菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Give any agent senses (video/audio/image understanding) and OSINT reach (search/capture/monitor) organized around an investigation case. Use when the user wants to analyze a video/audio/image, scan or monitor sources for a target, or ask/brief over accumulated findings. Drives the `overcast` CLI (built on pi + the tinycloud/Cloudglue perception backend); see reference/verbs.md for the full verb surface.
Install and configure overcast for this harness: install the CLI, verify the system ffmpeg, and configure reusable provider profiles. Use once per machine/profile before driving the `overcast` skill.
Turn a case into a visual evidence board — materialize face and object crops, link the same person across clips, connect themes, and render the corkboard as a CSI brief plus a live monitor wall.
Make unreadable footage legible — denoise/upscale a marked moment, re-run OCR and detection on the enhanced output, and record what was recovered with honest provenance (interpolation is a lead, not proof).
Work out where a photo or clip was taken — check embedded EXIF/GPS first, then pull signage, landmarks, and terrain clues, reverse-image-search the strongest ones, and corroborate to a location with cited evidence.
Run a standing surveillance watch on public sources for a target — auto-sense new media, auto-suggest matches into a triage queue, and keep a live control-room wall — so new evidence surfaces itself over time.
| name | overcast-where |
| description | 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. |
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.
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
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.
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
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.
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.