Overcast where

Skill kdr/overcast/skills/overcast-where

Locate WHERE in a frame a target is — open-vocab detect, crop the box, then VLM-verify the crop so hallucinated boxes don't survive.From its SKILL.md

Install

npx -y skills add kdr/overcast --skill overcast-where

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

12 stars12 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

2.4 KB, 576 tokens by cl100k_base, as published. Nobody here has run it

overcast-where

Use this skill to turn a moment into spatial evidence: a bounding box on the target plus a verified crop. It follows the detector-proposes / VLM-verifies pattern (open-vocab detection like OWLv2, then confirm the crop) — because chat VLMs are unreliable at emitting raw coordinates, so a real detector draws the box and the VLM only judges the crop. Use the broad overcast skill and overcast/reference/verbs.md for exact flags.

Setup

see --detect needs a detection provider bound (boxes come from OWLv2, not the brain LLM):

scripts/visual-db-uv.sh --detect     # once: prints DETECT_PY (the venv python)
export DETECT_PY="$DETECT_PY"; overcast provider setup apply --preset owl-local --yes --json  # persists a portable shipped: ref for detect.py + the venv python

Workflow

Have the moment (use overcast-pinpoint / overcast-frame-grid): timestamp T on record REC.
Detect the target in that frame, then materialize + verify the box:

overcast see frame://REC@T --detect "<target phrase>" --json      # -> see record with detections[]
overcast crop <see-record-id> --all --class "<target phrase>" --pad 0.15 --json
overcast see <crop-path> --prompt "Does this crop show <target>? yes/no + describe" --json

The re-see of each crop is what kills false positives — open-vocab detectors emit confident boxes for almost any phrase at low thresholds.

Optionally sharpen the exhibit and record the finding:

overcast enhance <crop-path> --ops upscale,denoise --json
overcast finding create "<target> located at T" --ref <see-record-id> --confidence medium --json
overcast brief --export ./where.md --json

Output

Per confirmed target: the timestamp, the box (from the see --detect record), the crop path, and the verification verdict. Cite the see detection record.id + media.at; note the crop is the durable, memory-friendly evidence artifact.

Caveats

Never ask the brain LLM for coordinates directly — bind a detector and verify crops. Detector confidence is not calibrated across free-form phrases: a high score on a rare phrase can still be wrong, so the crop re-check decides. For a specific PERSON, use face --match instead of --detect. Boxes are per sampled frame, not tracks.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.