Skip to main content

Run contextual AI skills

AI skills are named actions with instructions and input requirements. Lotti filters the Generate… menu to the current entity, so a task does not offer an audio-only transcription action and an image does not offer an action that requires a recording.

All screenshots
On a habitat image linked to the task, the applicable action looks for pressure-gauge anomalies and visible seal damage.
ScreenshotMobile view
Project Waddle habitat image with the Generate modal offering Inspect habitat photo

Run a skill​

Narrated walkthrough: the Generate Coding Prompt skill assembling a task's description and checklist into a ready-to-paste prompt.
  1. Open a task or supported journal entry.
  2. Select the outlined assistant action in the entry header. This opens Generate….
  3. Read the skill description and select the action you want.
  4. If Lotti offers a model or thinking-mode picker, confirm the default or choose a one-off override.
  5. Keep working while the inference runs; the result is attached to the relevant task or entry when complete.

The available menu changes with context:

  • tasks can offer text and task-context actions;
  • audio entries can offer transcription and audio-aware actions;
  • images can offer visual analysis; and
  • text entries can offer prompt or generation actions that accept text.

Some skills need a linked task because they read broader task context or write their result back to that task. If an action is absent, first check the entry type and task link, then check that the active inference profile has a model for the required capability.

Model overrides are temporary​

When a skill offers a model picker, that choice applies to the current invocation. It does not silently rewrite the category default or inference profile. To change the normal route, edit the profile under Models and inference profiles.

Treat generated output as a draft​

Review generated prompts, summaries, transcripts, image analysis, and artwork before relying on them. Task context can be incomplete, providers can return incorrect output, and a capable model can still misunderstand a very serious sardine emergency.