Practice & Reference · Reference

Reference & Certification

A GH-300 study map, a glossary, a one-page cheat sheet, and where to go next — everything to revise from and to keep within reach.

☺ Explain it like I’m 10

Think of this page as the cheat-sheet at the back of a board-game box: the quick reminders, the word list, and the “what beats what” tables all in one place, so you can look things up fast instead of re-reading the whole rulebook.

🐘🐿️Your hosts for this topic: Ellie the Elephant (remembers every term) and Nutty the Squirrel (keeps the cheat sheet stashed).

GH-300 study map

☺ Like you’re 10: A checklist of everything the big test might ask, so you can tick off what you already know and spot what still needs practice — like a packing list you check before a trip.

If you’re working toward the GH-300 GitHub Copilot certification, here’s how this course maps to what the exam emphasizes. Use it as a revision checklist — make sure you can answer each prompt cold. Aiming at Anthropic’s CCAR-F instead? The same foundations apply — that hub carries its own scenario-based study map.

Fundamentals & responsible use

Using Copilot effectively

Agents

Building & extending

Administration & governance (heavier on Business/Enterprise)

Exam tactic: the test likes “which feature for which scenario” questions. When a question describes a scenario, locate it on the autonomy spectrum first — that usually points straight at the answer. Drill three decision maps: the autonomy ladder, the four-ways-to-build-an-agent map, and agent-mode-vs-cloud-agent.
⌁ Before the exam

Certification objectives change. Confirm the current GH-300 objectives on the official certification page before sitting it — and note this course goes beyond the exam in places (the SDK capstone, multi-agent ops), which is good for real work even where it exceeds what’s tested.

Glossary

☺ Like you’re 10: A mini dictionary — every tricky word from the course explained in one line, so if you forget what one means you can look it up in a second.

🎬 At the AI Academy
🦊

Foxy: Wait — what does “agent” even mean again? I keep mixing it up with a plain chatbot!

🐘

Ellie the Elephant: I remember it word for word: an agent is Model + Tools + Memory + Planning. No tools, and it’s just a chatbot.

🐿️

Nutty the Squirrel: And it’s stashed right here in the glossary and the cheat sheet — one line, always within reach.

🦉

Professor Owl: That’s the whole point of this page: forget a term, look it up in a second, and get back to building.

TermMeaning
AgentAn LLM wrapped in a loop that can use tools, keep memory, and act on the world. Model + Tools + Memory + Planning.
Agentic AISystems that don’t just respond but plan, act, observe results, and iterate toward a goal.
Agent modeCopilot’s in-editor autonomous mode: plans, edits multiple files, runs commands, iterates. Claude’s in-editor/terminal equivalent is Claude Code.
Cloud agentCopilot’s async background worker; turns an assigned issue into a draft PR in a sandboxed Actions environment.
Claude CodeAnthropic’s agentic coding tool in your terminal/editor: plans and edits across a codebase, with project memory (CLAUDE.md) and MCP. Claude’s parallel to Copilot’s agent mode + CLI.
CoworkAnthropic’s async knowledge-work app: hand off a longer multi-step job and review the result. Claude’s parallel to Copilot’s cloud agent.
Agent HQA control plane that orchestrates coding agents from multiple providers across GitHub surfaces.
Agent SkillA SKILL.md folder teaching an agent a repeatable specialized procedure (with optional scripts).
BYOKBring Your Own Key — using your own model-provider API keys with the Copilot SDK instead of the subscription.
Copilot ExtensionAn external capability plugged into Copilot Chat via @mention, built as a GitHub App; either a skillset (you provide API endpoints, Copilot handles the AI) or an agent (you control the full interaction).
Copilot SDKA library exposing the same agent runtime behind Copilot CLI, to embed agentic workflows in your own apps (TS, Python, Go, .NET, Rust, Java). Anthropic’s counterpart is the Claude Agent SDK.
Custom instructionsAlways-on project conventions an assistant loads as context: copilot-instructions.md / AGENTS.md for Copilot, CLAUDE.md for Claude.
EvalCriteria-based testing of non-deterministic AI output; the “unit test” of AI systems; gates releases.
MCPModel Context Protocol — open standard for connecting external tools/data (“MCP servers”) to an agent. See MCP — Tools for Agents.
NESNext Edit Suggestions — predicts where your next edit will be, not just what comes after the cursor.
Premium requestA budgeted call to a frontier model; agentic, multi-step work consumes these per step.
Prompt fileA reusable, version-controlled prompt stored as Markdown in the repo.
Prompt injectionMalicious instructions hidden in content an agent reads, attempting to hijack its behavior; dangerous because tool-wielding agents can then take real actions.
ReActThe Reason+Act loop most agents use: think, call a tool, observe, repeat.
SpacesCopilot Spaces — curated project-specific knowledge bundles used to ground answers.
TokenThe unit of text a model reads/generates (~¾ of a word). Context windows and pricing are measured in tokens. See How models work.
Context windowThe maximum tokens (prompt + reply) a model can consider at once — its short-term memory. Overflow pushes older text out.
HallucinationA confident, plausible, but false model output; it predicts likely text, not verified truth. Reduced by grounding and verification.
TemperatureThe sampling randomness knob: low = focused/repeatable, high = varied/creative.
PromptingGoal + context + constraints + examples + output format — the highest-leverage, tool-transferable skill. See Prompting & Context.
Few-shotSteering a model by showing a few input→output examples rather than describing the format (zero / one / few-shot = 0 / 1 / several).
RAGRetrieval-Augmented Generation — fetch relevant documents and add them to the prompt so the model answers grounded in real data. See Retrieval & RAG.
EmbeddingA vector capturing the meaning of text, so similar meanings sit close together — enables search by meaning (semantic search).
Vector databaseStores embeddings and returns the top-k most similar to a query vector; the retrieval engine behind RAG.
GroundingBasing an answer on supplied real documents instead of the model’s memory — the main defense against hallucination.
Fine-tuningFurther training that bakes knowledge/behaviour into a model’s weights. One of three ways to add knowledge (vs. long context and RAG).
MultimodalA model that handles more than text — images, audio, video — as input and/or output. See Multimodal & Generative Media.
Diffusion modelThe architecture behind most text-to-image generation: denoise from random noise toward an image. Distinct from the next-token LLM.
Machine learning (ML)Systems that learn patterns from data instead of being explicitly programmed. A subset of AI; deep learning is a subset of ML. See The AI Landscape.
Neural networkLayers of simple units with weights learned from data; “deep learning” means many layers — the substrate under modern models.
Transformer / attentionThe architecture behind modern LLMs; attention weighs which earlier tokens matter most for predicting the next one.
AGIArtificial general intelligence — hypothetical human-level ability across any task. Today’s systems are narrow / task-specific.
Reasoning modelA model that generates step-by-step reasoning before answering, trading time for accuracy on hard problems. See Reasoning & Test-Time Compute.
Test-time computeSpending more compute at inference (more “thinking”) to improve hard answers without a bigger model — the key 2025–26 shift.
Chain-of-thought (CoT)Reasoning through intermediate steps before the answer; once a prompt trick, now native to reasoning models.
Agent memoryContinuity for agents: working memory (the context window) plus long-term episodic / semantic / procedural memory in a store. See Agent Memory.
JailbreakCoaxing a model past its safety guardrails to produce disallowed output. See AI Security.
Lethal trifectaPrivate data + untrusted content + the ability to act = the data-exfiltration danger zone for agents.
Responsible AIFairness, transparency, privacy, accountability, human oversight, and safety — engineering practice, not PR. See Responsible AI & Safety.
Automation biasOver-trusting confident machine output; a key reason to keep humans reviewing consequential AI decisions.

One-page cheat sheet

☺ Like you’re 10: The whole course squished onto one page of quick reminders — the kind of note you’d love to have taped to your desk during the test (if you were allowed one!).

The autonomy ladder: Completions → Ask → Edit → Agent mode → Cloud agent → Multi-agent orchestration. (Claude’s rungs: chat → Claude Code → Cowork → multi-agent — same ladder, different names.) Pick the rung that matches how much you want to hand off.

Which chat mode? Ask = understand. Edit = change what you know. Agent = do what you describe. Scope in Ask, execute in Agent.

Agent mode vs. cloud agent? Agent mode = synchronous, in your editor, you watch. Cloud agent = asynchronous, in the cloud, returns a draft PR. Claude parallel: Claude Code (synchronous) vs. Cowork / a cloud agent (asynchronous).

An agent = Model + Tools + Memory + Planning. No tools → it’s just a chatbot.

Model basics: it predicts the next token; measure context and cost in tokens; low temperature = repeatable, high = creative; it hallucinates when it lacks a fact — so ground and verify.

Give a model your knowledge: long context (stuff it in the window) · RAG (retrieve just what’s relevant — the usual pick) · fine-tune (bake into the weights). Grounding beats a bigger model for “it doesn’t know my stuff.”

Reasoning vs speed: for hard math / logic / planning / tricky debugging, reach for a reasoning model (it thinks step-by-step — test-time compute); for lookups and formatting, a fast standard model. Dial reasoning effort up only when accuracy is worth the latency and cost.

Four ways to build an agent: ① Agent Skill (Markdown) → ② Custom Agent (cloud-agent config) → ③ Copilot Extension (GitHub App: skillset or agent) → ④ Copilot SDK (embed the runtime). With Claude the same rungs are: a Skill / CLAUDE.md → a specialized config → an MCP server → the Claude Agent SDK. Pick the lowest rung that solves it.

Skillset vs. extension-agent? Skillset = you give ≤5 API endpoints, Copilot does the AI (start here). Agent = you control prompts/model/flow (more power, more work).

Multi-agent patterns: orchestrator–worker, pipeline, reviewer/debate, hierarchical. Use when tasks genuinely decompose; don’t over-decompose.

The two pipelines: A = agents help ship your software (issue→plan→code→review→test→deploy). B = your AI product’s own request-time flow (input→context→model→tools→response→log).

New AI pipeline stages: evals (replace exact-match tests), prompt/version management, model routing + fallbacks, tracing/observability, AI-specific security (prompt injection, OWASP ASI).

Cost levers: right-size the model (biggest lever), scope tasks tightly, monitor usage weekly, mind both premium requests and Actions minutes.

Guardrails (always): human gate on irreversible actions, least-privilege tools, sandboxing, full audit/trace, treat external content as untrusted.

AI security reflexes: everything the model reads is untrusted data, not commands (prompt injection); beware the lethal trifecta (private data + untrusted content + ability to act); least privilege + human gates + the OWASP LLM/Agentic Top-10 before prod.

Responsible AI: fairness, transparency, privacy, accountability, human oversight, safety. Match autonomy to blast radius; watch for automation bias; you own what you ship.

The ops flywheel: production failure → add it as an eval → fix prompt/tool → eval gate passes → redeploy. Failures become permanent tests.

Which model when

SituationReach for…
Everyday coding, best balanceAuto — let Copilot route by task & availability
Hard multi-file reasoning, architecture, tricky bugsA top frontier reasoning model (Claude Opus / GPT-5-class); raise reasoning effort
Fast, simple, high-volume editsA fast/cheap model (or just Auto)
Refactor across a very large codebaseA 1M-context model with extended context
Agentic / tool-using taskA tool-calling-capable model (local: Llama 3.1+, Qwen2.5-Coder, Phi-4)
Sensitive/proprietary code — nothing leaves the machineLocal model via BYOK (Ollama / Foundry Local)
Offline / air-gappedLocal model + VS Code BYOK (no completions or semantic search)
Out of premium-request budget / cost-sensitiveA 0×–1× model, Auto, or a local model
Inline completions (ghost text)The completions model — set separately; local not supported

Downloads & study aids

Downloadable files (in the course zip):

Interactive study aids (built into the course):

(In the single-file version, the interactive aids are in the sidebar; the downloadable files live in the multi-page course zip.)

Where to go next

Beyond this course — advanced & adjacent AI

☺ Like you’re 10: This course taught you to drive and build with AI helpers. There’s a whole other side — learning to build the engines themselves, plus kinds of AI that aren’t chatbots at all. Here’s the map of what’s next if you want to go deeper.

This course deliberately focused on applied, agentic AI — using and building with LLM assistants. That’s one important slice of a much bigger field. The areas below were intentionally left out of scope; when you’re ready to go wider or deeper, this is where each leads. (🦎 Cami the Chameleon’s world, it turns out, is bigger than one course.)

Advanced / adjacent areaWhat it isReach for it when…
Classical ML & data scienceRegression, classification, clustering, decision trees, feature engineering — the non-LLM machine learning that still powers most production systems.You need prediction on structured / tabular data, not text generation.
Deep learning from scratchThe math and craft under the models: neural nets, backpropagation, gradient descent, PyTorch / TensorFlow.You want to understand and build models, not just use them — the theory beneath How AI models work.
Training & fine-tuning your own modelsPretraining, fine-tuning, LoRA / PEFT, RLHF & alignment, distillation, quantization.RAG and prompting aren’t enough and you must bake knowledge or behaviour into the weights.
Computer vision & classic NLP (as fields)CNNs, object detection, segmentation; tokenization, parsing, and traditional NLP pipelines — beyond what a general multimodal model gives you.You need specialized perception or language tooling a general model can’t do well.
Reinforcement learningAgents that learn from reward through trial and error — games, robotics, control, and the basis of RLHF.Behaviour must be learned from feedback rather than prompted or retrieved.
MLOps & AI infrastructureModel serving, GPUs, scaling, deployment, and monitoring model drift — the infra beneath production models (this course covered agent ops, not model-serving infra).You operate models — not just agents — at scale.
Data engineering for MLPipelines, labeling, feature stores, synthetic data, and data quality — 🐿️ Nutty’s deeper world.Your models are only as good as the data feeding them.

None of these are prerequisites for what you learned here — you can build real, useful agentic AI without them. But they’re the natural next mountains to climb. The Further Reading library and the wider landscape are good on-ramps.

A closing word

You started thinking these assistants were autocomplete. You’re ending able to build, ship, and operate an autonomous AI system. The single most important idea to carry forward is the one we opened with: everything is a point on the autonomy spectrum, and the engineering skill is choosing how much to hand off and how to stay safely in the loop. Completions hand off a line. The cloud agent hands off a task. A multi-agent system hands off a workflow. Your job — as the human — is never going away; it’s moving up the ladder from writing every line to directing, reviewing, and operating the systems that do.

Build the capstone. Reading about agents teaches you the vocabulary; operating one teaches you the judgment. Good luck.