When Things Break: Troubleshooting
This is a member-only chapter. Log in with your Signal Over Noise membership email to continue.
Log in to readModule 6 · Section 3 of 6
When Things Break: Troubleshooting
Even with a solid style guide and an established editing workflow, specific problems keep recurring. Here’s a diagnostic breakdown of the most common ones — what you’re seeing, why it’s happening, and what to actually do about it.
Issue 1: AI output feels too formal
What you’re seeing:
“One might consider” instead of “You should think about.” “It is advisable to” instead of “You should.” “Individuals” instead of “people.”
Diagnosis:
Your style guide lacks specificity about casualness level, or AI is interpreting “professional” as “formal.” These aren’t the same thing — professional means competent and clear; formal means distant and stiff. You want the first, not the second.
Fix:
Add explicit casualness instructions. “Contraction usage: always. Write ‘don’t’ not ‘do not.’ Write ‘you’re’ not ‘you are.’”
Then add a before/after example directly in the guide: “Too formal: ‘One might consider examining the implications’ → My voice: ‘Here’s why this matters.’”
Test with a sample paragraph immediately. If the fix worked, you’ll know. If it didn’t, the instruction needs to be more direct.
Issue 2: AI uses jargon you would never use
What you’re seeing:
“Leverage,” “synergy,” “utilise,” “robust,” “seamless,” “stakeholder,” “touch base,” “circle back.”
Diagnosis:
Your forbidden list is incomplete. The list needs to grow as you catch things — it’s never finished.
Fix:
Build a comprehensive forbidden words list and update it after every editing session where you catch jargon. Aim for 30-50 items within the first month. Categories to include:
- Corporate jargon: leverage, synergy, utilise, robust, seamless
- Buzzwords: game-changing, revolutionary, cutting-edge, innovative
- Academic formality: moreover, furthermore, henceforth
- Meeting culture: stakeholder, touch base, circle back, action item
Every word you catch in editing belongs on the list. Don’t rely on memory — document it.
Issue 3: Every piece sounds the same regardless of topic
What you’re seeing:
Blog posts sound like newsletter issues sound like LinkedIn posts. Technical topics get the same treatment as personal stories. Everything has the same rhythm, the same structure, the same energy level.
Diagnosis:
Your style guide is too rigid. It describes your voice accurately but doesn’t account for how your voice shifts across contexts. That shift is normal and intentional — you should sound different when you’re teaching versus when you’re arguing a point.
Fix:
Add context-specific guidelines to your guide:
- “For technical topics: more specific terminology, longer sentences for complex explanations, fewer conversational asides.”
- “For opinion pieces: more personal examples, stronger stances, shorter sentences for emphasis.”
- “For tutorials: shorter sentences overall, more imperatives, strict step-by-step structure.”
Then create separate prompt templates for each major content type that reference the appropriate context-specific guidelines. One master voice, multiple appropriate expressions of it.
Issue 4: AI loses your personality in longer pieces
What you’re seeing:
The first 300 words sound like you. The next 1,500 words drift into generic territory. By the end, it could be anyone.
Diagnosis:
Your style guide doesn’t have enough examples of extended writing, or AI is losing context over long pieces. Both are fixable.
Fix:
Add longer writing samples to your initial analysis — 2,000 words minimum. Include examples that show how you maintain voice across sections, not just in short bursts.
Add specific guidance: “Personality refreshers — include conversational asides every 300-400 words. Long pieces need rhythm variation every 2-3 paragraphs to avoid monotony.”
For pieces over 2,000 words, work in 500-word chunks with editing passes after each chunk rather than drafting everything then editing. Don’t hand AI a 3,000-word outline and expect consistent voice throughout.
Issue 5: Editing takes as long as writing from scratch
What you’re seeing:
You’re rewriting 60-70% of AI output. The time savings aren’t materialising. You’re starting to wonder if this is actually worth it.
Diagnosis:
The AI first draft is too far from your voice. This indicates one or more of: inadequate style guide, lazy prompting, or AI losing context over long pieces.
Fix:
Multi-part solution — you likely need all of these:
- Write your own 200-300 word core first before AI touches anything. Give it something to work with that’s already in your voice.
- Be more specific in prompts. Don’t say “write about X.” Say “expand MY outline about X using MY style guide.”
- Reference the style guide explicitly in every prompt: “Review my style guide before generating anything.”
- For very long pieces, work in chunks rather than generating the whole thing at once.
- Accept that substantial editing is normal and valuable. The efficiency gains show up across volume — 50 pieces across a year — not necessarily in every single piece. Some pieces will need heavy editing. That’s fine.
The goal isn’t to stop editing. It’s to edit toward your voice rather than away from the AI’s.