AI Cockpit

Prompts editor

The Prompts tab in the AI Cockpit is the V28-native editor for the llm_prompts table — the set of named LLM prompts the app uses for analysis, classification, and assistant scaffolding.

This is the V28 sibling of the legacy Admin → LLM Prompts page. Both edit the same data.

Named prompt slugs

Each prompt has a stable slug — code calls loadConfigurablePromptText('message_analysis') etc., gets the current DB row, and falls back to a hardcoded default if the row is missing or invalid.

Common slugs you'll see:

  • message_analysis — classifies inbound guest messages
  • wa_chat_assistant — drives Pulse Copilot's high-level routing
  • task_template_validator — validates template edits (see Task Templates editor)
  • cleaner_walkthrough_intent_classifier — classifies cleaner WA free-text + photos (gh#212)
  • … (full list visible in the editor)

Editor surface

For each slug:

  • Prompt body — textarea, with {{placeholders}} highlighting.
  • Model — dropdown (same shape as the Agents editor's model field).
  • Save → DB → live on the next call.

Same audit-log treatment as the Agents editor: every save is logged with actor + diff + timestamp.

When to use Prompts vs Agents

Tab Use when
Agents You want to tune Pulse Copilot's specialist agents (accounting, messaging, etc.) — system prompt + model + thinking budget + tool rounds.
Prompts You want to tune a one-shot LLM call (message analysis, validator, classifier). These are not multi-turn agents.

A rule of thumb: if the LLM is called in a single round (input → output, done), it's a Prompt. If it's an agent that can call tools and loop, it's an Agent.


Related: gh#415 (Agents editor), gh#357 (V28 Webhook Secrets card cleanup — sibling V28-native operator surface).

Source: the FlatsBratislava operator manual.