When Assigning Tasks to AI, Define "Done" First

"Help me organize this weekly report"—if you send this, you’ll likely receive a mind-reading attempt. The biggest pitfall when using AI isn’t its capability; it’s ambiguity over who gets to decide what "done" looks like.

My recent fixed routine is simple: **Before sending the task, write one line defining "what done looks like."**

For example, instead of saying "Organize the interview notes into a document," say: "Format as Markdown, with a Level 2 heading for each interviewee. Every conclusion must include a direct quote from the original text. Append three unresolved questions at the end. Keep the total length under 800 words." This raises the usability rate of the first draft from around 50% to a state where you mostly just need to fix typos.

**When to use it**: When the task has a clear deliverable (document, email, report, code file) and that deliverable will be reviewed or used directly by others. The tighter the deadline, the more valuable "defining done" becomes—because there’s no chance to go back and ask clarifying questions.

**When not to use it**: For brainstorming, seeking inspiration, or tasks like "just throw out a few ideas." Strict formatting constraints can stifle creativity in these cases. Also, when you aren’t sure yourself what you want, it’s more efficient to let the AI produce a rough draft first, then add acceptance criteria after reviewing it, rather than forcing yourself to craft a rigid definition upfront.

**Checklist** (Review before sending the task):

1. What is the output format? (Markdown / table / plain text / code)

2. Are there structural or length requirements? (Heading levels, word count, number of items)

3. What content must be included? (Data, quotes, source links)

4. What content is prohibited? (Fabricated data, pleasantries, disclaimers)

5. Who will review the deliverable? The audience determines tone and professionalism.

**Common pitfalls**:

- **Only writing "be more detailed."** "Detailed" has no acceptance criteria, so it’s effectively meaningless. Change it to "Include a real-world example or data point for each key point."

- **Treating aesthetics as strict standards.** AI can only guess at vague terms like "sophisticated feel." Instead, describe specific elements: "Use short sentences, avoid parallel structures, and minimize exclamation marks."

- **Making the standards too long.** If acceptance criteria exceed five lines, the AI will struggle to balance them all. Pick the three most critical rules, and leave the rest for second-round feedback.

- **Treating first-draft standards as final-draft standards.** In the first round, focus only on structure and factual accuracy. Leave tone and polishing for the second round. Iterating twice is cheaper than trying to craft a perfect prompt in one go.

In short: **Defining "done" means moving acceptance criteria upstream.** Every round of rework you save is paid for by that single line of clarification.