. The simulated guide lets you compare exercises without installing an agent.

Q5: What is CLAUDE.md and do I need one? A: CLAUDE.md is a project configuration file that Claude Code reads at the start of every conversation. It contains your project's tech stack, coding conventions, and rules. Every project using Claude Code should have one. The equivalent files for other tools are .cursor/rules/ (Cursor), copilot-instructions.md (Copilot), and GEMINI.md (Gemini).

Q6: How long does it take to complete the full guide for one tool? A: Each tool has 4-6 modules with 3-5 lessons each. A typical module takes 5-10 minutes. You can complete an entire tool's guide in 30-60 minutes, or work through it across multiple sessions since progress is saved automatically.

Q7: Is my progress saved in the AI Agent Starter Guide? A: Yes. Lesson completion and quiz scores are saved in your browser's local storage. When you return, you can resume from where you left off. Each tool tracks progress independently.

Next Steps

• Try the interactive guide: Start with the AI Agent Starter Guide and pick your primary AI tool
• Compare AI models: Use the AI Model Picker to find the best model for your specific workflow
• Track your usage: Monitor costs with the Claude Code Usage Analyzer
• Learn universal skills: Complete the "Universal AI Skills" module in the guide for techniques that work with any tool