PATTERN 01 Conversational Foundation

Begin with the End in Mind

Define what "done" looks like before you describe what you want AI to do. A common source of wasted iterations isn't bad prompts - it's no destination.

A professional points to a bicycle sketched on a whiteboard reading Move 1 person 5km, Low cost, Simple - while AI-generated outputs show a helicopter, submarine, and luxury sedan all marked with a red X, illustrating that AI optimizes for what you ask, not what you need

Problem

Output arrives at the wrong destination

Most AI interactions start with a task description but no definition of done. The model produces output - often impressive output - that doesn't fit the actual destination: wrong format, wrong audience, wrong level of detail.

The user iterates. Tokens accumulate. Quality drifts. Time is lost. The model wasn't wrong; it was just never told what "right" looks like.

Value

First drafts arrive ready to use

Output arrives at a usable state in fewer iterations. Measure it: count edit rounds per AI-assisted deliverable before and after you write the destination first.
Format mismatches, audience misreads, and length issues - the most common revision causes - disappear before they occur
Downstream steps (review, approval, publishing) move faster because the first draft is already shaped for its destination

Context

When this pattern applies

Any AI-assisted task that produces a deliverable: a document, analysis, code, plan, recommendation, or decision. Applies at the start of a conversation, prompt sequence, or agentic workflow.

It matters most when someone other than you will read or use the output - a sponsor, a customer, a reviewer - and when the format is fixed by where the output is going.

When not to use it: exploratory work, where you can't name the destination until you've seen a few drafts. There, start loose and let the destination emerge (see Iterate to the Destination, coming soon).

Solution

Specify the destination before describing the task

Before describing what you want the AI to do, describe what you want to receive. Four things to specify:

  • FormatDocument, table, bullet list, code, slide outline, email
  • AudienceTechnical team, executive sponsor, customer, general public
  • ConstraintsLength, tone, reading level, terminology to use or avoid
  • Success criteriaWhat would make this output immediately usable

Then give the task. The model aligns its generation to the destination, not just the request.

Quick check: could someone else tell whether the output met your destination? If not, the destination isn't specific enough yet.

Example opening
  • FormatA two-page executive summary
  • AudienceA CFO with no technical background
  • ConstraintsPlain language, no acronyms
  • Success criteriaConcludes with three specific cost implications
"Then: summarize our Q3 AI pilot results..."

Notes

Patterns that build on this one

Several other patterns build on this one, including Avoid AI Slop with an LLM Interview (Pattern 05). Any technique that helps specify the desired output before prompting is an instance of this pattern.

Use this pattern

Paste this into any AI system's custom instructions or at the start of a conversation to run Pattern 01.

Put this pattern to work

Greg works with teams building AI into their practice. If you want to go deeper on any pattern, start a conversation.

Book a meeting Connect on LinkedIn