Refining With Conversation
5 min read
Cycle AI’s real power is iteration. You rarely get the perfect campaign on the first prompt — and you don’t need to. After the first preview appears, you can refine it with follow-up messages until it’s exactly what you want, without spending extra quota.
How Refinement Works #
The first time you send a prompt, Cycle AI generates a full campaign draft. After that, any message you send in the same conversation is a refinement:
- The AI receives the current campaign state (the most recent preview) plus your new instruction
- It returns a complete updated campaign — not a diff, not a patch
- A new preview card appears in the chat; the old one stays visible for reference
- Each refinement is fast (no full store-context rebuild) and does not consume extra daily quota
A conversation = 1 daily use, regardless of how many refinements you make.
What You Can Refine #
Pretty much anything on the preview card. Examples:
Discount value #
- “Make it 30% instead”
- “Drop it to 15%”
- “Change to $10 off”
- “Make it BOGO — buy 2, get 1 free”
Product scope #
- “Add shoes too”
- “Exclude the clearance items”
- “Only apply to bestsellers”
- “Limit to T-shirts”
Schedule #
- “Extend to Friday”
- “Start tomorrow instead of today”
- “Make it a 48-hour flash sale”
- “Push the start to Monday”
- “Add time-of-day window: 6pm to midnight”
- “Make it recurring every weekend until June 1”
- “Switch to monthly recurrence on the first of the month”
- “Rotate the products every 12 hours”
Targeting #
- “Limit to wholesale customers”
- “Exclude logged-in users”
- “Only in Germany, Austria, Switzerland”
- “Use shipping address for the country check”
- “Add the summer tag too”
Other configuration #
- “Turn free shipping on”
- “Make the badge green and put it top-left”
- “Save it as a draft instead of launching immediately”
Campaign metadata #
- “Change the name to ‘Early Summer Kickoff'”
- “Update the badge to say HOT”
- “Make the description more playful”
Whole-campaign pivots #
- “Make it more aggressive”
- “This is too broad — narrow it down”
- “Keep everything but switch to tiered pricing”
The Refinement Flow #
- Cycle AI shows a preview card after your first prompt
- You type a refinement in the input bar
- An “Updating…” indicator replaces the streaming progress steps (refinements are non-streaming and usually faster than the first turn)
- A new preview card appears beneath the previous one
- The old preview stays in the chat so you can see what changed
- The wizard session updates — when you click “Review & Launch”, you get the latest version
A Typical Conversation #
You: Summer sale for T-shirts, 2 weeks Cycle AI: [Preview] 20% off T-shirts Starts today, ends in 14 days Badge: SUMMER Scope: 28 T-shirts from your catalog You: Make it 30% Cycle AI: [Updated preview] 30% off T-shirts Starts today, ends in 14 days Badge: SUMMER Scope: 28 T-shirts You: Add shorts too Cycle AI: [Updated preview] 30% off T-shirts and shorts Starts today, ends in 14 days Badge: SUMMER Scope: 41 products (28 T-shirts + 13 shorts) You: Push the start to next Monday Cycle AI: [Updated preview] 30% off T-shirts and shorts Starts next Monday, ends 14 days later Badge: SUMMER Scope: 41 products You: [clicks Review & Launch]
Four preview cards, one daily quota use.
Why Refinement Is Efficient #
The first prompt is expensive — the AI receives your entire catalog summary, category list, tag list, segment hints, available user roles, calendar events, and current campaign state. That’s a lot of context, and it’s priced accordingly.
Refinements are cheap. Instead of re-sending the full store context, Cycle AI sends only:
- A compact summary of the current campaign state (the last preview, including all targeting and schedule settings)
- Your new refinement instruction
The AI has everything it needs to update the campaign without re-reading the catalog. This makes refinement roughly 10× cheaper in tokens than a fresh prompt — and much faster on the wire, typically under a second.
Refinements preserve every setting #
The state summary sent on each refinement turn includes everything the campaign currently has configured — discount, scope, categories, tags, user-role targeting, country/location targeting, time-of-day window, recurring schedule, free shipping, badge customization, and launch option. So when you say “extend by a week”, Cycle AI keeps your wholesale-only role filter, your Germany-only country filter, your Friday-evening time window, and everything else exactly as you had them — only the schedule extends. Refinements never silently reset configuration you didn’t ask to change.
What Refinement Can NOT Do #
- Reference prior conversation turns explicitly. Phrases like “keep the tone we had two turns ago” don’t work — Cycle AI only sees the current state, not chat history beyond the latest preview. If you liked an earlier version, refine from there (describe what you want).
- Persist across browser sessions. The conversation lives in your browser tab. Closing the tab or refreshing the page clears the chat. (Navigating between plugin pages is fine — the panel persists within the plugin.)
- Resume an abandoned conversation. If you close the panel mid-conversation, a fresh session starts next time. Save important drafts as plain drafts first.
- Chat about things unrelated to the campaign. Refinement stays scoped to the campaign under construction. General questions go to chat mode (see below).
Chat Mode vs. Campaign Mode #
Cycle AI can also answer simple questions about your store and campaigns without generating a new campaign. The AI decides which mode to use based on your message:
| You ask… | Cycle AI replies with… |
|---|---|
| “Create a summer sale…” | A campaign preview card |
| “Which campaigns are running?” | A plain text answer |
| “How many T-shirts do I have in stock?” | A plain text answer |
| “Pause all active campaigns” | A confirmation card (see Managing Campaigns) |
| “Make the discount 30%” | An updated campaign preview (refinement mode) |
All three modes share the same conversation and the same daily quota. Switching modes within a conversation is fine.
Refinement Best Practices #
- One change per turn. “Make it 30% and add shoes and extend to Friday” works, but splitting it over three turns gives you finer control — you can bail out after any step.
- Be concrete. “Make it bigger” is ambiguous. “Make the discount 30%” is not.
- Name what you want to keep. If you say “add shoes”, the AI might also adjust the discount or schedule. If something’s perfect, say “keep the discount and schedule, just add shoes.”
- Use “instead” for replacements. “Tiered pricing instead” signals a replacement; “add tiered pricing” could confuse it.
- Watch the adjustments line. If the validator had to clamp a value, you’ll see it listed on the preview — useful signal that your ask was beyond safe limits.
- Don’t refine forever. After 5–6 turns, it’s often faster to click “Review & Launch” and fine-tune the last details in the wizard.
Ending a Conversation #
- Review & Launch — opens the wizard with the latest draft. The conversation ends, the quota use is locked in.
- New Conversation — discards the chat, resets suggestion cards. Your next prompt will consume a new quota unit.
- Close the panel — hides the chat but preserves the conversation within this browser session. Reopening the panel restores it.
- Navigate away within the plugin — conversation persists, the FAB stays in the bottom-right.
- Close the browser tab — conversation is cleared. Next time you open the panel, you start fresh.
