The problem
Keeping up with AI takes time. And an automatic summary is worthless if you can neither trust it nor read it at a glance.
What I did
CollectionEvery morning, n8n reads 14 sources and drops anything already sent.
AgentA Claude agent sorts, rates the impact for a PM and writes two sentences per article.
PermissionsThe agent that reads the web has no tools: the routine reads and sends, the agent writes.
TemplateThe email follows a fixed template, enforced by n8n rather than by the AI.
Sign-upOne page, prisme.krady.fr, with Tri, the mascot: double opt-in by email, one-click unsubscribe.
MeasureEvery email ends with a “Useful / Not useful” button.
Versions
- For whom
- Me first, then PMs from Noé’s cohort 35 as beta testers, before opening it to all PMs.
- Finding
- Over 12 days of use: an email that changed shape daily, news 3 to 4 days old, the same articles coming back.
- Change
- Refocused on AI for PMs: n8n collects, the agent writes, a fixed template lays it out, one email a day.
- Result
- In test since 2 October: one issue every morning, hands-off. I measure the share of issues rated useful.
Before · plain text, a new shape daily
After · Prisme, fixed template
- For whom
- Just me.
- Change
- A news agent on AI in healthcare, launched with one sentence: from a page, by email or every evening.
- Result
- 13 digests in 12 days, no two in the same format. Then no articles at all once the network blocked the pages.
At a glance
14sources read every morning
5updates at most per email
8:30amthe email lands
Links
What I took away
“Reliability is built into the pipeline, not into the prompt.”How I work →