A recurring challenge with AI-generated content is that it can feel generic, hallucinate facts, or miss client-specific context — leading to client complaints like "this doesn't sound like us." The techniques below address these failure modes by grounding AI output in real documents, using the right tool for the right job, and cross-checking outputs across multiple models.
These practices were shared during the [1] and are intended to be formalized in a team-wide AI professional development session planned for October.
The most common mistake is jumping straight to a writing prompt without first establishing context. Instead, build a knowledge base from real client materials, then interact with that grounded context.
Prompt technique for chat-based tools (no documents):
1. Give the AI the client's website URL and ask it to summarize what the company does. Confirm it understood correctly.
2. Introduce the specific topic (e.g., "Now let's talk about reverse ATMs — do you know what that is?") and coax it until you're confident it has the right context.
3. Only then ask it to write.
Best for: Project-specific knowledge bases drawn from existing documents.
How it works:
- Go to NotebookLM and create a new notebook named for the client or project.
- Upload all relevant source materials: PDFs, TXT files, Markdown, Google Docs, website URLs, or pasted email text. (Convert PowerPoints to PDF first.)
- NotebookLM generates a summary of everything you've given it, then answers questions and writes content drawing only from those sources — it does not go out to the internet.
Key benefit: Eliminates hallucination and off-topic content. The AI stays within the bounds of what you actually know about the client.
Additional features: Mind maps, in-depth reports, flashcards.
When to use it: Any time you have a collection of client documents and want AI to synthesize or write from them without going off-reservation.
Best for: Document-grounded interaction with strong writing and reasoning quality.
How it works:
- In Claude, go to Projects → New Project.
- Upload client files to the project. Claude will reference these files in all subsequent conversations within that project.
- Claude is particularly strong at writing, editing, strategy, and code.
Key benefit: Like NotebookLM, it keeps the AI focused on your documents rather than the open internet. Claude's writing quality is generally considered superior to ChatGPT for most content tasks.
When to use it: Default choice for most writing and strategy work (~90% of use cases per team experience). Also excellent for reviewing and editing output from other tools.
"I'll have ChatGPT do something, then take whatever it gave me over to Claude and say, 'ChatGPT gave me this — what do you think?' And Claude will edit it." — Mark Hope, 2025-09-30 Ops Sync
Best for: Internet research with cited sources; competitor discovery; fact-checking claims.
How it works:
- Perplexity functions like a high-speed, AI-powered browser. Every claim it makes is backed by a citation you can click through to verify.
- Use it to research industries, find competitors, surface what others are saying about a topic, or discover sources you wouldn't have found manually.
Key benefit: Eliminates unsourced hallucination. If it can't cite something, it won't say it.
Limitation: Not creative. It can't synthesize or reason the way Claude or ChatGPT can. Use it for research, not for drafting.
Pro tip — force citation discipline in any AI tool:
Add this to your prompt: "Don't give me any facts, statistics, or details unless you can support them with a citation." This changes how the model writes and significantly reduces made-up statistics.
Spaces: Perplexity also supports project-like "Spaces" where you can attach documents and maintain ongoing research threads.
Best for: Tasks that require direct integration with Google Workspace.
When to use it over Claude/ChatGPT:
- Creating or editing Google Docs or Google Sheets directly
- Accessing files in Google Drive
- Any output that needs to live natively in the Google ecosystem
ChatGPT and Claude cannot create Google Sheets or write directly to Drive. Gemini can.
One of the most effective techniques for improving AI output quality — and making it less detectable as AI — is to bounce the same content across multiple models:
Each model will make different edits, which naturally diversifies the writing style and reduces the homogenized "AI voice" that clients notice. This technique is especially powerful for code review but works equally well for written content.
Clients will notice if content sounds generic or off-brand. Common causes and fixes:
| Problem | Fix |
|---|---|
| AI doesn't know the client's voice | Build a NotebookLM or Claude Project with real client materials before writing |
| Output contains made-up statistics | Add citation-requirement language to your prompt; verify with Perplexity |
| Content is technically correct but misses nuance | Ask the client for more input; use their own words as source material |
| Writing sounds robotic | Cross-check across multiple models; edit the output yourself before sending |
Important: Always read and internalize AI-generated content before presenting it. You need to be able to speak to it in client conversations — don't just skim and send.
A structured training session led by Mark Hope is planned for October 2025 (after the client health check). Attendance is expected for account managers, Raphael, and Gavin. Isalia Ramirez is coordinating scheduling.
Topics will include live walkthroughs of the techniques above and hands-on practice with real projects.
→ See [1] for action items and context.