Beginner Glossary (30 terms) + I'm Stuck Guides (23)

The thirty terms cover the course's whole vocabulary: foundations (LLM, model, token, embedding, self-attention), prompting and context (prompt, context window, context engineering, system instructions), quality (hallucination), knowledge systems (RAG), the Claude workspace (Project, Project instructions, knowledge file, artifact, Skill), assistants and agents (bot, custom assistant, AI agent, AI employee), integrations and security (API, API key, credential/OAuth), automation (workflow, n8n, node, trigger, JSON), and delivery (vibe coding, GitHub/version control).

The definitions are precision-tooled for beginners: 'the model is not the whole chat application'; 'anyone with the key may be able to use the associated service and spend its allowance'; 'a bot may retain instructions or knowledge but does not automatically become autonomous'; ''AI employee' is a business framing — the useful part is the operating design, not the label.'

The paid-plan guide is honest about subscriptions: long inputs, long outputs, large files, retries and parallel agents all consume capacity regardless of plan wrapper. The disciplines: decide whether the task needs the full source or selected sections, ask for a plan before a long run, limit output length and alternatives, test one item before scaling, and reuse a concise approved summary instead of re-supplying the same large material.

The context guide completes it: use the smallest complete evidence set for the CURRENT task; separate facts, examples, constraints and open questions; keep a short approved project brief and decision log; start a fresh chat when the conversation holds unrelated work. The troubleshooting prompt makes the model sort supplied context into required / optional / unrelated / missing — 'do not solve the task yet.'

The drift fix: state the goal in one sentence, separate background from requirements, NUMBER every non-negotiable, specify the output structure, state what must not be added, and require a final compliance check. The audit prompt closes the loop: 'list every mismatch first, then return a corrected answer that introduces no additional scope.'

The reuse taxonomy settles a real beginner confusion: a template is a fill-in-the-blanks shape; an instruction is a behavior rule; a Project is a workspace grouping one job's chats, files and rules; a Skill is a reusable method that should work across jobs. The promotion path: start as instructions inside one Project, package as a Skill only after it proves reusable — and test any template on three different inputs before trusting it.

Three guides form one safety doctrine. Hallucination control: identify which claims require evidence, supply approved sources, require flagging of missing facts instead of gap-filling, save output as a draft, and verify before anything publishes. Private data: classify before upload, redact or use fictional data while learning, check retention/training/sharing/deletion controls, smallest permissions, human approval on high-impact actions, and the responsible privacy owner involved in regulated settings.

Support automation gets its own boundary map: automate intake, classification, routing, summarization and drafting FIRST; substantive replies stay behind human review until policy, quality, escalation and monitoring are proven; auto-send only a narrow approved low-risk acknowledgement, if authorized at all. The audit prompt hunts every point that 'can send, promise, refund, delete, disclose data, or hide an urgent case.'

Five guides cover the life of an AI-built app after the demo. Generic design: give a design system (palette, type scale, spacing, max width) and fix ONE visual dimension at a time — the audit prompt asks a 'senior product designer' for the five highest-impact reasons it looks generic, with one small testable change each. Maintenance: version history, tests, docs, backups, rollback — reproduce a failure before requesting a fix, demand root-cause analysis and a minimal patch, verify in preview, roll back regressions.

Architecture literacy: map frontend/backend/database/files/auth/APIs/secrets/repo/hosting/domain, with provider, owner account, backup path and unresolved risk per layer — 'each one still needs an owner and recovery plan.' Portability: GitHub stores code, but portability also depends on env vars, proprietary services and a clean-environment build test. And the bug-report guide is the master pattern: expected, observed, steps, exact error, last working state, last change — 'do not make changes yet.'

The vocabulary guide de-mystifies: workflow = the whole automation, node = one step, trigger = the starting event — 'software-workflow concepts, not neural-network layers' (a real learner confusion). Credentials doctrine: keys live in the credential manager, attach to compatible nodes, one provider's key is not valid for another's node, and diagnosis runs provider → billing → permissions → expiry → documentation, in that order.

The Gmail/Sheets guide isolates integration failures: confirm the owning account, reconnect, approve only needed scopes, test the trigger ALONE, the draft ALONE, the append ALONE, match exact document/tab/header names, and inspect the first empty or red node — 'find the first broken assumption. Do not redesign the workflow.' Cost estimation separates line items: hosting, executions, model calls, input/output volume, connected services, safety margin — with self-hosting allowed only if you can own updates, backups, secrets and uptime.

The beginner-plan guide is the portal's own priority ordering: pick a task you already understand, define input and output, write a five-line requirements brief, build the MANUAL AI-assisted version first, test on three examples, save the working prompt, and automate only after the manual version is reliable.

The reduction prompt does the scoping for you: 'one user, one input, one output, one tool, three acceptance checks, and what must remain manual — do not add automation or agents yet.' It's the antidote to tool-list paralysis, and it's the same smallest-working-version doctrine the Practice Lab, the roadmap's deterministic-first sequence, and the workbooks all repeat.