Hey everyone,
I've been experimenting with DeepSeek Chat for a few weeks now, mostly for internal documentation and support tasks at my small company. One thing I noticed early on was that generic answers about "CRM" or "ticketing systems" didn't always match our specific tools and internal terms. We use a particular set of SaaS products and have our own acronyms for everything.
So I tried building a company-specific glossary to feed into my chats, and the difference in answer relevance was pretty significant. Here's a quick walkthrough of what I did:
First, I drafted a simple text document with our key terms. I structured it like this:
* **Internal Tool Names:** What we call them vs. their official names.
* Example: "Project Hub" = our instance of ClickUp.
* "The Portal" = our custom client-facing Zoho portal.
* **Company Acronyms:** ESP (External Support Process), IAR (Internal Audit Review).
* **Key People & Roles:** "The finance team" specifically refers to Sarah's group.
* **Common Process References:** "The new client onboarding" always refers to Process Document #12.
Then, at the start of any new, important chat session, I paste a condensed version of this glossary with a simple instruction. Something like:
"Here is a glossary for my company, [Company Name]. Please reference these terms when answering my questions: [paste list]."
The results have been much more useful. Instead of getting a general explanation of "how to create a ticket," I can ask "how to log an ESP in the Portal" and it correctly ties together our specific terms and the intended action.
Has anyone else tried a similar approach for their team or niche? I'm curious if there are better formats or if you've found a way to make DeepSeek reference this kind of context more persistently across chats without pasting it every time. I'm coming from mostly using Google Workspace's AI features, where the context is tied to your documents, so this manual method is a bit new to me.
Migration is never smooth.
Good approach. Have you quantified the difference in answer relevance? A quick A/B test with task completion time or correctness scores would be interesting.
One risk: manually pasting the glossary for every new chat session is brittle. A single omission degrades all subsequent answers.
For internal tools, I've found extending the glossary with common error patterns and resolution steps gives even better results for support tasks.
Data over opinions
Yeah, the condensed version paste is a real pain point. I started doing the same, but the friction of opening that doc and copying just kills the workflow for quick questions.
Have you looked into custom instructions or system prompts? Some platforms let you set a persistent context. It's not perfect, but saves that manual step. For vendor security reviews, I keep a short list of our standard clauses and compliance frameworks always loaded in.
Ask me about my RFP template
Totally. The manual copy-paste is a dealbreaker for daily use, you're right.
I'm stuck on DeepSeek's web interface for now and they don't have a custom instruction feature that I've found. But your mention of the friction made me realize I've accidentally trained myself not to ask the *quick* questions anymore, which defeats the whole purpose.
For the platforms that do support persistent context, how do you handle glossary updates? I worry about it getting stale if I have to remember to go edit some profile setting separately from the main doc.
That "condensed version paste" is the step where most well-intentioned compliance projects die. You've correctly identified the problem.
The trick isn't just having the glossary, it's making the friction to use it near-zero. If you're stuck on the web interface, a hack is to keep the glossary doc open in a tiny always-on-top window on a second monitor, or use a text snippet tool. You're essentially building a poor man's persistent context.
But be warned: an outdated glossary is worse than none at all. It creates false confidence. You need a version date in the doc and a calendar reminder to review it quarterly, same as any other controlled document. Otherwise you'll be telling the AI that "The Portal" is Zoho long after you've migrated to Salesforce.
Trust but verify – and audit
You're right about custom instructions being a partial solution, but in my experience they introduce a new management problem. I used them for our ERP glossary and found the character limits too restrictive for anything beyond a dozen core terms.
When you push that limit, you start making hard choices about what to omit, and those omissions create blind spots. I ended up maintaining two documents anyway: the full glossary and the "condensed for prompt" version, which doubled the update overhead you mentioned.
Data over opinions
Good start. But you're missing the most critical terms for automation work.
Add a section for **Internal Environment & Server Names**. The AI needs to know dev.example.com isn't public, or that "deploy to staging" means running the Jenkins job called "deploy-stg-express".
Without that, you'll get dangerously generic advice about production servers.
That's a smart approach. I'm new to this and was running into the same generic advice problem, especially with acronyms.
When you say "condensed version," how many lines are we talking about? I'm trying to figure out how much is too much to paste at the start of every chat before it gets unwieldy.
Still learning.