How do I install the Overcast Visual Target Search skill?

Run npx skillmds@latest add kdr/overcast-visual-target-search in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Overcast Visual Target Search safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Overcast Visual Target Search?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Overcast Visual Target Search free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Overcast Visual Target Search?

kdr (@kdr) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Overcast Visual Target Search by kdr

Name: Overcast Visual Target Search
Author: kdr

File contents

overcast-visual-target-search

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.

Workflow

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

Output

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.

Caveats

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.

---name: overcast-visual-target-searchdescription: Find a person, logo, object, landmark, or visual reference across local clips or captured media with timestamped Overcast evidence.---# overcast-visual-target-searchUse 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.## WorkflowFor a person with a reference image:```bashovercast doctor --jsonovercast case init --jsonovercast face ./clip.mp4 --match ./person.jpg --json # a match ≥75% auto-suggests a findingovercast 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 --jsonovercast ask "where does the reference person appear, with timestamps and confidence?" --jsonovercast 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 --presetowl-local --yes`, which persists a portable `shipped:` ref for detect.py and uses thevenv python, NOT system `python3` which lacks torch/transformers):```bashovercast see ./clip.mp4 --detect "red backpack" --jsonovercast crop <see-record-id> --all --class "red backpack" --jsonovercast ask "list target detections with timestamps, confidence, and crop paths" --json```For logos, landmarks, or near-duplicate visual references:```bashovercast index create refs --type image-ransac --local --jsonovercast index add ./reference-logo.png --to <index-id> --jsonovercast image match ./clip.mp4 --index <index-id> --json # a RANSAC hit auto-suggests a findingovercast finding list --state triage --json # then accept/dismiss the lead```## OutputReturn 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 anun-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.## CaveatsFace 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, thensynthesize with `ask` and `brief`.

Frequently asked questions

Overcast Visual Target Search