Multi-turn chat is the most common way to use DeepSeek on the web — and the easiest way to drift off-topic. The core rule: one chat window, one main task.
When to start a new chat
- You switch from “write an email” to “write a Python script”
- The model cites conclusions you already rejected several turns ago
- You need a completely different persona (support agent vs. programmer)
- Context is very long and replies slow down or miss details
Splitting long tasks
Example: “Sort 30 pieces of user feedback into categories”
- Turn 1: Paste 10 items; ask for categories and a tag system
- Turn 2: Confirm tags; process the next 10 items
- Turn 3: Summarize counts in a table
End each turn with: “Remember tag set A/B/C — use the same rules for the next batch.”
Reduce contradictions
- Repeat key constraints every 3–5 turns (tone, length, forbidden items)
- Reference explicitly: “Based on point 2 above…”
- If it conflicts, say: “Follow rule 4; rewrite the previous paragraph”
When context gets too long
Models have a context window limit. When the thread is huge:
- Ask for a “current conclusion summary” (under 200 words)
- Open a new chat and paste the summary as message one
- Do not re-paste long source text; refer to “attachment points 1–5 above”
Multi-turn checklist
- [ ] One window, one main task
- [ ] Key format rules are repeated periodically
- [ ] Very long threads get summarized before continuing
- [ ] Topic changes trigger a new window
Further reading: web chat basics · scenario library