A field curriculum for anyone using Claude Code, Codex, Cursor, Windsurf, Hermes or any coding agent to ship AI systems. Every module ends in an artifact your agent can inspect, test and improve — self-paced, artifact-driven, open.
Field thesis: most AI projects fail because the operator lets the agent code before the problem is scoped, measured, evaluated and owned. The Academy teaches the operating discipline that prevents that failure.
Domain dossier, scoping report, prototype spec, eval plan, production handoff, productization memo.
Domain research, 6-Q interviews, ROI sizing, architecture trade-offs, eval design, runbooks.
Prompt theater, vague requirements, eval-less demos, no-owner handoffs, SaaS-first confusion.
Self-paced. Every artifact maps to a template or script in the open-source repo.
Builders using coding agents to ship real projects. Control context & scope, demand concrete artifacts, use evals before trusting output.
A repeatable delivery method for client-facing AI work. Run discovery without vague questions, translate pain into scope, hand off maintainable systems.
Contributors improving the Skill, templates, scripts and examples. Add tests, sharpen rubrics, productize repeated field patterns.
Not "watch videos and feel inspired." You pass by producing artifacts that survive the FDE rubric.
What the Skill is, what it is not, and how to keep agents in co-founder mode instead of autocomplete mode. Pass artifact: local install + one vague prompt rewritten into an FDE request.
Scoping → Prototyping → Production → Feedback as the backbone of every engagement. Pass artifact: one-page operating memo (outcome, owner, timeline, artifact chain).
Market, pains, regulations, stack, benchmarks, news and talent — before discovery. Pass artifact: domain dossier with 5 insights that change the questions you ask.
Force process, output, data, error cost, current system and success metric before architecture. Pass artifact: completed 6-Q sheet with quantified answers and named assumptions.
Stakeholder map, pain matrix, ROI, risks, architecture sketch and 90-day plan in one decision artifact. Pass artifact: 5-page scoping report with a conservative ROI threshold.
Generate candidate architectures, validate against held-out constraints, prune weak paths, keep rejected hypotheses as lessons. Pass artifact: prototype spec with diagram, failure modes, held-out results, data flow.
Golden cases, adversarial cases, regression checks, human-review rules, metrics and drift alerts. Pass artifact: eval framework with target metrics, test categories, fail cases and owner.
Deployment, rollback, observability, security, runbooks, cost projections, ownership, incident paths. Pass artifact: handoff another team could operate tomorrow.
Extract repeated patterns into templates, scripts, adapters, examples and eventually MCP tools. Pass artifact: productization memo ranking assets by effort, ROI and strategic fit.
| Lab 01 | Convert a vague founder prompt into an FDE request. |
| Lab 02 | Build a domain dossier for one vertical. |
| Lab 03 | Write six stratigraphic discovery questions. |
| Lab 04 | Complete a quantified 6-Q interview. |
| Lab 05 | Calculate ROI and sensitivity cases. |
| Lab 06 | Write a 5-page scoping report. |
| Lab 07 | Generate and score 4 architecture candidates. |
| Lab 08 | Draft a prototype spec with failure modes. |
| Lab 09 | Create an eval suite (happy, edge, adversarial, regression). |
| Lab 10 | Write a production handoff runbook. |
| Lab 11 | Score your own work with the 6-trait rubric. |
| Lab 12 | Extract reusable IP into a productization memo. |
Generic work fails. A passing artifact scores ≥3 on every trait, with Ownership and Decomposition at 4+.
| Trait | Excellent looks like | Auto-fail signal |
|---|---|---|
| Customer Curiosity | Specific domain reality and stakeholder context. | Generic "AI can help" language. |
| Ownership | Concrete outcome, owner, date and timeline. | "Could", "maybe", "explore". |
| Decomposition | 6-Q answered with numbers. | Vague restatement of the problem. |
| Empathy | Adoption, politics, maintenance and constraints named. | Ignores the customer's operating reality. |
| Product Sense | Shippable artifact, failure modes, production path. | Slides, theory, or no runbook. |
| Communication | Executive summary plus technical detail. | Jargon or oversimplification. |
Run the full loop on one realistic workflow: research the dossier → 6-Q + ROI → scoping & GO/NO-GO → prototype spec, eval suite & handoff → score with the rubric and revise. You don't build the whole system to pass — the package must make implementation obvious, measurable and ownable.
Install the Skill, then use the Academy as the path. The first prompt should not ask the agent to code — it should ask the agent to scope.
# from the repository root mkdir -p ~/.claude/skills ln -s "$(pwd)/skill" ~/.claude/skills/fde-consultant # first Academy exercise /fde-consultant turn this vague AI idea into a 6-Q scoping interview: "We want an AI agent for customer support operations."