AI-Assisted Embedded Development

Objectives

• Use AI coding assistants (Claude Code, Copilot and Codex)

• Prompting to validated Firmware

• Spec-driven development

• Functional Specification Document (FSD)

• Configure MCP servers

• Design AI agents

• Build reusable Skills

• Validate firmware on target

• Apply IP and data rules

  
  

Prerequisite

• C language

• Embedded target experience

• Git basics

• Command line

• No AI experience needed

Course Environment

• Theoretical course

PDF material in English (printed for face-to-face); online over Teams.

Trainer assistance throughout.

• Practical activities (40-50% of duration)

  • Code examples, exercises and solutions.

  • Remote: one online Linux PC per trainee, with emulated or physical board as needed.

  • Face-to-face / on-site: one PC (one per two beyond six trainees), target board and install manual as needed.

• Downloadable preconfigured VM to redo the labs afterwards.

• Each session starts with a trainee check-in.
  

Target Audience

• Any embedded systems engineer or technician with the above prerequisites.

Course Outline First Day The AI Tooling Landscape for Embedded Engineers

• AI assistant families

• Interaction modes

• Why embedded is different

• From coding to orchestrating

• Strengths and limitations

  
  

Exercise: give the same driver prompt to AI tools and compare the results

Prompting for Embedded Problems

• Anatomy of a prompt

• Few-shot prompting

• Iterative refinement

• Prompting with embedded artefacts

• Reviewer-mode prompting

• Plan before code

• Detecting hallucinations

  
  

Exercise: rewrite three requests as structured prompts and measure the improvement

Spec-Driven Development and Functional Specification Documents

• Why specs come first

• Anatomy of a spec

• From requirement to spec

• Spec as contract

• Living specifications

• Role of the FSD

• FSD structure

• Generating a first-draft FSD

• Reviewing the FSD

• Iterating the FSD

• Handing off the FSD

  
  

Exercise: turn a feature request into a structured specification

Exercise: find the hidden assumptions in a requirement and make them explicit

Exercise: generate a first-draft FSD with Claude

Exercise: review and harden the FSD into a ready-to-build version

AI Assistants in Your IDE and CLI

• GitHub Copilot in VS Code

• Claude in VS Code

• Configuring the IDE

• Privacy and licensing

• What a CLI assistant adds

• Scaffolding a new firmware project

• Claude Code

• OpenAI Codex CLI

• Project memory files

• Encoding the constraints

• Permissions and autonomy

• Slash and custom commands

• Hooks (build, clang-format, lint)

• Plan Mode and Extended Thinking

• Claude Code on the web

• Context-window economics

Exercise: set up Copilot and Claude in VS Code, then build an I²C driver with each and compare

Exercise: configure CLAUDE.md and test hooks, then let Claude Code build a small driver

Second Day

Datasheets, Reference Manuals and

• The datasheet problem

• Feeding the right pages

• Per-project knowledge corpus

• Trust heuristics

  
  

Exercise: generate a DMA config

Model Context Protocol (MCP)

• MCP as the agent’s interface

• MCP server categories

• Designing the MCP toolbox

• Writing an MCP server

• Sharing MCP configs

Exercise: connect Claude Code to three MCP servers and verify each one

Exercise: write a small MCP server

Designing AI Agents for Embedded Development

• What an agent is

• The agent loop

• Single vs multi-agent

• Subagents and delegation

• Autonomy boundary

• Agent configurations

• Failure modes

  
  

Exercise: build a two-agent workflow (one implements the FSD, the other reviews it)

Exercise: run an agent on a build-flash-test loop, inject a failure, and watch it recover

Third Day

Skills, Reusable Configuration and Multi-Tool Orchestration

• The Skill concept

• Skills for embedded teams

• Installing and using plugins

• Plugins vs Skills

• Project-level config files

• Encoding team standards

• Versioning and sharing Skills

• Three tools, one workflow

• Where each fits

• Switching tools mid-task

• Decision guide

• Cost and licensing trade-offs

• IP and data residency

  
  

Exercise: write a Skill that turns a one-line request into an FSD, and test it on three cases

Exercise: write a CLAUDE.md of coding conventions and see how Claude Code’s output changes

Exercise: take one feature through all three tools (Claude for the FSD, Codex to build, Copilot to refactor)

Validating AI-Generated Embedded Code

• What review means

• Hardware failure modes

• Concurrency failure modes

• Timing failure modes

• Review workflow

  
  

Exercise: find and fix three planted defects (register, concurrency and timing) in a generated driver

Exercise: have a second AI tool review the first’s output and compare what each missed

Data Handling and IP

• IP and data handling

• Your company’s AI policy

  
  

Exercise: classify ten scenarios as safe, borderline or forbidden, and draft a one-page house rule