About Horizon

The 5-minute version of what this product does.

What Horizon is

Horizon is a loop that finds patterns in your Chatbot’s conversations, lets you edit the Playbook (Metaprompt) to fix them, and proves the fix works before you publish.

The loop, in 5 stages

Stage 1·Run (Eval Run)

Luna runs N synthetic conversations against test customers you’ve pre-configured. The chatbot uses the currently-published Playbook. Each chat is scored automatically.

New Run →

Stage 2·Diagnosis

After a Run completes, Horizon analyzes what went wrong across all chats and surfaces patterns— repeated failure modes with examples, severity, and plain-English descriptions. The top pattern becomes the run’s headline.

See recent diagnoses →

Stage 3·Draft (Playbook Editor)

You edit the Playbook section-by-section in the Playbook Editor. Claude can suggest edits per section. Saving creates a Draft / Candidate Playbook version. The live Playbook is not affected until you Publish.

Open Editor →

Stage 4·Sanity Check / Verification

Horizon replays the original chats under your Draft / Candidate Playbook and compares before/after. You see which conversations improved and whether any regressed against your reference conversations.

Open a run →

Stage 5·Publish (Deploy)

If the Sanity Check / Verification looks good, you publish the Draft / Candidate Playbook. It becomes the live Playbook for all new conversations.

Most recent run →

Supporting tools

Roleplay Sandbox— type lead messages and see how Luna replies under your Draft, with no database writes. Inside the Playbook Editor as a tab.
Your Review (Human Annotation)— on any session, disagree with the auto-label. Your reviews count 3× the auto-label when re-analyzing a Run.
Safety Rule Fired (Guardrail Triggered) feedback— rate each fire as good-catch or false-positive. Builds per-trigger precision. On the Safety Rules page.
Side-by-Side viewer— from any Sanity Check, open original vs replay for a single conversation, with the first divergent turn highlighted.
Weekly Digest— on Settings, preview or send a weekly summary of runs, deploys, fires, and reviews to Slack.

Developer mode

Toggle in Settings. When on, Horizon reveals internal identifiers (rule ids, run statuses) below each label. When off, only the plain labels show. The vocabulary pairs like “Playbook (Metaprompt)” are always visible regardless of mode.

Glossary

Every pattern, safety rule, and signal Horizon tracks, in plain English: Open the glossary →

Multiple chatbots

This page lives at /l/horizonbecause it’s the Luna-specific view. Each chatbot the engine supports gets its own /<product>/horizon path with the same vocabulary and the same 5-stage loop.

Keyboard shortcuts

? — open this page’s tour
n — new Run (from Horizon home)
/ — focus search (on Sessions)
g h — go to Horizon home
g g — go to Glossary
g a — go to Activity
g p — go to Playbook
Esc — close tour / modal