I would prepare for a Forward Deployed Engineering role by practicing the journey from an unclear customer request to a small delivery someone else can inspect.
A tutorial can teach a framework. A coding exercise can expose a gap in implementation. Neither, on its own, tells me whether I can discover the right problem, negotiate its boundaries, build something useful and explain what remains unresolved. Those are the transitions I would put into a training plan.
This is not a shortcut around engineering fundamentals. It is a way to connect them to a person trying to get work done. For an experienced builder moving toward an FDE role, the question is not only which technology to study next. It is which part of that customer-to-delivery journey the next practice session will make observable.
Read the work behind the title
OpenAI's current FDE description combines discovery, technical scoping, system design, building and production rollout with customer collaboration. It also names tradeoffs and direct coding. That is one employer's description, not a universal definition, but it is a useful corrective to treating the job as either a sales conversation or an isolated coding task.1
Start with the actual role you are considering. Highlight its verbs. Separate what you can already demonstrate from what you have only read about. If the role asks for production engineering, do not substitute a polished slide deck. If it asks for customer discovery, do not assume a technically impressive implementation proves you understood the customer's need.
Anthropic's October announcement offers another concrete reference. Its Frontier Deployed Engineer program describes a simulated enterprise deployment through security review and handover, followed by a practical on a new scenario and a residency. That is a provider's curriculum description, not evidence that my proposed approach produces the same results or that one program defines the profession.2
The useful design question is what the learner has to do when the instructions stop being complete. I would keep technology-specific study, but place it inside an exercise with an unclear request, a real constraint and a receiving person. The purpose is to expose a missing skill, not to decorate a familiar demo with a customer persona.

Conceptual scene. The equipment-rental exercise is synthetic; these are not customer records or a training result.
Practice the transitions, including the uncomfortable ones
Consider an explicitly synthetic exercise: an equipment-rental team asks for an AI assistant that answers availability questions. Use invented inventory records. Before building, ask whose decision the answer supports, how current the records are, what unavailable means and whether the tool may only report information or also reserve equipment.
For this exercise, choose a read-only availability report, not an autonomous booking system. Write the agreed inputs, exclusions and checks. Build a small path that reads the sample records and shows both the answer and its basis. Include a missing record, a stale timestamp and an ambiguous equipment name. An honest unresolved answer is more useful practice evidence than silently inventing stock.
Then change one requirement: the same item is described differently at a second depot. Keep the original test cases. Record what the change breaks, which question you should have asked sooner and whether the new solution still serves the original decision. The exercise is no longer only about finishing code; it reveals how you revise a model of the customer's work.
A November 2 preprint examines AI agents attempting requirement discovery and delivery in one enterprise text-to-SQL case, using a simulated customer grounded in that engagement. The authors report incomplete requirements and unreliable execution, while explicitly limiting generalization beyond the case. It is not a study of human training or interview success. I take it as a prompt to inspect those transitions, not as validation of this curriculum.3
I would separate a coached attempt from a transfer attempt. In the coached attempt, use permitted tools and explanations to learn. In the transfer attempt, change the records or constraint and state the assistance you will allow. If you need help, record it. The aim is to discover what you can now explain and execute, not to manufacture an independence score.
This distinction matters with AI assistance. A useful suggestion can improve the artifact while leaving your understanding unchanged. Ask yourself to explain the chosen data model, reproduce a failing case and identify a condition that would change your design. If you cannot, that is the next learning task, not a reason to hide the tool.

Original conceptual model. Transferring an approach is a proposed practice exercise, not a validated competency or employment measure.
Keep a practice record a colleague could challenge
I would organize the work into four passes, repeating the weak pass rather than buying a larger syllabus:
- Discover: capture the reader's decision, unanswered questions, allowed data and agreed boundaries.
- Deliver: build one executable slice, preserve failing cases and explain the engineering choices that matter.
- Transfer: change a meaningful requirement, disclose assistance and document what you can resolve without the original walkthrough.
- Hand over: give another person the setup, checks, limitations and next decision; ask them to identify what they cannot reproduce or understand.
The record can be a small repository and a plain-language brief. Keep the initial request, your clarification, the implementation, actual check results and an assistance log together. Label a proposed check as proposed until it runs. If nobody else has tried the handover, write that down. Do not turn an intended collaboration into an invented peer review.
My own VerShep FDE training project is exploring practice around decisions, evidence and recovery. Its public record remains an in-development prototype, not a released course, credential or validated hiring assessment. That boundary matters: a training environment can make work inspectable without establishing that a person is ready for every customer's production system.4
For interviews, this record gives you material to discuss honestly: what you asked, built, tested, changed and still do not know. It does not replace preparation for the employer's actual coding, design or collaboration assessment. Never present a synthetic exercise as client experience, or a tool's contribution as work you performed independently.

Conceptual handover, not an observed learner outcome. No customer deployment, assessment pass or hiring result is depicted.
Start with one small case you can afford to understand thoroughly. After the first run, choose the next session from the evidence: a discovery gap, a weak test, a brittle design or a handover nobody can follow. That is a more useful training decision than collecting another finished tutorial and assuming the difficult transitions will take care of themselves.
Research Record
References and Evidence
The four-pass practice record is my proposed method, not a validated curriculum or employer rubric. The rental scenario and changed depot are synthetic. Employer and provider statements describe their own work and programs; the research is author-reported and not independently replicated here. No hiring guarantee, accreditation, learner outcome, private CPF implementation or production authorization is claimed.
- Forward Deployed Engineer (FDE), SFPrimary employer role description · OpenAI; publication date unstated; reviewed November 14, 2026
About the role and responsibilities describe customer discovery, scoping, engineering and production delivery. One employer's current description, not a universal FDE specification, interview rubric or promise of an opening.
- Claude Frontier AcademyPrimary provider announcement · Anthropic; October 2, 2026; reviewed November 14, 2026
Describes a simulated enterprise deployment, security review, handover, a new-scenario practical and residency. Provider-reported program design and intentions, not independently demonstrated training effectiveness or a universal credential. Its program uses the term Frontier Deployed Engineer.
- Can LLM Agents Be Forward-Deployed Engineers?Research preprint, bounded case study · Marianne Menglin Liu and colleagues; November 2, 2026; version 1 reviewed November 14, 2026
Sections 3 and 4 and Limitations study agent discovery and delivery in one enterprise text-to-SQL engagement with a simulated customer. Reported omissions and execution failures are task-bounded; not a study of human learning, interviews, job replacement or this practice plan.
- VerShep FDE Training PlatformFirst-party portfolio record · Nick Holbrook; project record reviewed November 3, 2026; checked November 14, 2026
Documents an in-development prototype direction around decisions, evidence and recovery. Not publicly released, not a completed training product, credential or validated hiring assessment. No learner outcome or private implementation is disclosed here.