AI Prompting — Field Guide

Every AI model operates with a context window — the total amount of text it can hold in view at once. Think of it as a very large but ultimately finite notepad. Every message, response, and instruction takes up space on it.

As a conversation grows longer, older content — including instructions set at the very beginning — begins to carry less weight relative to recent exchanges. In very long conversations, early content may effectively fall out of active influence altogether. This is not the AI being unreliable; it's a predictable consequence of how the models work.

• The AI contradicts or ignores a constraint set earlier (e.g. tone).
• The persona you established begins to revert to a generic assistant-like voice.
• The AI asks questions or makes suggestions it already addressed earlier.
• Output quality or consistency drops without any change in your prompting.
• Responses feel generic, as though your project context is no longer applied.

If you notice any of these signs, don't just rephrase your request — reconfirm the context.

Reconfirmation means explicitly asking the AI to surface and verify the key facts, decisions, and constraints established so far. Use this checkpoint prompt at any point — particularly when you notice drift, or proactively after a significant number of exchanges.

Once the AI provides its summary, review it. If anything is missing or has drifted, issue a correction prompt immediately.

Use reconfirmation: after roughly 20–30 exchanges · when switching task phases · when you notice drift symptoms · before producing a final or polished output.

For long, multi-session projects, begin each new session with a summary briefing rather than continuing an existing thread indefinitely. This removes reliance on the AI having retained earlier context.

• Project name and objective — what you're ultimately trying to produce
• Persona or role — who the AI should be
• Audience — who the output is for
• Tone and style — formal, conversational, technical
• Key decisions made so far — choices not to be revisited
• Constraints — word count, format, things to avoid
• Current stage — where you are and what's needed next

Keep a brief running document with this information, updated as the project evolves, and paste it into the first message of each new conversation.

Custom Instructions define standing preferences applied automatically to every conversation — your typical audience, preferred tone, or output type. They sit outside the context window and are re-applied consistently, making them immune to context drift.

Project Files let you attach documents and reference material to a project so the AI has access throughout your work without re-pasting them. A well-prepared project file containing your brief, persona, constraints, and key decisions is a permanent context anchor.

• Not fact-checking: AI can confidently state incorrect information
• Sharing sensitive data: Don't put confidential information in free AI tools
• Assuming AI knows your company: AI needs context about your specific situation
• Not checking for bias: AI can reflect biases from training data
• Using outdated information: Verify current facts, laws, or prices
• Ignoring your judgement: AI assists but doesn't replace human decision-making

• Expecting perfection immediately: AI is a tool that requires practice
• Giving up too quickly: Refine your prompt rather than abandoning the task
• Not experimenting: Try different approaches to find what works
• Thinking "AI can't do my job": AI augments your work, doesn't replace your expertise
• Comparing to experts: Compare your AI-assisted work to doing it manually