Overcast Pinpoint

overcast-pinpoint

Use this skill to answer "exactly when does X happen?" in a clip and back it with evidence the model actually looked at. It mirrors the temporal-search pattern from the VLM video literature (T\* / VideoAgent): score cheaply over the whole clip, then spend expensive VLM calls only on a few candidate frames and zoom in. Use the broad overcast skill and overcast/reference/verbs.md for exact flags.

Two rules that make the answer trustworthy:

• Report a window, not a frame. Frame-exact localization is unreliable; emit

[t1-t2] plus one verified keyframe.

• Every timestamp must trace to a frame you see-verified. Never emit a

time the model merely guessed — models answer correctly while grounding on the wrong moment, so confirm by looking at the frame.

Workflow

1. Make the clip local and get a record id (see frame:// needs media on disk —

capture a remote clip first). watch also gives per-shot timestamped content to search:

overcast doctor --json
overcast case init --json
overcast watch ./clip.mp4 --json         # -> video.analysis record id (REC)

1. Get COARSE candidates cheaply (pick what's available):

overcast ask "moments where  happens, with timestamps" --json      # over watch shots/notes
overcast grid ./clip.mp4 --count 16 --json                            # one contact sheet ...
overcast see  --prompt "which numbered cells show ? give cell numbers" --json
overcast similar search "" --index  --json          # if a local CLIP index exists
overcast ask "moments  happens" --index  --probe --json  # remote index

For grid, translate the chosen cell number to a time via the grid record's payload.cells[n].at (don't trust a model-guessed time). CLIP/shots only SHORTLIST — CLIP is weak on actions/order — so verify next.

1. VERIFY + zoom on each candidate time T (expensive, precise):

overcast see frame://REC@T --prompt "Is  happening here? answer yes/no and what you see" --json
# refine: sample T-d and T+d, halve d each round until adjacent frames flip yesno
overcast see frame://REC@ --prompt "Is  happening?" --json
overcast see frame://REC@ --prompt "Is  happening?" --json

1. Record the verified window and eyeball it:

overcast note " occurs" --ref REC --at  --confidence medium --json
overcast view REC --at  --json
overcast brief --export ./pinpoint.md --json

Output

The tightest [t1-t2] window, the single see keyframe that confirms it (its record.id), and how it was found (shots / grid / CLIP / probe). Cite record.id + media.at for every claim; state the window, not a false-precise single frame.

Caveats

Needs the video local (a URL-only watch record can't extract frames). CLIP and shot text shortlist but don't decide — the see frame check does. If X recurs, pinpoint each occurrence separately; don't collapse them into one window.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kdr
• Source: kdr/overcast
• License: Apache-2.0
• Homepage: https://overcast.video

Install and usage instructions live in the source repository linked above.