Sources

S1 — Anthropic Academy

authority: official-docs
supports: Confirms that Anthropic offers structured, non-developer AI Fluency courses, validating that agent skills can be taught as a literacy track rather than only as developer tooling.
key-fact: "Featured Course: Claude Code in action; Introduction to Cowork; Build with Claude; Claude for work; Claude for personal."

authority: official-docs
supports: The public syllabus frames the three core skills taught in this course: giving clear instructions, reviewing outputs, and responsible use.
key-fact: "People new to AI — practices giving clear instructions, adding context, and reviewing outputs."

authority: official-docs
supports: The public syllabus maps directly onto the checking-agent-work half of the course: providing context, defining outputs, setting boundaries, and reviewing agent drafts.
key-fact: "People already comfortable with AI chat who want to direct agents through structured work — providing context, defining outputs, setting boundaries, and reviewing agent drafts."

authority: official-docs
supports: Defines the agent/workflow distinction, the need for ground truth, checkpoints, and clear evaluation criteria, and the value of human feedback loops.
key-fact: "Agents are systems where LLMs dynamically direct their own processes and tool usage... During execution, it's crucial for the agents to gain 'ground truth' from the environment at each step... pause for human feedback at checkpoints or when encountering blockers."

authority: official-docs
supports: Explains that effective instructions, context, constraints, and output format are the levers that make a model's output reliable.
key-fact: "Prompt engineering is the process of writing effective instructions for a model, such that it consistently generates content that meets your requirements."

authority: official-docs
supports: Recommends clear, direct instructions, specific output formats, and asking the model to self-check against test criteria.
key-fact: "Claude responds well to clear, explicit instructions... Be specific about the desired output format and constraints... Ask Claude to self-check."

authority: blog
supports: Catalogs agent failure modes visible to a non-coder, including hallucination, goal drift, loops, error snowballs, and premature success.
key-fact: "Hallucination in a chatbot means a wrong answer. Hallucination in an agent means a wrong action... A hallucinated tool call is the same mechanism as any other hallucination — the model generating a plausible-looking continuation."

authority: blog
supports: Explains hallucinated actions, scope creep, cascading errors, context loss, and tool misuse as production failure modes.
key-fact: "Scope creep is when the agent expands beyond its mandate... Agents without defined scope boundaries will expand their actions based on what seems useful, not what was requested."

authority: blog
supports: Distinguishes three end-user failure types: outdated knowledge, missing context, and true hallucination, with a symptom table that matches the course's checking framework.
key-fact: "AI gives wrong answers mainly for three reasons: outdated knowledge (knowledge cutoff), missing context in prompts, or hallucination where information is invented."

authority: video-transcript
supports: Registry-listed beginner explainer for framing the agent concept; transcript pull was attempted but blocked by YouTube sign-in, so this entry is used as a topic pointer only, with technical claims cross-checked against S4 and S8.
key-fact: "YouTube sign-in required during transcript extraction; technical claims in the course were cross-checked against S4 and S8 instead."

authority: video-transcript
supports: Registry-listed first-run tutorial for non-coders; transcript pull was attempted but blocked by YouTube sign-in, so this entry is used as a topic pointer only, with technical claims cross-checked against S4 and S8.
key-fact: "YouTube sign-in required during transcript extraction; technical claims in the course were cross-checked against S4 and S8 instead."