Pattern recognition with a fancy suit on. AI predicts the next token from what you give it — say "twinkle, twinkle, little", it returns "star." Better input, better prediction. Use four parts every time:

[ROLE]    Act like a world-class strategist focused on conversion.
[CONTEXT] Product, audience, current copy, ICP notes …
[COMMAND] Rewrite the homepage hero to maximise booked demos.
[FORMAT]  Table — Headline · Subhead · CTA · Reasoning.

Output never exceeds input quality. The prompt caps the ceiling.

The context window is the river your output flows from. Pollute it and the output is polluted. Curate ruthlessly — every extra token steers the next one.

Length

Token budget is finite. Past it the model truncates or finds the relevant bits poorly. Pin what matters — don't dump entire docs.

Compaction

When the window fills, the agent summarises older messages. Details smooth out — exact wording, numbers, quotes. That's when AI "forgets." Save precise items outside the chat before compaction.

Pollution

Irrelevant or contradictory info derails confident answers. One wrong fact buried in long context surfaces in the output.

Tone

Aggressive or negative wording produces statistically worse outputs. Models mimic register — feed it craft, not contempt.

Don't argue

When wrong, don't reply "you're wrong, redo." Edit the original prompt or branch fresh — corrections + the mistake stay in context and skew what follows.

Reset

Don't keep one chat going forever. Old goals bleed into new tasks. New goal → new session.

Edit the prompt, not the response.

Like a GPS, not turn-by-turn directions. Engineers call this spec-driven development.

1. Set the role and context.
2. State the outcome — be specific. "A sequence that converts cold leads into booked calls."
3. Add: "Ask all the questions you need. Return as <format>."
4. Answer conversationally.
5. If off-target: "refine — ask more questions." Iterate.

Spec-driven — destination first, route second.

Use AI to scaffold its own work. Two reusable patterns for the same problem — quality output without grinding 30 iterations.

1. "Write a plan first. Steps, dependencies, edge cases. Don't execute yet."
2. Review the plan. Refine in plain English: "add this, restructure that."
3. Say "go" when it's right. Catches wrong assumptions before any output is generated.

1. "Senior prompt engineer. Goal: <X>. Write me a prompt — role, context, constraints, format."
2. Tweak the draft to fit your situation. Run it.
3. If output disappoints: "Why did this fall short? Rewrite the prompt." Save the polished version, reuse it.

You direct — the model plans and drafts.

A recipe that runs on repeat — same ingredients, same sequence, predictable result instead of 17 minutes of back-and-forth.

Agent

A system prompt packaged for reuse — a Custom GPT, Claude Project, Gemini Gem. Define role, constraints, tone and tools once; talk to it forever.

Skill

A smaller, composable capability the agent invokes when relevant — "summarise contract", "extract invoice", "format SQL." Skills stack inside agents.

1. Describe the agent: "Help me build an agent that does <X>. Ask what you need to know."
2. Answer its questions. Try the prompt it gives you on a real task; refine in plain English.
3. Save as an Agent — the exact method differs from platform to platform.

Browse aitmpl.com for ready-made Agents and Skills.

Single-pass output is fragile. Models polish confidently — even when wrong. Push it to challenge its own work before you accept anything. The model only catches its blind spots when you make it look.

Devil's advocate

"Argue against your previous answer as a skeptical senior reviewer. What's weakest? What's missing?"

Code review

"Review what you just produced — bugs, edge cases, weak assumptions, missed requirements, security holes."

Failure modes

"List five ways this could fail in production. Now fix the top three." Forces it past the happy path.

Fresh eyes

Paste the output into a clean chat with no prior context: "Critique this." Removes accumulated sycophancy from a long session.

First answers are hypotheses, not conclusions.