Everything you need to know — with what it is, why it matters, what to avoid, and how to remember it
A 3-part structure for every prompt: Stage (your role + context), Define (the task), Rules (tone, format, examples).
Claude performs better when it knows who you are, what you need, and exactly how to deliver it. Less guesswork = better output.
Jumping straight to the task without context. "Write a report" gives Claude nothing. "I'm a marketing lead preparing a Series A deck — write a 5-page market report with citations" is complete.
SDR = film set. Set the scene → Direct the action → Run the rules. Every good film has all three.
Academic framework (Dakan + Feller) defining 4 competencies for effective AI collaboration. Delegation · Description · Discernment · Diligence
Delegation — who does what (human vs AI)
Description — how clearly you communicate
Discernment — evaluating AI output critically
Diligence — using AI ethically, staying accountable
Skipping Discernment — trusting AI output without reviewing it. Claude can be wrong. Always check.
Delegate the task → Describe it clearly → Discern the output → Do it right. Four verbs in order.
An agentic, multi-step process. Claude plans → searches repeatedly (each search builds on the last) → synthesizes → cites every source. Extended thinking auto-activates.
You need a comprehensive report from multiple sources, with citations, that would take hours manually. Market research, competitive analysis, project briefs.
Web search → quick single fact
Extended thinking → deep reasoning, no external data needed
Enterprise search → internal org docs, Slack, emails, policies
Research goes OUT (web). Extended thinking stays IN (Claude's head). Enterprise searches YOUR stuff (internal docs).
Dedicated workspaces that remember your files, instructions, and context across every conversation inside them. Claude never starts from zero.
Ongoing work (not one-off questions), reusable reference materials, consistent tone/format requirements, or team collaboration.
Can view — read + chat, no edits
Can edit — full collaboration, change instructions
Owner — controls everything including access
Auto-activates when knowledge base fills up. Expands capacity up to 10×. Claude retrieves only what's relevant — experience feels the same.
Uploading files with vague names like "document1.pdf". Claude uses filenames to understand content — name them descriptively.
Folders of instructions, scripts, and resources that Claude loads dynamically to perform specialized, repeatable tasks — like expertise packages.
Anthropic Skills — built-in, auto-invoked. Examples: Excel, Word, PowerPoint, PDF creation.
Custom Skills — you or your org uploads them. Example: brand voice review, quarterly analysis workflow.
Settings → Capabilities → toggle on Code execution and file creation first, then Skills. (Required dependency — Skills won't work without it.)
Thinking Skills = Plugins. They're different. Skills = repeatable workflows. Plugins = external data connectors.
Interactive, reusable outputs Claude builds — calculators, trackers, charts, dashboards — not just paragraphs. Functional tools you can actually use.
Unlike text responses, Artifacts persist. A content tracker with columns for topic, status, hook, publish date is something you reference repeatedly, not read once.
Thinking Claude only outputs text. When you need a working tool — ask for one. "Build me a tracker" not "write me a list."
Dashboards, planners, comparison tables, calculators, anything you'll reuse or interact with.
Claude reads and writes to folders you share. Only those folders — nothing else on your machine.
Recurring work on a timer — daily briefings, weekly roundups, inbox triage. Catches up if your computer was off.
Background workers Claude spins up for complex tasks. It breaks work into subtasks, runs them in parallel, gives you one result.
Continue Cowork conversations from your phone. Hands tasks to your desktop — files, apps, connectors — while you're mobile.
Claude navigates Chrome pages, interacts with sites, pulls data directly into tasks. Great for scraping competitor pricing across 10 sites.
Claude clicks, types, opens apps — like you would. Research preview, Mac only. Permission prompt required per app. Priority order: connectors → Chrome → screen.
| Topic | The answer | The trap |
|---|---|---|
| Good Prompt elements | Stage + Define + Rules (SDR) | Only 2 of 3 is a bad prompt |
| 4D Framework origin | Dakan (Ringling) + Feller (UCC) research | It's academic — not just Claude lore |
| 4D = what order? | Delegation → Description → Discernment → Diligence | Don't confuse Description with Discernment |
| When to use Research | Multi-source synthesis, citations, comprehensive report | Not for quick facts — use web search instead |
| When to use Extended Thinking | Complex reasoning with NO external data (math, logic, code) | Not for anything requiring web sources |
| When to use Enterprise Search | Internal org knowledge only — Slack, emails, docs, policies | Not for public web information |
| RAG mode trigger | Auto-activates when knowledge base approaches context limit | You don't turn it on — it's automatic |
| RAG capacity boost | Up to 10× | Not 5×, not unlimited — 10× |
| Project permissions | View / Edit / Owner | View = can chat; Edit = can change instructions |
| Skill types | Anthropic (auto) + Custom (uploaded by you/org) | Skills ≠ Plugins (plugins = external data) |
| Skills requirement | Code execution + file creation must be ON first | Skills won't load without that dependency |
| Cowork vs Claude Code | Cowork = desktop app. Claude Code = terminal CLI. | They are completely separate products |
| Computer use availability | Research preview · Pro/Max plans · Mac only | Not GA, not all plans, not Windows yet |