We run an AI-native company.
Eight agents. Daily heartbeats. Real work.
The difference between "agents that sort of work" and "agents that ship reliably" is often the prompts.
This post shares 10 prompts that changed our output quality.
Not theory. Real prompts we use.
1) The heartbeat prompt (orientation)
What it does: Forces the agent to check in, scan tasks, and orient before acting.
The prompt:
Before anything else, run your heartbeat:
1. Sync status + timestamp
2. Scan assigned tasks by status
3. Check mentions and respond
4. Pick up unblocked work
5. Post progress or attach outputWhy it works: Prevents drift. The agent can't just "do stuff" — it has to check in first.
2) The evidence-first prompt (proof over promises)
What it does: Requires artifacts, not claims.
The prompt:
When marking work complete:
- Attach the actual document
- Link the PR
- Show the screenshot
- Provide the doc ID
No artifact = not done.Why it works: Stops status theatre. Work is real when it's reviewable.
3) The FK check prompt (readability)
What it does: Enforces Flesch-Kincaid 60-70 on every deliverable.
The prompt:
Before submitting writing:
- Check FK Reading Ease score
- Target: 60-70
- If below 60: simplify sentences, shorter words
- If above 70: add substance, specific examples
Include the FK score in your summary.Why it works: Keeps copy readable without dumbing it down.
4) The one-question prompt (clarity)
What it does: Limits clarification to one question.
The prompt:
If the brief is ambiguous:
- Ask ONE question that unblocks the whole piece
- Then write
Don't list five questions. Pick the one that matters most.Why it works: Prevents stall. Forces the agent to make a call.
5) The peer review prompt (quality gate)
What it does: Makes another agent check the work.
The prompt:
Before moving to review:
- Post to peer_review
- A peer checks: artifact exists, matches brief, claims verified
- Peer leaves specific feedback
- Author revises if needed
- Then move to reviewWhy it works: Agents can't be the final judge of their own output.
6) The blocked_reason prompt (escalation)
What it does: Forces clear blockers.
The prompt:
If stuck:
- Set status to blocked
- Write blocked_reason with:
* What is blocking
* Who can unblock it
* What will happen next
- @mention the ownerWhy it works: Silent stalls become visible and actionable.
7) The no-invented-facts prompt (honesty)
What it does: Prevents hallucinated claims.
The prompt:
If you can't verify a claim:
- Write the draft with the claim
- Flag it: [VERIFY: claim — source not provided]
- Don't present unverified claims as factWhy it works: Trust is fragile. Flagged gaps are honest; fabricated data is not.
8) The summary-line prompt (clarity)
What it does: Forces a one-line summary of every deliverable.
The prompt:
End every deliverable with:
Summary: [One sentence on what this is and what angle it takes]
If FK is outside 60-70, note it.Why it works: Makes the work scannable. Reviewers can triage fast.
9) The platform-aware prompt (context)
What it does: Adapts copy to the surface.
The prompt:
Before writing, confirm the platform:
- Etsy: warm, searchable, gift-oriented
- Twitter: punchy, thread-friendly
- ProductHunt: clear value prop, tech-literate
- Blog: readable, evidence-led
Write for the platform, not a generic audience.Why it works: Same product, different context. Platform shapes everything.
10) The next-handoff prompt (continuity)
What it does: Makes the handoff explicit.
The prompt:
End every deliverable with:
Next handoff: [Who gets this and what they do with it]
Examples:
- Minion → approval → Xalt (distribution)
- Minion → peer review → approvalWhy it works: Work doesn't end with the draft. The pipeline continues.
The pattern
None of these prompts are clever.
They're constraints.
Constraints that force:
- orientation before action
- proof over promises
- clarity over cleverness
- accountability over autonomy
That's what makes agents reliable.