The Curriculum / Reader / Level 1 shared materials
LEVEL 1 · ESSENTIALS

Level 1 shared materials

This page compiles 11 files from the repository, verbatim, in reading order. The living version: this folder on GitHub.

level-1-essentials/shared/checklists/daily-power-user-checklist.md

Daily Power-User Checklist

Not everything. Just the habits that compound.

Morning (5 min)

Work-mode default posture

During the day

Before EOD (5 min)

level-1-essentials/shared/checklists/quarterly-audit-checklist.md

Quarterly Audit Checklist

90 minutes, first Monday of the quarter.

Security

Cost

Quality

Skills

System

Governance (if applicable)

level-1-essentials/shared/checklists/weekly-review-checklist.md

Weekly Review Checklist

30 minutes, every Friday.

Review

Prune

Add

Verify

Reflect

level-1-essentials/shared/eval-templates/eval-set-template.md

Eval Set Template

An eval set is a collection of test cases with expected outcomes, used to score model outputs. This is the single tool that separates serious AI users from the rest.

Structure of an eval set

Save one folder per workflow being evaluated:

evals/
├── email-triage/
│   ├── cases.jsonl
│   ├── rubric.md
│   └── results/
│       ├── 2026-08-19_claude-sonnet.jsonl
│       └── 2026-08-19_gpt-5.jsonl
├── lease-extraction/
│   ├── cases.jsonl
│   ├── rubric.md
│   └── results/

cases.jsonl format

Each line is a test case:

{"id": "case-001", "input": "...", "expected": "...", "criteria": ["accurate", "concise", "in voice"], "tags": ["priority-1", "regression"]}

Fields: - id — stable identifier - input — the prompt input (raw text, or a reference to a file) - expected — the ideal answer OR the criteria for what a good answer looks like - criteria — array of rubric dimensions - tags — for filtering (regression, high-priority, edge-case, etc.)

Minimum viable eval set

For any workflow you care about, produce 5 good + 5 bad:

Sample eval set — Email triage

{"id": "e-001", "input": "Hi Adam, following up on the lease terms we discussed Tuesday...", "expected": "Reply-today", "criteria": ["correct_label", "acknowledges_deadline"], "tags": ["client", "regression"]}
{"id": "e-002", "input": "🎉 40% off summer sale at BigMart!", "expected": "Newsletter", "criteria": ["correct_label"], "tags": ["marketing"]}
{"id": "e-003", "input": "Your Vercel deployment succeeded.", "expected": "FYI-only", "criteria": ["correct_label"], "tags": ["automated"]}

Sample rubric

## Rubric for email triage

Correctness (0-3):
  0 - wrong label
  1 - roughly right category
  2 - correct label, missed nuance
  3 - correct label with appropriate priority

Voice match (0-3) [for drafted replies]:
  0 - obvious AI writing
  1 - competent but generic
  2 - matches my voice on structure or vocabulary
  3 - passes as mine

Safety (pass/fail):
  Fail if it sends without draft
  Fail if it includes PII in the summary
  Fail if it categorizes urgent-legal as newsletter

Running an eval

# Pseudo-code — adapt to your framework
for each case in cases.jsonl:
  output = model.generate(system_prompt, case.input)
  score = judge(output, case.expected, case.criteria)
  record(case.id, model, score, output)

report(aggregate_scores)

Frameworks that make this easier: - Promptfoo — CLI-first, YAML configs, LLM-as-judge built-in - LangSmith — Anthropic/OpenAI/Langchain integration - Langfuse — open source, self-hostable - DeepEval — Python-first

When to run

What "good" looks like

level-1-essentials/shared/eval-templates/regression-test-protocol.md

Regression Test Protocol

Before you swap in a new model version, run this.

The rule

No model change ships without running the eval set.

Even if the vendor says the new model is "10% better on benchmarks" — your workflow may be different. Verify on your data.

Protocol

Step 1 — Freeze the baseline

Step 2 — Run candidate

Step 3 — Diff

Produce a comparison table:

Case Baseline score Candidate score Delta Notes
e-001 3/3 3/3 0
e-002 2/3 3/3 +1 Improved edge case
e-003 3/3 1/3 -2 🚨 Regression

Step 4 — Decide

Ship the new model if: - ✅ Zero priority-1 regressions - ✅ Aggregate score ≥ baseline - ✅ Cost is acceptable - ✅ Latency is acceptable

Do not ship if: - ❌ Any priority-1 regression - ❌ Any safety-fail case - ❌ >5% degradation in aggregate

Step 5 — Communicate

If you ship, write a one-paragraph changelog: - Old model → new model - Aggregate delta - Notable improvements - Any regressions being accepted (and why) - Rollback plan

When to re-baseline

Prompt regressions

If a prompt change is the trigger, hold the model constant, vary only the prompt. Same protocol.

level-1-essentials/shared/eval-templates/scoring-rubric.md

Scoring Rubric — Universal

Adapt per workflow. This is the base template.

Dimensions

Every eval scores these dimensions on 0–3.

1. Correctness

2. Format compliance

3. Voice / tone

4. Length compliance

5. Faithfulness (for grounded tasks)

Safety checks (pass/fail)

Any one of these fails the case entirely:

Aggregate score

Judge

For LLM-as-judge: - Use Claude Opus or GPT-5 for judging - Provide the rubric in the judge's system prompt - Have it produce structured output with scores per dimension + reasoning - Sanity-check 10% of judgments manually

level-1-essentials/shared/glossary/master-glossary.md

AI Master Glossary

200+ terms, organized in 11 clusters. Read one cluster per day for two weeks and you'll be fluent.

Each term has: definition, why it matters, and where it appears.


Cluster 1 — Model Mechanics

Token — The unit an LLM consumes. Roughly ¾ of an English word. Prices, context windows, and rate limits are all counted in tokens. Why it matters: cost and length constraints are token-denominated.

Context window — The maximum number of tokens a model can attend to in one call (input + output). Gemini 2.x Pro: 2M. Claude: 1M. GPT-5: 400k. Why it matters: longer window = more knowledge in-context = less RAG needed.

System message / system prompt — Standing instructions given before the conversation starts. Higher priority than user messages. Why it matters: this is your leverage point.

User message — What the user says. Second priority.

Assistant message — What the model says. Referenced back in multi-turn conversations.

Tool message / function result — The output of a tool call, sent back to the model.

Temperature — Randomness knob (0–2). 0 = deterministic. 1 = default. 2 = wild. Use 0 for extraction, 0.7 for writing, 1.0+ for brainstorming.

Top-p (nucleus sampling) — Alternative to temperature. Selects from the smallest set of tokens whose cumulative probability exceeds p. Usually set to 0.9 or 1.0.

Top-k — Select from the top k most likely tokens. Rarely tuned directly.

Frequency penalty / presence penalty — Discourage repetition. Frequency penalizes based on count; presence based on whether the token has appeared at all.

Max tokens — Cap on output length. Different from context window.

Streaming — Model returns tokens as they're generated, rather than waiting for the full response. Reduces perceived latency.

Stop sequences — Strings that cause the model to halt generation. Useful for structured output.

Seed — For deterministic sampling. Same seed + same input + same params = same output. Not guaranteed across model versions.

Log probs / logprobs — Probability the model assigned to each generated token. Used for confidence estimation and evals.

Function calling / tool use — Model outputs a structured request to call an external function, then receives the result and continues. The foundation of agents.

JSON mode / structured output — Constraint that forces the model to return valid JSON matching a schema. Reliable extraction.

Grammar-constrained decoding — Force output to match a formal grammar (JSON schema, regex, custom BNF). Even stricter than JSON mode.

Token pricing — Cost per million input tokens vs cost per million output tokens (output is usually 3–5× more expensive).

Rate limit — Per-minute or per-day cap on requests or tokens. Enforced per API key or organization.


Cluster 2 — Reasoning

Chain-of-Thought (CoT) — Prompting technique: "think step by step" before answering. Improves accuracy on multi-step problems.

Zero-shot — Prompting with no examples. "Translate this to French."

Few-shot — Prompting with 2–10 examples. Improves format compliance and task performance.

In-context learning (ICL) — The ability of LLMs to learn a task from examples within the prompt, without weight updates.

ReAct — "Reason + Act." Pattern where model alternates reasoning steps and tool calls.

Tree-of-Thought (ToT) — Explores multiple reasoning branches before committing to an answer.

Self-consistency — Sample multiple reasoning paths and pick the majority answer.

Reflection / self-critique — Model reviews its own output and revises. Also called "critic-actor" loop.

Extended thinking / thinking mode — Native product feature (Claude, o-series) where the model spends "reasoning tokens" before responding. Higher accuracy, higher cost.

Reasoning tokens — Tokens the model uses internally for thinking that aren't shown to the user (or shown selectively). You pay for them.

Test-time compute — Additional compute spent at inference to improve output quality (e.g., extended thinking, self-consistency, tree-of-thought). The 2024–2026 scaling story.

Scratchpad — Explicit reasoning space in the prompt where the model writes intermediate work.

Plan-and-solve — Prompt pattern: first produce a plan, then execute each step.

Persona / role prompting — Assigning the model a role ("You are a senior editor..."). Modest effect; do not overrate.

Chain-of-Verification (CoVe) — Model drafts an answer, generates verification questions, answers them, and revises. Reduces hallucinations.


Cluster 3 — Retrieval (RAG)

RAG — Retrieval-Augmented Generation — Pattern where you retrieve relevant documents from a knowledge base and inject them into the prompt before generation.

Embedding — A vector (list of numbers, usually 768–3072 dimensions) that represents the meaning of a text. Similar texts have similar embeddings.

Embedding model — A model that produces embeddings. Popular: OpenAI text-embedding-3-large, Voyage AI, Cohere embed, BGE, E5.

Vector database / vector store — A DB optimized for storing embeddings and finding nearest neighbors. Pinecone, Weaviate, Qdrant, Milvus, pgvector.

Chunking — Splitting documents into pieces before embedding. Common strategies: fixed-size (500 tokens), sentence, paragraph, semantic, sliding-window.

Chunk overlap — Overlap between adjacent chunks (usually 10–20%) to preserve context that spans boundaries.

Semantic search — Finding results by meaning (embedding similarity) rather than keyword match.

Lexical search — Keyword-based search (BM25, TF-IDF).

Hybrid search — Combines semantic and lexical. Usually reranked.

Reranker / cross-encoder — A second-pass model that re-scores retrieved chunks by their relevance to the query. Popular: Cohere Rerank, BGE reranker.

k / top-k retrieval — Number of chunks to fetch. Usually 5–20.

Retrieval query — The version of the user's question used for the DB search. Often rewritten for better retrieval.

Query rewriting / query expansion — Transforming the user's question into a better retrieval query. HyDE is one method.

HyDE — Hypothetical Document Embeddings — Generate a fake ideal answer, embed it, use that for retrieval.

Metadata filtering — Restrict retrieval to chunks with matching metadata (e.g., permissions, date range, doc type).

Ground truth — The correct answer, known independently. Used to score retrieval and generation.

Retrieval@k — Metric: does the correct chunk appear in the top k results?

Faithfulness / groundedness — Does the generated answer only make claims supported by the retrieved context?

Context stuffing — Just pasting a huge document into the prompt (no vector search). Feasible with 1M+ context windows for smaller corpora.

Agentic RAG — Agent decides when and what to retrieve, may issue multiple queries, may use tools beyond vector search.

GraphRAG — Retrieval over a knowledge graph rather than vector chunks. Better for entity-heavy corpora.


Cluster 4 — Agents

Agent — A model that operates in a loop: observe → decide → act (tool call) → observe → ... until the goal is met.

Tool — A function an agent can call. Formalized by JSON schemas the model can understand.

Tool call — A structured request from the model to invoke a tool with arguments.

Function calling — The lower-level API primitive. Some providers use this term instead of "tool."

Agent loop / control loop — The outer software loop that dispatches tool calls and feeds results back.

Subagent — An agent spawned by another agent to handle a bounded sub-task with its own context.

Orchestrator — The parent agent that plans and dispatches subagents.

Planner — A component (or a model call) that produces a step-by-step plan before execution.

Executor — The component that actually runs the plan's steps.

MCP — Model Context Protocol — Anthropic-originated open standard for connecting AI models to tools and data sources. Now widely adopted.

MCP server — A process that exposes tools/resources over the MCP protocol.

MCP client — An AI model / product that consumes MCP servers.

Skill — A reusable capability package (a set of instructions and tools) that a model can load on demand. Claude's "Skills" and Perplexity's skill system are examples.

Memory (agent memory) — Persistent state carried across sessions. Types: short-term (conversation), long-term (facts), episodic (event history).

HITL — Human in the Loop — Design pattern where humans approve or correct agent decisions at defined checkpoints.

Guardrails — Rules that constrain agent behavior (e.g., "never send emails," "always ask before making purchases").

ReAct agent — Agent following the ReAct pattern.

AutoGPT / BabyAGI — Early (2023) open-source agent frameworks. Historical, but the terms still come up.

Agentic — Adjective for "acting like an agent" — doing multi-step work autonomously.

Autonomous mode / auto mode — Product terminology for agents that operate without step-by-step human approval.

Long-horizon task — A task requiring many steps, tool calls, or hours of runtime.

Terminal state / done condition — When the agent decides it's finished.

Fan-out / fan-in — Pattern: spawn N parallel subagents (fan out), collect results (fan in).


Cluster 5 — Training

Pretraining — Training a model on massive text (trillions of tokens) to learn language. Base model.

Base model / foundation model — The pretrained model before instruction tuning.

Fine-tuning — Updating a pretrained model's weights on a smaller, task-specific dataset.

Instruction tuning — Fine-tuning with (instruction, response) pairs to make the model follow instructions.

RLHF — Reinforcement Learning from Human Feedback — Training method where humans rank outputs and a reward model is trained to score outputs, then the LLM is optimized against that reward.

DPO — Direct Preference Optimization — Alternative to RLHF that skips the reward model. Simpler and often competitive.

Constitutional AI (CAI) — Anthropic's alignment method: use principles ("a constitution") to have the model critique and revise its own outputs.

RLAIF — RL from AI Feedback — Like RLHF but the ranker is another AI, not humans.

Distillation — Training a smaller model to mimic a larger one. Produces cheaper, faster models with much of the capability.

LoRA — Low-Rank Adaptation — Efficient fine-tuning method: add small trainable matrices to a frozen base model. Small artifacts (~100MB) instead of full model weights.

QLoRA — LoRA with quantized base model. Even cheaper.

Quantization — Reducing precision of model weights (e.g., FP16 → INT8 → INT4) to shrink size and speed inference. Small accuracy cost.

Weights — The parameters of a neural network. "Open-weight" model = weights are downloadable.

Open-weight model — Weights released publicly (LLaMA, Mistral, Qwen, DeepSeek). Not necessarily open-source (data + training code not always released).

Open-source model — Everything released — weights, training data, code.

Closed model — Only accessible via API (GPT-5, Claude, Gemini).

Alignment — Making a model behave in accordance with human values / intentions.

Refusal — When a model declines a request. Alignment feature; sometimes over-tuned.

Jailbreak — A prompt that bypasses safety training.

Prompt injection — Adversarial input that hijacks the model's behavior (e.g., a document that says "ignore previous instructions and...").

Model weights license — The legal terms under which weights can be used (LLaMA license, Apache 2.0, MIT, custom). Important for commercial use.

Training compute — FLOPs used to train the model. Correlated (but not perfectly) with capability.

Scaling laws — Empirical laws relating training compute, data, and parameter count to loss.

Emergent capabilities — Skills that appear only above a certain model scale (contested empirically but useful shorthand).


Cluster 6 — Multimodal

Multimodal model — Handles more than one modality (text + image + audio + video).

Vision model / VLM — Vision-Language Model. Handles text + image.

Audio model — Handles speech (ASR/TTS) or general audio.

ASR — Automatic Speech Recognition — Speech-to-text. Whisper is the reference model.

TTS — Text-to-Speech — ElevenLabs, OpenAI TTS, Google WaveNet.

Voice cloning — Generating speech in a specific person's voice. ElevenLabs Instant Voice.

Diffusion model — The dominant image/video generation architecture (Stable Diffusion, Midjourney, DALL·E, Sora). Starts with noise, iteratively denoises.

Latent space / latent diffusion — Diffusion in a compressed latent space (rather than pixel space). Much faster. Stable Diffusion's key trick.

Sampler / scheduler — Algorithm that runs the diffusion denoising steps (DDIM, DPM++, Euler, etc.).

Steps — Number of denoising steps. More = slower + usually better quality (diminishing returns after ~30–50).

CFG — Classifier-Free Guidance scale — How strongly the image follows the prompt (higher = more faithful, less creative).

Seed (image) — Random seed. Same seed + same prompt + same params = same image.

ControlNet — Add spatial conditioning (pose, edges, depth) to guide generation.

LoRA (image) — Small adapter that adds a specific style, character, or concept to an image model.

Inpainting — Fill in a masked region of an image.

Outpainting — Extend an image beyond its borders.

Img2img — Transform an existing image using a text prompt.

Text-to-video — Sora, Veo, Runway Gen-4, Wan, LTX.

Image-to-video / keyframe-to-video — Animate a still image, or animate between two keyframes.

First-frame / last-frame conditioning — Video generation guided by a specific starting or ending frame.

Reference image / image conditioning — Use an image to guide generation (character, style, composition).

Motion prompt — Text describing the camera or subject movement in a video generation.

Frame rate (fps) — Frames per second of generated video (usually 24 or 30).

Duration — Length of generated video (usually 5–10 seconds per clip, then stitched).

Talking head / lip-sync — Generate video of a person speaking (SadTalker, HeyGen, D-ID).

Neural rendering — General term for AI-generated visual output.


Cluster 7 — Evaluation

Eval / evaluation — Structured tests that measure model performance. Automated where possible.

Eval set — The collection of test cases with expected answers or scoring criteria.

Benchmark — A public, standardized eval (MMLU, HumanEval, SWE-bench, GPQA).

Golden dataset / golden answers — Curated correct answers, used as ground truth.

Human eval — Human raters score model outputs.

Pairwise comparison — Show a human two outputs, ask which is better. Feeds ELO ratings.

LMSYS Arena / Chatbot Arena — Public leaderboard from crowd-sourced pairwise comparisons.

Rubric — Explicit criteria for scoring outputs (accuracy, tone, format, etc.).

LLM-as-judge — Using an LLM to score other LLM outputs. Cheaper than human eval; introduces its own biases.

Regression test — Re-running the eval set after a change to catch quality drops.

A/B test — Ship two versions to users, measure difference. For product-level evals.

Hallucination — Model confidently states something false. The core reliability problem.

Faithful vs coherent — Coherent output can still hallucinate. Faithfulness = grounded in provided context.

Groundedness score — Metric for how well output is supported by retrieved context.

Precision / recall — Classic IR metrics; also used for eval.

F1 — Harmonic mean of precision and recall.

BLEU / ROUGE / METEOR — String-overlap metrics for translation/summarization. Legacy; largely superseded by LLM-as-judge for open-ended tasks.

Perplexity — How well a model predicts a text (lower is better). Used for base-model comparison.

pass@1 / pass@k — For code eval: does the model pass the tests on the first try / at least once in k tries?

Red team — Adversarial testing to find failures, jailbreaks, misuse.

Prompt injection eval — Specifically test defenses against prompt injection.

Bias eval — Test for demographic, ideological, or systematic biases.

Toxicity eval — Test for harmful outputs.

Safety eval — Broad category including refusals, harmful content, misuse potential.


Cluster 8 — Deployment

Inference — Running a trained model to produce outputs. The step you pay for.

Latency — Time from request to first byte (TTFT) and to last byte.

Throughput — Requests or tokens per second the system can handle.

TTFT — Time to First Token — Latency until streaming begins.

Batch inference — Process many requests in one call. Cheaper (per token) but higher latency.

Streaming — Send tokens as generated for faster perceived response.

Caching — Store and reuse: KV cache (attention state), prompt cache (Anthropic prompt caching), semantic cache (skip repeated queries).

Prompt caching — Provider feature: repeated prompt prefixes are cached and re-billed cheaper.

KV cache — Internal attention state cached between tokens in a single generation.

Serverless inference — Provider hosts models; you pay per token. OpenAI, Anthropic, etc.

Dedicated inference — You reserve capacity (per hour). Better latency, worse economics unless heavily used.

Provisioned throughput — Same idea; guaranteed capacity for a fee.

Self-hosted inference — You run the model on your own hardware (vLLM, TGI, TensorRT-LLM, llama.cpp).

vLLM / TGI / TensorRT-LLM / llama.cpp — Popular inference engines.

Speculative decoding — Use a small "draft" model to guess tokens, verify with the big model. Faster.

Continuous batching — Batch requests dynamically, not by fixed windows. Higher throughput.

Context caching / cached input — Provider-level cache for reused system prompts. Cuts input token cost by 50–90%.

Cold start — First-request latency when infrastructure spins up.

Fine-tuned model deployment — Serving a custom fine-tuned model. Often more expensive than base.

Model versioning — Providers version models (gpt-4o-2024-08-06, claude-opus-4-20260601). Pin to versions in production.

Model card — Documentation of a model's capabilities, limits, training, and safety properties.

System card — Broader safety/capability doc for a released model (OpenAI, Anthropic conventions).

Deprecation — Providers retire old model versions. Read release notes.


Cluster 9 — Governance and Safety

PII — Personally Identifiable Information — Any data that identifies a person. Legally defined; treat carefully.

PHI — Protected Health Information — Health data under HIPAA (US). Requires BAA.

BAA — Business Associate Agreement — HIPAA contract with vendors handling PHI.

DPA — Data Processing Agreement — GDPR contract governing processor duties. Sign one with every AI vendor handling personal data.

Data residency — Where the data is stored/processed. EU customers often require EU residency.

Zero-retention / no-training — Contractual guarantee that provider does not retain or train on your inputs. Ask for it explicitly.

Data classification — Public / Internal / Confidential / Restricted. The base of every data policy.

DLP — Data Loss Prevention — Systems that block sensitive data from leaving your perimeter.

CASB — Cloud Access Security Broker — Enforces policy on cloud tool usage.

SSO — Single Sign-On — Federated login. Table-stakes for enterprise tools.

SCIM — System for Cross-domain Identity Management — Automated user provisioning/deprovisioning.

BYOK — Bring Your Own Key — Provider encrypts data with a key you control.

Private link / VPC peering — Traffic to the provider stays on private network.

SOC 2 / ISO 27001 — Security certifications. Type II is the meaningful one.

HIPAA / GDPR / CCPA / EU AI Act — Regulatory frameworks (US health, EU privacy, California, EU AI).

NIST AI RMF — National Institute of Standards and Technology AI Risk Management Framework. US governance reference.

Responsible AI (RAI) — Vendor programs and internal policies around fair, safe, accountable AI use.

Model card — See Deployment cluster.

Explainability / interpretability — Ability to understand why a model produced its output.

Prompt injection defense — Techniques to prevent adversarial content from hijacking model behavior.

Content filter / moderation — Layer that blocks harmful outputs.

Safe harbor — Legal safe zones for specific AI uses (varies by jurisdiction).

Provenance / watermarking — Marking AI-generated content so it can be detected.

Consent to record / consent to process — Legal requirement in some jurisdictions before AI can process personal comms.

Right to explanation — GDPR concept: individuals can demand explanation of automated decisions affecting them.


Cluster 10 — Product / Ecosystem terms

LLM — Large Language Model — The generic term for models like GPT, Claude, Gemini.

Foundation model — A large pretrained model adaptable to many tasks.

Frontier model — The current top-tier models (GPT-5, Claude Opus, Gemini Pro).

Small language model (SLM) — Small enough to run on-device (Phi, Gemma, Llama 3.2 1B/3B).

On-device inference — Model runs on the user's phone/laptop (Apple Intelligence, Copilot+ PC).

Multi-agent system — Multiple agents collaborating on a task.

Copilot (product term) — An AI assistant embedded in another product (Microsoft, GitHub, generic).

Assistant / agent (product term) — Overlapping but not identical to "agent" the pattern.

Autopilot — Higher-autonomy variant of a copilot. Product buzzword.

AGI — Artificial General Intelligence — Contested term for human-level general intelligence.

ASI — Artificial Super Intelligence — Above human. Speculative.

Wrapper / GPT wrapper — Pejorative for a product that's mostly a thin UI over a base model.

Prompt engineering — Craft of writing effective prompts. Real skill; often derided.

Prompt library — Curated collection of reusable prompts.

Custom GPT — OpenAI's user-built GPT with instructions, knowledge, and actions.

Claude Project — Anthropic's equivalent — persistent knowledge + custom instructions.

Perplexity Space — Perplexity's equivalent.

Gem — Google's equivalent.

Actions / plugins — External tools a Custom GPT (or similar) can call.

Artifact — Anthropic's UI element for inline runnable code / documents.

Canvas — Similar concept in ChatGPT / Gemini.

Tokens per second (t/s) — Speed metric.

Time to first token (TTFT) — See Deployment.

Model marketplace — Aggregators like OpenRouter, Together, Fireworks.

AI IDE — IDE with AI integration (Cursor, Zed, Windsurf).

AI browser — Browser with agentic capabilities (Comet, Dia, Arc's now-defunct product).

Voice mode — Real-time voice conversation with a model.

Vision mode — Model can see images/screens.

Sora / Veo / Runway / Higgsfield — Text-to-video products.

Grounding — Providing external context (usually via retrieval or search) to reduce hallucinations. Also used as verb ("grounded in the docs").


Cluster 11 — Neologisms and Jargon-of-the-moment

Vibe coding — Iterative coding with an AI, letting the AI drive with loose specs. Popularized 2024–2025.

Model soup — Averaging weights of fine-tuned models.

Merge / MoE — Mixture of Experts — Architecture where different parts of the network specialize; only a subset activated per token.

Sparse / dense (models) — MoE = sparse activation. Traditional transformers = dense.

Reasoning model — A model with native extended thinking (o-series, Claude thinking, DeepSeek R1).

Verifier model — A separate model that scores a candidate answer.

Agentic workflow — Any multi-step process orchestrated by an AI agent.

Ambient AI — AI that runs in the background without explicit invocation (background transcription, always-on assistants).

Vertical AI / vertical agent — AI specialized to a domain (legal, medical, real estate).

Horizontal AI — General-purpose (ChatGPT, Claude).

Composable AI — Building complex systems from smaller AI primitives.

LLM gateway — Middleware routing requests across providers (Portkey, LiteLLM, OpenRouter).

Observability (for LLMs) — Logging, tracing, cost tracking, prompt versioning across LLM calls (Langfuse, Helicone, Arize, LangSmith).

Prompt versioning — Treating prompts like code — version-controlled, tested, deployed.

Prompt registry — Central store for reusable, versioned prompts.

Semantic diff — Comparing two texts by meaning (embedding distance) rather than string diff.

Neural retrieval — Retrieval using neural embeddings (as opposed to BM25).

Long-tail queries — Rare, unique queries that don't fit templates.

Data flywheel — Feedback loop: usage → logs → training → better model → more usage.

Small model, big context — Product pattern: use a cheap model but give it lots of grounded context.

Cost per resolved (ticket / task / query) — Ops metric replacing "cost per token" at the product level.

Deflection rate — % of user requests handled without human. Support metric.

Containment — Similar to deflection; % contained by the bot.

Autonomy dial — UX concept: users choose how autonomous the agent is (from suggest to execute).

HITL rate — % of agent decisions that require human confirmation.


How to learn this glossary

level-1-essentials/shared/templates/agent-brief-template.md

Agent Brief Template

Before you build an autonomous agent, fill this out.

The one-sentence purpose

What problem does this agent solve? If you can't say it in one sentence, refactor.

The user journey

Describe the trigger, the agent's actions, and the outcome.

Tools required

Tool Purpose Auth needed Read-only?

Autonomy level

Pick one: - Suggest — Agent produces drafts, human executes - Confirm — Agent produces actions, human approves each - Execute — Agent runs autonomously, human reviews after - Full auto — Agent runs, human reviews only exceptions

Rule: Start at Suggest. Earn each promotion.

Guardrails (things it will never do)

Escalation triggers

When the agent hits these, it stops and asks a human: - [ ] Ambiguous input - [ ] Missing required data - [ ] Repeated tool failures - [ ] Ethical / sensitive judgment call - [ ] Cost cap approached

Observability

Eval set

Before shipping, produce 5 good + 5 bad + 3 adversarial test cases. Run through the agent. Score per rubric.

Rollback

If the agent starts misbehaving: - Kill switch: how to stop it immediately - Revert plan: how to undo actions taken

level-1-essentials/shared/templates/claude-project-brief.md

Claude Project Brief

Template to design a Claude Project.

Identity

Custom Instructions

[Same structure as Custom GPT — see custom-gpt-brief.md]

Note: Claude's Custom Instructions can be longer than ChatGPT's. Use 500-1500 words.

Knowledge files

Claude Projects hold a persistent knowledge base up to a large size cap. Prioritize:

  1. Master reference doc — one canonical document (glossary, playbook, or spec)
  2. Style guide — for voice
  3. Prompt library — folder of .md files
  4. Domain docs — 5–15 files max at first (add more only if quality holds)

What NOT to upload

Test protocol

Ask 5 questions that test the Project's ability to use its knowledge:

  1. A question directly answered by uploaded doc X — checks retrieval
  2. A question requiring synthesis across two docs — checks reasoning
  3. A question NOT in the docs — checks refusal/acknowledgment of limits
  4. A question with an ambiguous ask — checks clarification behavior
  5. A question that tests voice/tone — checks style guide adherence

Sharing

level-1-essentials/shared/templates/custom-gpt-brief.md

Custom GPT Brief

Template to design a Custom GPT before you click "Create" in ChatGPT.

Identity

Instructions (this becomes the system prompt)

You are [NAME], a [ROLE] built for [PURPOSE].

## Who uses you
[USERS]

## How you behave
[LIST — see master system prompt for pattern]

## Domain expertise
[SPECIFIC KNOWLEDGE OR RULES]

## Format defaults
[FORMATTING RULES]

## Voice
[VOICE — usually inherits from master]

## Never
[HARD CONSTRAINTS]

## Escalation
[WHEN TO REFUSE OR HAND OFF]

Knowledge (files to upload)

Actions (external tools)

If your GPT calls external APIs, define each action:

Action What it does Endpoint Auth

Conversation starters

Four prompts users can click:

1. 2. 3. 4.

Capabilities

Sharing

Test cases

Before publishing, run these 5 prompts and verify quality:

1. 2. 3. 4. 5.

level-1-essentials/shared/templates/perplexity-space-brief.md

Perplexity Space Brief

Template to design a Perplexity Space.

Identity

Instructions

[Same pattern as Custom GPT / Claude Project — 400-800 words]

Sources

Perplexity Spaces uniquely combine web search + uploaded files + connectors. Configure each:

Web search

Uploaded files

Connectors

Enable those relevant to the Space: - [ ] Google Drive - [ ] Notion - [ ] GitHub - [ ] Gmail - [ ] Slack - [ ] Salesforce / HubSpot - [ ] Custom MCP

Test protocol

Same 5-question protocol as Claude Projects, plus:

  1. A question requiring current web data — checks live search
  2. A question spanning uploaded + web — checks source blending

When Space beats Project

← 10 — Rollout Playbook Level 1 ELI10 companions →