Advisor & Session Summaries

Prism generates two distinct LLM-authored artifacts to help you understand and improve a session.

Session summaries

After a session’s telemetry lands, the Prism Engine runs a summary worker (apps/prism-engine/src/summary_worker.rs) that produces a short narrative of what happened: the goal you started with, the path the turns took, the outcome, and the two or three turns that mattered most.

Summaries appear on the session detail page (drill into any session from /my/explorer) and inside the coaching cards on your report. They aren’t a transcript — they’re meant to be scannable in 10 seconds.

The advisor

The Advisor produces per-session coaching cards — short, specific suggestions keyed to the weakest Skill input on that session. Each card has:

• The weakness (e.g. “PES — low information density”)
• One concrete habit change (“add file paths and a line number”)
• A rewrite of a prompt from the session that would have scored higher

Advisor output is stored in prism.advisor_cards and shows up in:

• /my/report under “Worst prompts” and the habits chapter
• The plugin command /prism:advisor when you pass a prompt directly
• Scheduled report emails / Slack, if configured

Real-time vs. after-the-fact

There are two advisor surfaces — don’t conflate them:

Surface	When it runs	Code path
Plugin advisor skill	Inside Claude Code, on each prompt you submit	Plugin (Claude-side skill)
Dashboard / email advisor	After the session, in the engine	apps/prism-engine/src/intelligence/advisor*.rs

The plugin version is fast and local and gives single-line nudges. The engine version is slower, richer, and designed to be read on the dashboard or in your inbox.

Opting out

You can disable LLM-generated summaries for an organization in the Settings page (governance tab). Disabling summaries does not disable scoring — the rubric runs independently.