Writing good prompts
Patterns that work, patterns that don't.
Updated · April 2026
The agent is only as good as what you ask it. A few hours of trial and error has produced patterns that consistently work — and patterns that don't.
The shape of a good prompt#
Three parts, in this order:
- Context — what you're trying to accomplish and why.
- Task — the specific thing you want the agent to do right now.
- Constraints — what "done" looks like and what to avoid.
Example#
I'm looking for 5–15 acre parcels in Hillsborough County, FL, for a build-to-rent development. Owner should not be institutional. I'd like 20 candidates ranked by likely cost and ease of acquisition. Don't skip-trace anything — just give me the shortlist with your reasoning. Summarize in a table.
This works because it supplies the motivation (BTR development), a crisp task (20 candidates, ranked), and explicit constraints (no skip tracing, table output).
Patterns that work#
Explicit output format#
"Return a markdown table with columns: APN, address, acres, owner type, your reasoning."
Narrow scope#
"Look at only these three ZIPs: 33609, 33611, 33612."
Guardrails#
"Don't spend more than 5 credits. Tell me if the task needs more."
Intermediate checkpoints#
"Before you pull the full list, show me the filter you plan to run and let me confirm."
Patterns that don't work#
Unbounded creativity#
❌ "Find me good sites somewhere in the Southeast."
The agent will make reasonable guesses, but they'll be your guesses, not its — and you could have narrowed the problem in ten seconds. Narrow before you prompt.
Asking for judgment it can't have#
❌ "Which of these sites should I buy?"
The agent doesn't know your cost of capital, your pipeline, or your internal strategy. It can rank on measurable criteria; it can't pick winners.
Assuming it remembers across sessions#
❌ "Continue what we were working on yesterday."
Sessions don't share memory by default. If you want continuity, paste the previous session's summary into the new session's first message.
Using agent output#
A useful rule: agent output is always a draft. Check its work, especially for:
- Counts and totals — verify against a manual filter if the number matters.
- Ownership claims — the agent may surface an LLC's registered agent as the "owner," which isn't the beneficial owner.
- Regulatory or legal statements — always verify with the jurisdiction or your attorney.
When in doubt, ask it to explain#
Ask the agent to show its reasoning before acting:
"Walk me through how you'd approach this before running anything."
This often catches misunderstandings before they cost credits.