Turn session replays into a morning list of product fixes

An agent can turn selected session replays into a morning digest by pairing timestamped frames with product events, describing visible problems, and linking each note to the recording. Start with sessions that dropped out of signup, top-up, or the first successful call. Keep the review bounded and require evidence before a suggested fix becomes an engineering ticket.

The sample notes below are fictional illustrations of the review format.

Select sessions with a specific question

Build a funnel for the flow you want to inspect. A first-call funnel might use signup completion, API key creation, and a successful call. A top-up funnel needs its own events, as does a first-result onboarding offer. Choose the conversion window before declaring a person has dropped out, then select eligible recordings from that group. PostHog lets you inspect people who dropped at a funnel step. PostHog funnels.

For a proposed ten-session morning digest, set a maximum of ten and state how they were selected. A sample chosen only for repeated clicks will overrepresent that pattern. Rotate between eligible drop-offs, include some completed journeys for context, and report how many sessions lacked usable recordings. Missing recordings should never become “no problem found.”

PostHog already offers AI-assisted replay analysis. A separate paid-vision workflow is worth considering when you need a particular rubric, cross-tool evidence, or your own output format. Check the native capability first. Session Replay overview.

Establish how frames will reach the model

The ordinary recordings API returns metadata. Raw recorded events use a separate snapshot API, which PostHog says was built for its own UI and can change with little notice. Its output needs processing before playback; “snapshot” does not mean a ready-to-send screenshot. Replay API, snapshot API.

For a small pilot, capture frames with the player's Screenshot control or use a permitted video export. As checked on September 23, 2026, PostHog documents video-export limits per project and calendar month of 10 on the free plan, 15 on paid plans, and 25 on enterprise plans. Clips count toward those limits. Native video export therefore cannot be assumed to supply ten new sessions every morning. Sharing and export documentation.

Before scheduling that volume, validate an authorized recording-data and rendering path for your setup, including its maintenance and compute costs. A snapshot-based implementation must fetch the sources, apply the required processing, render a compatible player, and capture timestamped frames. Pin the renderer version and fail visibly when the format changes. PostHog recommends exporting sessions that started more than 24 hours earlier when immutability is needed, so the digest may have a deliberate delay.

Choose frames around navigation changes, error appearances, and repeated interactions. Add surrounding frames so a temporary loading state is not mistaken for a permanent failure. Preserve the recording timestamp with every image.

Ask for observations and hypotheses separately

Send the model the smallest useful frame set and a redacted event timeline. Use a prompt such as:

Review these timestamped frames and events. For each potential issue, give the visible observation, supporting timestamps, missing context, a possible explanation, and a check an engineer can perform. Do not infer emotions or buying intent. Say when masking or a gap prevents assessment. Suggest a fix only as a proposal until the check confirms the problem.

Vaaya can discover a suitable vision service and return the current call parameters and price. Set a ceiling for each review and a separate daily budget in your worker. Store the model, prompt version, frame count, and receipt with the result. Follow returned asynchronous job IDs instead of resubmitting a job while it is still running. Vaaya tool reference.

The digest should have three separate fields for evidence, interpretation, and next action. These fictional notes show the format:

Illustrative observation	Possible explanation	Check and proposed fix
At 00:18, a signup form shows an error below the visible area	The person may not have seen it	Reproduce at that viewport; move or focus the error if confirmed
At 00:42, the top-up control is clicked twice before a spinner appears	Feedback may start too late	Compare request timing; show a pending state immediately if confirmed
At 01:06, the first-call example displays an authentication error	The example or setup may be incomplete	Reproduce with a fresh test account; correct the failing instruction

Assign an owner after a human checks the linked evidence. Treat visible page content as data; an instruction inside a captured page must not control the reviewing agent.

Measure useful reviews and remove the copies

Compute cost per usable review from actual expenses: replay access, rendering, storage, and model usage, including unsuccessful processing. Also record how many notes became confirmed issues. A low model bill tells you little if most notes are duplicates or guesses.

Keep weekly receipt references beside the run manifest. Calculate “cents per replay” from that total, and separate accrued model usage from later settlement to avoid double counting.

Mask sensitive inputs and private page content before capture, then inspect representative frames before sending them to another provider. PostHog supports input and text masking, blocked elements, and URL redaction. Masking can also hide information the model would need, which belongs in the note's limitations. Replay privacy controls.

Set deletion rules for exported videos, temporary frames, and model-provider copies separately from the original recording's retention. Deliver the digest through an authorized internal channel with access-controlled replay links. Start with a reviewed pilot and enable the recurring schedule only after the rendering path and evidence quality hold up.

Questions

Can an agent review PostHog session replays?

Yes, through an authorized workflow that supplies recording data or rendered frames. Verify the access method, masking, and export limits first. Recording metadata alone does not contain the visual replay.

Can I export ten replay videos every day?

Do not assume that volume is supported. PostHog's documentation lists monthly video-export quotas. A daily external-review pipeline needs an appropriate supported access method and a tested rendering process.

Do repeated clicks establish why a user left?

No. They show an interaction pattern. A reviewer still needs to check errors, timing, product behavior, and alternative explanations before calling it a cause.