Your AI Newsletter Reads Like a Database Dump (This is the Fix)
Our AI-generated digests had an 18% open rate. Five structural rules from Morning Brew, TLDR, and Ben's Bites doubled it. The template, the prompt trick, and the before/after.
At 3:17 a.m. last Thursday I pulled the latest digest from our agent pipeline. It contained fourteen RSS items, three YouTube transcript summaries, and two market-signal alerts. Every bullet began with "The model shows" or "OpenAI released." The facts were correct. The output was still a database dump.
Our open rate sat at 18 percent. Replies were zero. I had built an automated system that produced accurate summaries but gave readers nothing to act on.
Five Rules Stolen from Newsletters People Actually Read
I studied the three newsletters that actually get opened: Morning Brew opens with a single implication before any data, TLDR never exceeds seventy-five words per item and uses bullets for the numbers, and Ben's Bites ends every entry with one sentence on what changes next. I extracted five structural rules and forced them into our pipeline.
1. Lead with the implication, not the fact. Readers scan for relevance first. Start with a one-sentence "why this matters" hook before any details.
2. Cap every summary at 75 words. Break it into two short paragraphs plus three bullets. Dense blocks kill engagement.
3. End each item with a forward-looking sentence. This creates momentum and makes the newsletter feel like a curator, not a feed.
4. Use consistent visual hierarchy. Bold section headers, bolded key phrases inside items, and one italicized "editor's take" line per item for voice.
5. Limit the total issue to 5-7 items maximum. Prioritize ruthlessly and group weaker items into a "Quick Hits" section of 20-30 words each.
Those five rules alone cut average read time from nine minutes to under four.
The Second-Pass Prompt That Changed Everything
The largest single improvement came from a second-pass prompt that runs after the raw agent output. The raw summaries are accurate but robotic. This prompt humanizes them:
"Rewrite the following summary. Start with one sentence stating the implication for builders. Use contractions. Never start with the company name. Sound like a knowledgeable colleague, not a press release."
Before and After
Before (raw AI summary):
"OpenAI released GPT-5 with improved reasoning capabilities. The model shows 40% improvement on MATH benchmarks and introduces native tool use. Pricing starts at $15/1M input tokens. Key implications: competitive pressure on Anthropic and Google, potential disruption to existing agent frameworks, new multimodal capabilities may obsolete current vision pipelines."
After (second-pass rewrite):
GPT-5 just made reliable multi-step agents practical. The model scores 40 percent higher on MATH benchmarks and now calls tools natively out of the box. Pricing begins at $15 per million input tokens.
- Agent scaffolding that once required brittle external loops now runs inside the model.
- Multimodal features are strong enough to replace several current vision pipelines.
- Expect Anthropic and Google to ship matching tool-use updates within ninety days.
*Editor's take: This is the first release that feels like it could actually compress the gap between today's agents and something production-ready.*
The second version is still factual, but a reader can decide in eight seconds whether to keep reading.
The Template
We now ship every digest with this fixed template. A sixty-word intro paragraph sets the week's theme. The main section contains three to five items, each under seventy-five words and formatted with bold claim, bullets, and one italic line. Quick Hits follows with three items at twenty-five words each. The closer is a single forty-word reflection that ends with a question.
After the first week under this structure our open rate rose to thirty-four percent and we received twelve direct replies. The pipeline still runs automatically, but the output no longer reads like a database dump.