A source-backed, production-first field guide for Forward Deployed Engineer interviews.
Current release: 1.0.0-rc.1 — the content and automation contracts are reproducible at A1; stable promotion remains closed until the independent reader, facilitator, reviewer, and bilingual-use gates have real evidence.
简体中文 · English reading map · Guided practice · Job targeting · Start here · Role playbooks (Chinese) · Field Case Lab (Chinese) · Contributing
FDE interviews do not only ask whether you can write code. They test whether you can enter an ambiguous customer environment, find the real workflow, ship a thin but valuable system, keep it reliable in production, and turn field learning into reusable product capability.
This repository is a living handbook for that whole job. It combines:
- a current, source-backed map of FDE role archetypes;
- a practical interview operating system rather than a leaked-question dump;
- production AI coverage: RAG, agents, context engineering, MCP, A2A, evals, observability, durable execution, security, and data operations;
- original walkthroughs and facilitator-ready case packs with staged evidence, role-separated briefs, and case-specific rubrics;
- a transparent update process so time-sensitive claims can be reviewed and refreshed.
Most preparation material over-indexes on one of two halves:
- generic software interviews: algorithms, APIs, and system design; or
- generic AI interviews: model vocabulary, prompting, and toy chatbots.
The actual field role sits between the customer, the product, and production engineering. Current official role descriptions make that boundary clear:
- OpenAI describes ownership from discovery and technical scoping through build, rollout, adoption, and eval-driven feedback.
- Anthropic asks FDEs to deliver production artifacts such as MCP servers, sub-agents, and agent skills.
- Scale AI emphasizes customer-specific data infrastructure and distributed systems.
- Vercel combines embedded customer work with production agents, MCP servers, migrations, and knowledge transfer.
- Diligent explicitly calls for golden datasets, regression infrastructure, guardrails, tracing, and judgment about when a workflow needs an agent at all.
- Palantir describes FDE as the human equivalent of backpropagation: field feedback must become product capability.
The evidence and freshness dates are recorded in data/sources.json. Quarterly radar data is versioned under data/role-radar, and protocol/security baselines live in data/technology-baselines.json. These are bounded snapshots, not claims that every employer or geography follows the same pattern.
The handbook uses one reusable line of reasoning across case interviews, system design, project stories, and production incidents:
F — Frame the mission Who decides? Which workflow? What outcome matters?
I — Inspect reality Data, systems, users, permissions, constraints, failure history.
E — Engineer the thin slice Smallest end-to-end path that proves value and risk controls.
L — Launch and learn Evals, rollout, adoption, telemetry, incidents, iteration.
D — Distill into product Reusable primitives, playbooks, platform feedback, handoff.
It is not a script to memorize. It is a safeguard against the most common FDE failure: drawing architecture before understanding the job that must change.
| Time available | Recommended path | Output |
|---|---|---|
| 60 minutes | Read the role map, then score yourself with the master rubric | A prioritized gap list |
| 7 days | Follow the 7-day sprint and run one Field Case Lab | One recorded mock plus one scored case memo |
| 30 days | Follow the 30-day plan, run three mock loops, and build a portfolio narrative | Interview-ready evidence across all dimensions |
| Already interviewing | Use the question bank, worked cases, and blind case labs | Targeted practice, not broad rereading |
| One target role | Turn its public JD into an evidence-based campaign, then select one role playbook | A role brief, evidence matrix, and ten scored sessions |
| I keep reading but do not practise | Choose one guided 7-, 14-, or 30-day path | Ordered artifacts, completion evidence, and a repair cycle |
English-first readers can use the English reading map for the complete core learner-outcome path. Full parity means equivalent outcomes, evidence boundaries, practice routes, and completion logic—not sentence-by-sentence translation; the machine-readable contract is in data/content-parity.json.
| Chapter | What you should be able to do afterwards |
|---|---|
| Start here | Diagnose your target role and create a study backlog |
| What an FDE actually owns | Distinguish FDE archetypes and explain the field-to-product loop |
| Interview loop and scoring | Understand what each round is trying to observe |
| Discovery and decomposition | Turn a vague request into a bounded mission and acceptance criteria |
| Coding, data, and delivery | Demonstrate fast, testable implementation and messy-data judgment |
| System design | Design from workflow and risk, not from a memorized architecture |
| Production AI in 2026 | Reason about RAG, agents, MCP, context, evals, security, and durability |
| Casebook | Walk through three realistic customer problems end to end |
| Question bank with guided answers | Practice high-signal answers without memorizing slogans |
| Behavioral, stakeholder, and demo recovery | Show ownership, judgment, conflict handling, and calm under pressure |
| 7-day and 30-day study plans | Convert reading into observable interview performance |
| Resume and portfolio evidence | Present proof of field impact rather than a technology inventory |
| Answer calibration pack | Compare weak, independent, and leverage-creating answers using evidence |
| Enterprise field operating playbook | Carry interview reasoning into discovery, launch, incidents, handoff, and product feedback |
The Chinese core path continues with job targeting and the guided-practice system.
- Start here
- FDE role map
- Interview loop and scoring
- Discovery and mission framing
- Coding, data, and delivery
- Workflow-first system design
- Production AI in 2026
- Field Case Lab
- High-signal question bank
- Behavioral interviews and field leadership
- Evidence-first study plans
- Resume, portfolio, and project evidence
- Answer calibration
- Field operating playbook
- Turn one job description into an evidence-based campaign
- Turn reading into guided, evidence-producing practice
- Field Case Lab: ten production case packs with staged evidence
- Case facilitation standard
- Master scorecard
- Customer discovery scorecard
- System design scorecard
- 90-minute AI FDE mock loop
- 60-minute classic FDE mock loop
- Reviewer calibration guide
- Blind scoring, adjudication, and re-score protocol
- Cross-role evidence anchor library
- Field delivery worksheet pack
- Role-targeting playbooks for AI, data-platform, and regulated deployment FDEs
- Job-targeting worksheet pack
- Three role-targeted blind mock loops and pilot protocol
- Practice journal for first attempt, evidence, score, repair, and retry
- Corpus audit and design decisions
- Quarterly role-radar archive
- Technology baseline changelog
- Source policy
- Claim-review process
- Documentation-site evaluation
- v0.1 validation record
- v0.5 role-targeting validation record
- v0.6 guided-practice validation record
- v0.7 production-case validation record
- v0.8 reviewer-calibration validation record
- v0.9 bilingual-parity and archive validation record
- v1.0 release-candidate validation and completion audit
- Machine-readable release manifest
- Bilingual maintenance and source-archive contract
- v0.6-to-v1.0 execution plan
- Roadmap
- Accessibility
- It does not claim that an unofficial interview loop is company policy.
- It does not republish paid PDFs, copyrighted bundles, or confidential interview questions.
- It does not promise that memorizing model answers will pass an interview.
- It does not equate a framework name with good judgment.
- It does not treat every workflow as an agent problem.
Time-sensitive claims carry a source and last_checked date. Quarterly role snapshots are immutable additions, while protocol and security baselines keep a dated change log. Role playbooks separately register first-party signals, interview hypotheses, evidence boundaries, and practice assets. Guided paths separately register ordered sessions, learner outputs, and completion evidence. Monthly source-freshness and weekly public-link audits open maintenance issues when evidence ages out or a URL is confirmed dead. Changed Mermaid diagrams are rendered in CI, and versioned research and practice data is machine-checked. Contributors can propose a role update or dispute an overbroad claim using structured forms. Material changes are recorded in the changelog.
The project follows three confidence labels:
- Official: employer posting, protocol specification, standards body, or vendor documentation.
- Corroborated: multiple credible sources agree, but the employer has not published the detail.
- Community: useful practitioner experience; never presented as official policy.
For a reproducible local audit, run python3 scripts/verify_release.py --full --network. Add --require-clean when checking a committed release candidate. Network-dependent lint package resolution, Mermaid CLI resolution, and public-link health remain separate from dependency-free content contracts.
Corrections, fresh role evidence, original cases, translations, and clearer explanations are welcome. Please read CONTRIBUTING.md before opening a pull request. Do not submit leaked or proprietary interview content.
Released under the MIT License. This is independent educational material and is not affiliated with or endorsed by any employer mentioned in the guide. Job descriptions and interview processes change; verify current details with the recruiter.