CCAO-F curriculum

A fluent, well-organized answer can still be wrong or missing half the request. Check it against the source and the original ask before you use it.

Hallucinations look specific and plausible, inconsistencies hide across sections, and bias shows in what's missing. Learn the tells and how to check each one.

Verify claims against authoritative sources, and send output to a qualified human when stakes, audience, regulation, publication, or irreversibility demand it.

A correct answer can still fail its reader. Adapt length, tone, and detail to the audience, and choose inline, an Artifact, or a table to fit how it'll be used.

Use Claude to surface gaps, conflicts and assumptions in requirements, then confirm them with stakeholders. Claude analyzes; people decide what is actually needed.

Research options, plan in stages and iterate designs in short loops with Claude — verify the facts a decision rests on and keep the decision with people.

Augment one workflow step before redesigning, set honest expectations with stakeholders, and escalate integration or automation to a Claude Developer or Architect.

Claude can assist with almost any task, but high-stakes decisions about people and final legal, medical or financial judgments stay with an accountable human.

Classify data before you share it, redact or anonymize anything restricted, and use only approved tools. Telling Claude not to retain data is not a policy control.

Check output for bias, disclose AI assistance where people expect it, own every outcome you sign off on, and keep your own judgment active rather than over-relying.

A strong prompt states the goal, audience, context and finish line, and adapts to whether the task is analysis, research, drafting or brainstorming.

Split a complex request into ordered steps you can check one at a time, so errors surface early instead of hiding inside one giant answer.

Improve a promising draft with targeted follow-ups that name what to keep and what to change, one clear change at a time, and know when it is good enough.

Match the feature to the work: chat for one-offs, Projects for recurring context, Artifacts for content you'll iterate on, research for multi-source questions.

Haiku for fast, low-cost simple work at volume; Sonnet as the balanced default; Opus for complex reasoning where quality justifies slower, costlier output.

Long chats drift and have limits. Know when to restart with a summary, when to keep going, and when to persist context in a Project or memory instead.

Project instructions say how Claude should work in every chat; Project knowledge supplies what it works from. Keep each focused, current and clearly scoped.

Connect only the sources a task needs, prefer live sources over stale copies, and give every Project an owner who reviews and updates its files and instructions.

When output keeps missing, find the cause (missing context, vague criteria, wrong framing, wrong tool, overloaded chat, bad sources) before rewriting.

Let feedback drive changes, then streamline: reuse context, right-size the model, cut wasted steps, and keep the human checks that matter.