An Agent's Results Change a Lot with "How You Ask"
An agent is like a capable but instruction-faithful worker. Vague instructions spin wheels; clear ones work as expected. Here we organize judging "what to delegate, what to do yourself" and tips for asking.
5 Principles of Asking
1. State the goal
Over "do research," specify the end condition: "5 competitors' price plans in a table; save the result as a file."
2. Limit the scope
Over "do everything," "only files in this folder." Close off areas the AI shouldn't touch up front.
3. Specify confirmation points
Over "show me midway," embed human judgment points: "report once after reviewing 3 sites; continue if OK."
4. Failure behavior
Instruct behavior when things go wrong: "ask if uncertain, don't guess," "stop if no progress in 30 minutes."
5. Request logs
"Record the steps you took." Auditable/reproducible, and you can trace causes on failure.
Judging Tasks OK to Delegate
| Axis | Delegate | Do yourself |
|---|---|---|
| Failure impact | Small (reversible) | Large (irreversible) |
| Judgment quality | Objective, patternable | Subjective, sense, tacit |
| Frequency | Repetitive | One-off |
| Confidentiality | Low | High |
| Creativity | Routine work | Creation/strategy |
Bad Asking → Good Asking
Bad (vague, spins wheels):
× Research our company's competitors … too broad; agent doesn't know "our company"
× Reply to the email … which email, what tone, who decides to send—unclear
× Organize the files … which folder, how, delete-or-not—unclearGood (concrete, bounded, with confirmation):
Research 5 companies offering "AI coding tools" on the web.
Per company: price, free tier, key features in a table. Result in Markdown.
Write "unknown" for unknown items; don't guess.
When done, show me the result; save after my approval.Operational Tips
- Always insert approval points: have it confirm "OK to execute?" before send/purchase/delete
- Templatize once routine: write once and reuse for weekly repeats
- Keep logs: so you can later trace "why did this happen?"
- Design to tolerate failure: agents aren't perfect. Deploy where 80–90% success suffices, with the final check held by a human
Asking to Avoid
- A huge task at once: split and stage it
- "Do it well": vague instructions yield vague results
- Hands-off with no supervision: don't walk away on "handle it"; check at milestones
Next Step
Once asking is second nature, next is "Handling Confidential Information." The line between internal rules and AI use is the key to safe, ongoing in-company use.
