Requirements

Clueso MCP connected. Nothing else.

Inputs

The recording — either:
- a file the user provides → upload_file, then poll check_uploads until processed; or
- a fresh capture → record_screen (available when the workspace has recording enabled — check with get_capabilities; if unavailable, ask for a file instead).
What the demo shows — the feature/flow name and the 1-line takeaway per major step. If the user can't list steps, derive them by watching the recording's structure (clip timestamps + any narration via analyze_audio) and confirm your step list with the user before editing.
Tone — narrated product demo (default) or captions-only silent demo.

Workflow

1. Confirm workspace, ingest the recording

find(type='workspaces') → confirm → create_project → ingest the recording (upload_file/record_screen, check_uploads until ready), add_clips to place it on the timeline.

If the recording has spoken audio, run analyze_audio first — the existing narration tells you where the step boundaries and dead time are.

2. Cut the dead time

Raw captures are mostly waiting. Use split_clip at each step boundary, then:

Remove or compress stretches where nothing meaningful happens (page loads, typing long form fields, mouse wandering) — remove_clip the dead segments or tighten with update_clips.
Target: no shot where the screen is effectively idle for more than ~2 seconds.
Keep one breath (~0.5s) after each completed action so viewers register the result.

3. Write the narration against the cut

One or two sentences per step: what's being done and why it matters. Present tense, second person. estimate_duration per step — where narration outruns footage, either tighten the words or let the footage of that step run slightly longer; never rush the voice.

Skip this if the user chose captions-only.

4. Direct the viewer's eye

Fetch get_element_schema, then per step:

Zoom & pan — keyframe scale/position so the camera pushes toward the region being used, holds, and releases. One move per step; constant zooming reads as seasickness, a static full-screen recording reads as unedited.
Highlights — a keyframed rectangle or ring landing on the control as it's clicked, timed to the narration's action word.
Captions — short step labels (3-6 words) entering with animation_setting presets (slide, masked_reveal), swapped out as steps change. For captions-only demos these carry the narration's job: slightly longer, but still one line at a time.

5. Bookends

Intro clip (~3s): product/feature name + one-line promise, kinetic type on the committed palette (brand colors if the workspace has them; otherwise get_design_guide and pick).
Outro clip (~3s): the CTA. One action.

6. Audio pass

Narrated: set_voice → voiceover_batch for all steps → auto_sync, then add_sync_point so each zoom/highlight lands on its action word.
Optional bed: add_audio for subtle background music; analyze_audio / update_audio to keep it well under the voice.
Captions-only: music bed slightly more present, sync captions to the action instead of a voice.

7. Verify and export

get_clip(render=true, timestamp=…) at each step's action moment: is the zoom framing the right region, is the highlight on the control, is the caption legible and current? Watch the pacing math: total runtime should be a fraction of the raw capture (a 4-min raw take usually makes a 60-90s demo). Fix, then export_project and hand over the link.

Fallbacks

record_screen unavailable → ask for an uploaded file; don't attempt any local recording tooling.
Upload fails or stalls → report the check_uploads state and ask the user to retry; don't silently proceed without the footage.
Recording quality too low to zoom (tiny text, low resolution) → keep zooms gentle, lean harder on highlights and captions, and tell the user a higher-resolution capture would let the demo push in closer.
User's step list doesn't match the footage → trust the footage; show the user the discrepancy before cutting.