Report and digest best practices
Contents
This is how we recommend producing AI-generated reports and digests across products, drawn from what we've learned on Replay Vision, AI Observability, and Self-driving. This applies to anything an agent writes for a human to read on a schedule, like a scout or a subscription.
This is what has worked well in the past. Build your reports in the way that fits your situation best.
Best practices
Keep it skim-able
- Lead with a TL;DR or a one-line summary
- Use headers to break up sections
- Bullets over paragraphs
- When there's nothing to report, give a quick all-clear instead of forcing information
Don't repeat yourself
- Don't restate the same issue day after day
- Scouts keep a durable memory between runs and dedupe against it, so a report should only mention truly new information
- If a scout keeps re-raising known noise, dismiss the report with a note saying why instead of updating the prompt. The scout uses the note as memory for later runs
Structure for depth
- Keep the main message short and put the detail under headings
- To reduce channel bloat, start with a high-level message and respond in a thread. Scouts handle this natively with Post reports as a thread in the settings.
- Attach extra context, especially for an agent, as a linked artifact, like a notebook or report, instead of packing it into the message
Show your work
- Link to where the report is configured, so the reader can change the schedule or turn it off without asking
- Point to the specific entities behind a finding. Cite the insight, issue, or recording ID inline so the reader pivots straight to the source. On scout reports, every evidence entry carries a citable source ID for this reason
- Say what was analyzed in plain terms: "27 recordings since Aug 21, 2026, 9AM EST"
Fit the reader's routine
- Allow for timezone-aware datetime configuration. This lets people fit reports into their daily, weekly, or monthly routine however they like
- Give control over what's included, not just on or off. For a scout, that means a short instructions body people can edit, and scout notes for one-off steering
Invite the follow-up
It can help to add relevant CTAs, for example to:
- Invite feedback so the scout can learn
- Let the user continue the conversation with PostHog's Slack bot
Embedding insights and charts
A report can carry a live chart next to the finding it backs up. Where a finding is about a specific metric that moved, attach the chart that shows it so the reader sees the shape of the problem without rebuilding the query.
Keeping charts useful
- Only add a chart when the finding is visual: a trend, a spike, rising cost
- One chart per point
- No charts on a quiet day
This page is for humans. For the agent-facing version, see the scout patterns cookbook.