메뉴
Find a person, logo, object, landmark, or visual reference across local clips or captured media with timestamped Overcast evidence.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
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-visual-target-search |
| description | Find a person, logo, object, landmark, or visual reference across local clips or captured media with timestamped Overcast evidence. |
Use this skill when the task is to locate a visual target across videos, images,
or captured case media. Use the broad overcast skill and
overcast/reference/verbs.md for exact flags.
For a person with a reference image:
overcast doctor --json
overcast case init --json
overcast face ./clip.mp4 --match ./person.jpg --json # a match ≥75% auto-suggests a finding
overcast finding list --state triage --json # triage the auto-suggested lead(s)…
overcast finding accept <id> --json # …accept a real match (or `dismiss <id>`)
overcast crop <face-record-id> --all --class face --json
overcast ask "where does the reference person appear, with timestamps and confidence?" --json
overcast brief --export ./visual-search.md --json # short by default; --full for the per-match timeline
For an object or open-vocabulary target (--detect needs a bound OWLv2 detector —
build it once with scripts/visual-db-uv.sh --detect (it prints DETECT_PY), then
export DETECT_PY=… and bind via the preset: overcast provider setup apply --preset owl-local --yes, which persists a portable shipped: ref for detect.py and uses the
venv python, NOT system python3 which lacks torch/transformers):
overcast see ./clip.mp4 --detect "red backpack" --json
overcast crop <see-record-id> --all --class "red backpack" --json
overcast ask "list target detections with timestamps, confidence, and crop paths" --json
For logos, landmarks, or near-duplicate visual references:
overcast index create refs --type image-ransac --local --json
overcast index add ./reference-logo.png --to <index-id> --json
overcast image match ./clip.mp4 --index <index-id> --json # a RANSAC hit auto-suggests a finding
overcast finding list --state triage --json # then accept/dismiss the lead
Return timestamped matches, similarity or confidence where available, source
record.id, media.at, and cropped evidence paths created by crop.
face --match / image match auto-suggest findings — resolve them with
finding list --state triage → accept/dismiss so a run doesn't leave an
un-triaged queue; the default brief is short, --full for the per-match timeline.
State whether the match came from face --match, see --detect, or local
image-ransac matching.
Face detections are sampled-frame detections, not unique-person counts. Use
face --match <image> for a specific person and include confidence caveats.
For exact evidence, use crop to materialize local image records, then
synthesize with ask and brief.