AI cheatsheet · the one-pager
atomeons.com/learn/cheatsheet · CC-BY 4.0
::which model for what · the short answer
- Claude · long writing · code review · long-doc analysis · honest critique
- ChatGPT · short tasks · variants · in-chat tools (interpreter, web)
- Gemini · multimodal · 1M-token context · research-heavy work
- Perplexity · live web grounding · current events · primary sources
- Ollama (local) · NDA / PHI / private data · offline · zero cloud exposure
- Midjourney · highest-quality image generation · paid only
- Whisper · audio → text transcription · open source, excellent
- Wolfram · math verification · symbolic computation
::the 7-part prompt · use this shape for any serious task
- 1. ROLE — “You are a senior X who has done this 100 times.”
- 2. CONTEXT — what the situation is, what you already know, what the constraint is.
- 3. INPUT — paste the actual content (draft, code, data, transcript).
- 4. TASK — the specific thing you want done. Be exact.
- 5. SHAPE — output format: list, table, prose, JSON, length.
- 6. AUDIENCE — who reads the output. “A skeptical board.” “A 12-year-old.”
- 7. VOICE — “Be direct. Skip the diplomatic version. No corporate-speak.”
::the verify rule · the only rule that matters
AI hallucinates with confidence. Always verify: any citation · any medical, legal, financial advice · any claim about current events · any piece of math. If the cost of being wrong is real (audit, lawsuit, harm, money), pull the source yourself.
::never paste into cloud AI
- ○ Social Security numbers, passport, driver license
- ○ Bank accounts, credit cards, routing numbers
- ○ API keys, passwords, recovery codes, SSH keys
- ○ Patient names + diagnoses (PHI / HIPAA)
- ○ Client names + matter facts (privileged)
- ○ Customer PII, addresses, phone numbers, emails
- ○ Trade secrets, unannounced product, source code under NDA
- ○ Employee performance / HR / IEP records
For any of the above, use local Ollama or your facility's vetted internal LLM only.
Cheatsheet · page 2 · prompts + lingo
::10 prompts that work forever
- 1. “Give me 3 versions, across [register A / B / C]. Tell me which to ship and why.”
- 2. “Don't answer yet. Outline first. I'll approve, then you write.”
- 3. “Critique this like a hostile senior. Skip the diplomatic version.”
- 4. “What am I not asking that I should be?”
- 5. “Explain it: like I'm 12 / like I'm a grad student / like a textbook worked example.”
- 6. “What's the strongest argument for the option I'm NOT picking?”
- 7. “Where in this would a skeptical reviewer push back?”
- 8. “Quiz me on this. Withhold the answers until I commit.”
- 9. “What's the trap most people fall into with this kind of task?”
- 10. “Be direct. I want to know what would embarrass me at [my high-bar peer audience].”
::AI lingo · the words you keep hearing
Token · roughly 0.75 of an English word. Models charge per token.
Context window · how much input the model can “see” at once. 200K+ now common.
Hallucination · the model makes up something plausible-sounding but false.
RAG · feeding the model documents at query time instead of training on them.
Embedding · numeric representation of meaning; used for semantic search.
Fine-tune · further-train a base model on your specific data. Often unnecessary.
Prompt engineering · the craft of getting good output from a prompt.
Chain-of-thought · ask the model to reason step-by-step before answering.
Agent · model + tools + a loop. Can take actions, not just respond.
MCP · Model Context Protocol. Standard way to give models tool access.
Temperature · how random the output is. Lower = more deterministic.
System prompt · the instruction-set that runs before any user message.
BYOK · Bring Your Own Key. You pay providers directly; no markup.
Local model · runs on your computer (Ollama, LM Studio). Zero cloud.
::cost reality check (June 2026)
- ▲ Free tier of Claude / ChatGPT / Gemini covers ~80% of daily-use cases for one human.
- ▲ Paid Claude ($20/mo) and ChatGPT Plus ($20/mo) raise the daily-cap meaningfully. Pick one — you don't need both.
- ▲ Pay-per-token via API: $3-$15 per 1M input tokens. Heavy daily use, BYOK is cheaper than subscription.
- ▲ Local Ollama: free forever, ~70% of frontier quality. Worth it for privacy alone.
- ▲ Subscription stacking ($260+/mo across 4 tools) is the most-common waste in 2026.
::when to put the AI down
- ○ When you've been prompting for 20+ minutes on a thing that should take 5. The prompt is wrong, not the model.
- ○ When the AI keeps agreeing with you. Switch to a different model OR ask “steelman the opposite.”
- ○ When you're using AI to avoid a hard human conversation. That's yours.
- ○ When the decision is irreversible (legal · medical · financial · relationship). Pull a human expert.
::AtomEons · /learn/cheatsheet · CC-BY 4.0 · share freely
atomeons.com
::print it · pin it · share it