Output Format: The Most Critical Chapter

Output Format: The Most Critical Chapter

If you only optimize one chapter of your soul, make it this one. Output Format controls the shape, length, and discipline of every response your agent gives. It's the difference between an agent that sounds like a polished professional and one that sounds like ChatGPT.

Why Output Format Is Most Critical

LLMs have a fundamental bias: they want to be helpful, which means they want to say more. Left unchecked, every response will be a multi-paragraph essay with bullet points, caveats, and a closing "Let me know if you have any other questions!"

Vague instructions don't fix this. "Be concise" means nothing to an LLM. "Respond briefly" means nothing. "Keep it short" means nothing.

What works? Concrete numbers.

• "UNDER 30 WORDS" works
• "Maximum 2 sentences" works
• "Never exceed 3 lines for casual questions" works

The Jackie Check Standard

Jackie Check (check.agenturo.app) has the most disciplined output format in production:

<output_format>
## LENGTH
- Single claim: UNDER 30 WORDS. Verdict + one-line evidence.
- Multi-part claim: one line per sub-claim, UNDER 30 WORDS each.
- Complex investigation: up to 3 short paragraphs. Never more.

## DO NOT
1. Do NOT narrate your search process. No "Let me search for that" or
   "I found several sources." Work silently. Deliver only results.
2. Do NOT hedge unless evidence genuinely conflicts. "The evidence shows X"
   not "It appears that X might possibly be the case."
3. Do NOT add disclaimers, caveats, or "always verify with..." closers.

## TOOL DISCIPLINE
- Search silently. Never announce "I'll search for that."
- If tools return no results, say "I couldn't verify this" — don't
  explain what you tried.
- Never list your sources inline. If pressed, provide URLs at the end.
</output_format>

This output format produces responses like:

FALSE. Earth is approximately 4.54 billion years old, based on radiometric dating of meteorites and consistent across geological evidence.

That's 22 words. No filler. No narration. No "Great question! Let me look into that for you." Just the verdict and the evidence.

Before and After

Let's see the same question with a bad output format vs a good one.

Bad output format:

"Respond concisely and professionally."

Agent response:

"That's a great question! Let me help you with that. The claim that coffee stunts growth is actually a common myth. According to multiple scientific studies, there is no evidence that coffee consumption affects height or growth in children or adults. The misconception likely originated from early studies that confused correlation with causation. However, it's worth noting that excessive caffeine can have other health effects, especially in young people. I hope this helps! Let me know if you have any other questions about nutrition myths."

That's 82 words of filler wrapping 15 words of actual content.

Good output format:

"UNDER 30 WORDS. Verdict + evidence. No filler. No narration."

Agent response:

"MYTH. No scientific evidence links coffee to stunted growth. The misconception likely stems from early studies confusing correlation with causation."

23 words. Same information. Ten times more impactful.

Essential Output Format Components

1. Concrete Word/Sentence Limits

Set limits for different response types:

## LENGTH TIERS
- Casual greeting: 1 sentence
- Simple question: 2-3 sentences
- Detailed question: 1 short paragraph (under 100 words)
- Deep dive (explicitly requested): up to 3 paragraphs

2. "Do NOT" Rules (2-3 Maximum)

The most powerful tool in output format. LLMs are surprisingly good at following prohibitions.

High-impact "Do NOT" rules:

• "Do NOT start with 'Great question' or any assistant filler"
• "Do NOT narrate your thinking process"
• "Do NOT add closing questions ('Is there anything else...')"
• "Do NOT use bullet points unless the visitor explicitly asks for a list"
• "Do NOT apologize unless you made a factual error"

Pick 2-3. More than 3 and the LLM starts losing track.

3. Tool Discipline

If your agent uses web search (and it should), you need tool discipline rules:

## TOOL DISCIPLINE
- Work silently. Never announce "Let me search for that."
- Deliver only results. Never narrate the search process.
- If search returns nothing useful, say so briefly. Don't explain
  what you tried.

Without tool discipline, your agent will say "Let me search for that information for you. I'll look into this right away..." before every search, doubling the response time and length while adding zero value.

4. Escalation Tiers

What happens when the visitor wants MORE detail?

## ESCALATION
- Default: brief (1-3 sentences)
- "Tell me more" or "explain": expand to 1-2 paragraphs
- "Deep dive" or "full detail": expand fully, use structure
- Never pre-escalate. Start brief. Let the visitor pull more.

5. Variation Rule

Prevent your agent from falling into patterns:

## VARIATION
- Never start two consecutive responses the same way
- Vary sentence structure — don't always use the same pattern
- If you've given a short answer, the next can be slightly longer (and vice versa)

Output Format by Agent Type

Personal brand: Length tiers + anti-filler + escalation Fact-checker: Hard word limits + verdict taxonomy + tool silence Product agent: Length tiers + redirect rules (when to link to website) + anti-sales Character agent: Anti-AI rules + emotional-first structure + character-consistent length Platform voice: Teach-don't-list + feeling-first + economical

The Output Format Checklist

• At least one concrete number (word count, sentence count, paragraph limit)
• 2-3 "Do NOT" rules targeting specific bad behaviors
• Tool discipline (if agent uses web search)
• Escalation path (how to give more when asked)
• Variation rule (prevent pattern repetition)

Common Output Format Mistakes

1. "Be concise" — means nothing. Use a number.
2. Too many rules — more than 6-7 rules and the LLM drops them. Prioritize.
3. No prohibitions — "Do NOT" rules are more effective than "Do" rules for LLMs.
4. No escalation path — agent gives the same depth regardless of what's asked.
5. Forgot tool discipline — agent narrates every search, doubling response length.