Move from software developer to FDE
Extend your engineering ability into customer discovery, evaluated AI systems, integration, and adoption.
Start here if: Developers who can build applications and want to own more of the customer delivery process.
Try the relevant skills checkIf you already ship software, the next step is to connect engineering choices more closely to a customer’s operation. That means learning how the work happens, agreeing on useful outcomes, and carrying an implementation through integration and handover.
AI introduces variable outputs and new evaluation needs. Your existing skills in testing, architecture, and debugging remain useful. Add model behavior to the system you inspect, while keeping the user’s problem central.
What to study, in order.
- 01
Start with discovery you can act on
- Study
- Learn to trace a workflow across users, systems, decisions, and exceptions. Ask how the customer measures delay or rework.
- Practice
- Interview a volunteer about a harmless recurring task. Write a narrow proposal and ask them to correct your assumptions.
- Keep as evidence
- Keep a workflow map, baseline, constraints, and the change caused by their feedback.
- 02
Implement one model-backed capability
- Study
- Take Building with the Claude API, or the Microsoft application path if Azure fits your goal. Learn structured outputs, retrieval, and tool interfaces.
- Practice
- Add a document-grounded answer or request classifier to an existing application.
- Keep as evidence
- Show the interface contract, model configuration, source handling, and a simple comparison baseline.
- 03
Evaluate the whole transaction
- Study
- Study representative test cases, human review, error categories, and safe interaction with external systems.
- Practice
- Test incomplete sources, malformed outputs, unauthorized access, provider failure, and repeated requests. Separate model failures from integration failures.
- Keep as evidence
- Preserve results by case type, transaction checks, and a reasoned decision about unresolved errors.
- 04
Run a supervised delivery exercise
- Study
- Learn adoption support, service ownership, incident handling, and release decisions.
- Practice
- Have another person use the feature and report a failure. Reproduce it, communicate its impact, fix it, and rehearse recovery.
- Keep as evidence
- Produce a release note, incident record, operating instructions, and handover feedback.
Use coding agents with engineering judgment
Claude Code 101 can help you learn context, permissions, and reviewing an agent’s changes. Treat that as development tooling. You still need to understand dependencies, inspect diffs, protect credentials, and run checks that test meaningful behavior.
When an agent produces an unfamiliar design, ask for an explanation and investigate the underlying code. In an interview or customer review, you should be able to explain why the implementation is suitable and where it can fail.
Demonstrate the decisions around the code
A useful portfolio includes a decision to narrow scope, a disagreement resolved with evidence, and a change prompted by a user. These examples reveal whether you can navigate ambiguity while protecting delivery.
Choose architectural complexity through comparison. Anthropic’s engineering guidance recommends starting with simpler solutions and adding agent behavior when it improves outcomes. A clear workflow with retrieval may meet the need; measure before adding a more autonomous loop.
A document assistant with an integration boundary
Build against a small public document collection and a mock ticket system. The application may draft an answer and propose a ticket update; a user reviews the specific change before execution.
What to produce
- Customer brief, baseline, and scope decisions.
- Application, mock integration, and setup instructions.
- Evaluation report covering answer quality and transaction behavior.
- Monitoring plan, recovery exercise, and user handover.
Ask a reviewer
- How did you separate retrieval errors from answer errors?
- What prevents an unauthorized or duplicate write?
- Which user observation changed your design?
Courses and resources for this path.
Choose resources for the gap you are working on. Course access and access to paid products or cloud services are separate.
Claude Code 101
Coding-agent setup, context, permissions, project instructions, development workflows, and reviewing changes.
- For
- New and experienced developers learning to work with a coding agent.
- Before you start
- Basic familiarity with a code editor and command line; a supported Claude account or API key. Check the provider’s current account requirements.
- Access
- Free course content. Claude product access or API usage is separate. Understanding and checking the code remains part of the learner’s work.
Building with the Claude API
API integration, structured outputs, evaluation, tool use, retrieval, agents, and implementation patterns.
- For
- Software engineers integrating AI into applications.
- Before you start
- Proficiency in Python, basic JSON knowledge, and access to an Anthropic API key.
- Access
- Free course content; API usage is separate. This is an application-development course, so complete programming foundations first if needed.
Develop generative AI apps in Azure
Model selection and evaluation, chat applications, tools, grounding, and responsible implementation.
- For
- Developers and AI engineers working with Microsoft Foundry and Azure.
- Before you start
- Programming experience and familiarity with fundamental AI concepts and Azure services. The path is intermediate.
- Access
- Microsoft Learn training content is free. Azure exercises require subscription access; former Learn sandboxes are no longer available. Check resource costs before running labs.
Evaluation best practices
Set evaluation objectives, choose representative cases, compare implementations, and calibrate automated grading with human review.
- For
- Developers and delivery teams assessing an AI feature.
- Before you start
- A defined task and examples of the outputs or behavior you want to assess.
- Access
- Public documentation, not a course or credential. Running model-based evaluations can incur API costs.
Common questions.
Do I need to train my own model?
That depends on the role. For this application-building path, learn to integrate and evaluate existing models first. Model training is a separate specialization and should follow a specific requirement rather than become a default first project.
Should I learn every AI framework?
Choose one implementation you can explain, test, and debug. Understand the underlying API and data flow. Expand your tools when another environment or a measured limitation justifies it.
What is the biggest change from regular development?
FDE assignments often put you closer to discovery, customer constraints, and adoption. The balance varies by employer. Demonstrate that you can turn an unclear operating problem into a bounded, working, maintainable result.
Sources and editorial basis.
Original guidance and practice projects by LockedIn Labs. Provider references describe their own offerings; inclusion is not an endorsement of this site.
Put the next step into practice.
Try the role-specific scenarios, keep the feedback, and bring a work sample into guided training.