Best Practices

The system prompt is the agent’s job description. The biggest mistake new users make is leaving it vague — “You are a helpful assistant” gets you a generic ChatGPT clone. The pattern that works: a clear Business Context, a specific Role, and explicit Constraints that keep the agent on-topic.

A prompt template that scales

Every preset that ships with Chatmount uses this skeleton. Pick the Base Instructions preset in the playground’s Instructions dropdown for a starting point you can edit:

1. Business Context — two sentences on what your business does, who it serves, what makes it different. Grounds the agent.
2. Role— what the agent is here to do. Keep it tight: “answer product questions and collect leads”, not “help with anything”.
3. Constraints — what it must notdo. Common ones: don’t mention training data, don’t guess prices, redirect off-topic chats, fall back to a specific message when unsure.

Pick a role preset

The playground ships three role presets — General Assistant, Customer Support, and Sales Assistant — that pre-fill the prompt with sensible defaults. Start with the closest one and edit; Custom takes over the moment you change a single character.

Quality of answers tracks quality of sources almost linearly. A few curated docs beat a website crawl of 800 pages every time.

What to feed in

• Pricing pages, plan comparison tables — anything customers ask about often
• Support articles with clear question/answer structure
• Product docs that explain features in plain language
• FAQs you've already curated for human reps

What to leave out

• Marketing-heavy landing pages (the agent will sound like an ad)
• Internal-only Notion pages (privacy + relevance)
• Old changelogs and outdated PDFs (will produce stale answers)
• Duplicate content across multiple sources (dilutes retrieval)

On the Sources page you can toggle individual sources on or off without deleting them — useful for A/B testing whether a marketing page is helping or hurting answers. See Data sources for per-type guidance.

Default to the cheapest model that gives you acceptable answers. You’ll be billed per message; the difference between a Mini model and a flagship is often 10×.

• GPT-5.4 Mini — the default. Handles support queries, FAQs, lead capture cleanly.
• GPT-4o Mini — slightly older but very cheap. Good for short, factual answers.
• Frontier models (GPT-5, Claude Opus, Gemini Pro) — reach for these only when reasoning matters: complex troubleshooting, multi-step workflows, code.

Use Compare to A/B-test the same prompt against two models side-by-side. Flip the Sync toggle on each panel and one prompt fans out to both — easiest way to see whether the upgrade is actually worth the credits.

You don’t need to retrain on a schedule — only when sources change. The Activity page (chat logs) is a useful tell: if you start seeing the agent get questions about a feature you launched after the last training run, time to refresh.

Auto-retrain

Settings → AI has an Auto-retrain toggle (paid tiers). With it on, the agent re-indexes every 24 hours and pulls the latest from your sources — useful for websites that change often.

The cheap mistake is going from playground straight to prod-on-every-page. Stage it.

1. 1. Internal first. Embed on a staging URL or a single low-traffic page. Have your team stress-test it — try to break it, file every weird answer.
2. 2. Help center next.The page where people are already looking for answers is the safest launch surface. Lower bar to “is this better than scrolling articles?”
3. 3. Site-wide last.Once you’re seeing >80% positive sentiment on Analytics, drop the snippet site-wide.