Prompt tips for marginal's generate_variants tool
The generate_variants tool turns a short prompt into a batch of subject line candidates ready for a subject line A/B test. The quality of those candidates depends almost entirely on what you tell it. These marginal generate_variants prompt tips help you get sharper output from your AI agent on the first call.
Give the tool real context, not a topic
A prompt like "write subject lines for a sale" produces generic candidates. generate_variants does much better when you describe the audience, the offer, and the action you want recipients to take. The more concrete the brief, the more distinct and testable your variants become.
Treat the prompt as a creative brief you'd hand to a copywriter. Include constraints up front so the agent doesn't waste variants on tones or formats you'll never ship.
- Name the segment: "lapsed trial users who signed up 30+ days ago"
- State the offer plainly: "20% off annual plans through Friday"
- Specify a desired tone: urgent, playful, plain, or curiosity-driven
- Add hard rules: "no emojis", "under 50 characters", "avoid the word free"
Ask for variety across hypotheses
A strong subject line A/B test compares ideas, not synonyms. If every variant says the same thing five ways, the test teaches you nothing. Prompt generate_variants to span different angles so the lift metrics you get from get_results actually point to a winning strategy.
When you frame each variant around a distinct hypothesis, the within-test lift data becomes a map of what your audience responds to, not just which wording edged out the others.
- Request one variant per angle: benefit, urgency, social proof, question, curiosity
- Ask for a control-style baseline alongside bolder swings
- Specify how many variants you want and keep it within your test budget
- Tell the agent to vary length so you can test short vs. descriptive lines
Wire it into your workflow
Because marginal is a hosted email MCP server, generate_variants is one call in a sequence your AI agent can run end to end. Once you're happy with the candidates, pass them to launch_test, then pull open and click tracking with get_results, and let recommend_next suggest the follow-up.
Connect from any MCP client — Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Zed, or OpenAI Codex — point it at https://marginal.sh/mcp, and authenticate with your Bearer API key. The free tier covers 100 experiments per month for AI agent email marketing experiments.
- Generate, launch, measure, and iterate without leaving your agent
- Reference prior winners in your prompt to build on what already worked
- See the full tool reference at https://marginal.sh/docs/
Common prompt mistakes to avoid
Most weak generate_variants output traces back to an underspecified prompt or an overloaded one. Keep your instructions concrete but focused so the agent can actually act on them.
- Don't bundle multiple offers into one prompt — run separate tests
- Don't omit length limits if your inbox preview matters
- Don't ask for 20 variants when you only have budget to test a few
- Don't skip the audience description — it's the biggest lever on quality
Get started with marginal
marginal is a hosted email marketing MCP server at marginal.sh. Sign up free, create an API key, and connect https://marginal.sh/mcp from Cursor, Codex, or Claude Desktop.
- 100 experiments/month on the free tier
- Four MCP tools: generate_variants, launch_test, get_results, recommend_next
- Listed in the MCP Registry as sh.marginal/mcp