> An installable agent skill for AI Content Drop. Machine-readable copy: > https://aicontentdrop.com/skills/video-agent/SKILL.md. > Ships with the aicontentdrop package, and is listed in > /skills/index.json.

Handing the whole job to Drop, the video agent

A drop is one piece of work that carries a brief from start to finish: it researches ads that already work in the category, proposes concepts, writes a storyboard, renders, and assembles, on a board the person can open. The drop an agent starts here and the one a person starts on the website are the same record, planned by the same agent and charged the same way.

The five tools are start_video_agent, message_video_agent, get_video_agent_run, approve_video_agent_spend and list_video_agent_runs. Only approve_video_agent_spend can make a drop charge credits.

1. Check that the drop is here

The drop tools appear on tools/list for every signed-in account; a caller without a sign-in does not see them. When they are not listed, let the host sign the person in. If that is not possible, say so in one sentence and do the job with the precise tools instead:

Never imitate a drop by chaining those tools while calling it one.

2. Offer the drop for the right size of job

A single clip is faster and cheaper through generate_video. Offer a drop when the person asks for a campaign, a week of videos, several videos from one brief, an ad modelled on what competitors run, or when they want the whole job handled.

3. Open it, then send the brief

1. start_video_agent with the goal in one sentence. It opens the drop and returns a

run_id; nothing is planned and nothing is charged yet. budget_credits is an optional ceiling for the whole drop. Setting it is a spending decision, so set it only when the person gives a number.

2. message_video_agent with that run_id and the brief. A useful brief says: the product

or subject, who it is for, what the video is for, the length, the format (vertical, square, wide), and any assets. Pass images or clips as public https URLs in attachments (at most 8). A finished product ad needs a product image and a presenter image; attach them or expect the drop to ask.

The drop may answer with questions before it plans. Relay them to the person as they are and send the answers with message_video_agent. Do not answer for the person.

Starting from a reference

When the person shows a clip or a still to copy rather than describing the result, read it into words first and start the drop from those words. The full walk is the recreate-reference skill; the short form:

1. Ask who holds the rights (own_work, licensed or permission) and pass

the answer as given, as rights_basis (CLI: --rights). Never choose a basis for the person. There is no default: a video read without rights_basis is refused (rights_required), and the answer is to ask.

2. analyze_reference with the url of the clip or still (a public https

link) and the rights_basis (CLI: acd analyze <file|url> --rights <basis>). Free, capped per day, words only: the reference_class with its confidence, a summary, the shots, cast, look and lines. Show the person what it saw; let them correct the class.

3. start_video_agent with the goal, reference_url exactly as the read

returned it, and reference_class when the person confirmed or corrected it (CLI: acd drop "<goal>" --reference <file|url>). Starting charges nothing.

When the reference tools are not listed, or the read answers vision_unavailable, say so in one sentence and run the drop from a written brief instead of describing the reference yourself.

4. Follow a turn that outlives the call

message_video_agent returns within about 40 seconds. When a turn runs longer, the call comes back with status: running and an empty answer; the turn keeps going. Do not send the message again (a second send while it runs is refused as "already working on the last message"). Call get_video_agent_run every 15 seconds instead; its answer carries the reply once it lands.

get_video_agent_run returns the phase (intake, research, concept, plan, produce, assemble, distribute, done), the status, the latest answer, artifacts keyed by kind, artifact_counts, the board_link, credits_spent, the drop's own narration in trace, and pending_approval. Report a change of phase or a finished artifact, not every poll. Stop after 15 minutes of no change and hand over the board link.

Reading the scorecard

Once the drop has made pieces, get_video_agent_run carries the scorecard:

• pieces: one entry per made piece with its score (0–100), verdict and

reasons, and a retry_offer while another take the director suggested waits for a yes (section 5).

• director_notes: the Director's notes for the whole run, each naming the

direction rule (skill) and why it fired (why).

• reference_class_label: for a drop started from a reference, its class in

plain words.

The verdicts: accept (stands as made), retry or switch (the director suggests another take, with a changed prompt or on another model; that take is a new paid step made only after the person approves it, and arrives as a new piece), accepted_with_notes (it stands and nothing waits on it; the director may suggest another take, made only if the person approves it), unscored (not judged; its score is not a grade). The CLI prints all of this with acd drop status <run_id> --notes.

Hand the score to the person with its reasons. It is the drop's opinion in words, never a reason to approve anything on their behalf.

5. Approve the exact step, with the number

When a step would charge credits, the drop stops with status: waiting_approval and pending_approval holds task_id, quote_id, credits (the ceiling for that step) and what (the step in plain words). Ask in words, with the number:

> The drop wants to render the two-shot ad now: up to 152 credits, charged only > if it succeeds. Go ahead?

Read the figure from pending_approval.credits, never from memory. Then call approve_video_agent_spend with the run_id, and the task_id and quote_id exactly as given. decision: "decline" cancels that step instead.

• Approving the same quote twice does nothing the second time; it never charges twice.
• confirm_each is the only approval mode: every paid step asks. Approving never

lifts budget_credits.

• When a price has lapsed, the approval is refused with QUOTE_EXPIRED and

nothing is released. The drop prices the step again by itself: read get_video_agent_run (the quote_id is null for about a minute while it re-prices) and ask again only if the number changed.

What the drop makes depends on the brief. A product ad is cut from two or three shots (10 or 15 seconds); a single take is one clip priced as one clip; research, concepts, storyboards and the board cost nothing. Each rendered step shows its own quote.

A retry offer is the same gate on another take the director suggested for a piece (the surfaces call it "Try the other model"): the piece's retry_offer carries the task_id, quote_id and credits of that step, and while it is the next decision it is also pending_approval. Read the number from the offer, ask the person with it, and only after their yes approve it with approve_video_agent_spend and the task_id and quote_id exactly as given (CLI: acd drop retry <run_id> <piece_id> --quote <quote_id> with the quote_id the person saw; without --quote the command refuses and sends nothing, and when the retry was priced again since, it approves nothing and prints the new quote and credits to put to the person). One offer per piece; a declined offer leaves the piece as it was, with its score.

6. Pick up an earlier drop

list_video_agent_runs returns the account's drops, newest first, with phase, status, credits_spent and board link. Resume with get_video_agent_run on the run_id. A drop with status: done takes no more messages; start a new one.

7. Refusals

What to hand back

The board link, the finished videos with their links, credits_spent for the whole drop, and any step still waiting for a decision with its number, so the person can finish it on the website.

---

Install the plugin: https://aicontentdrop.com/plugin · All skills: https://aicontentdrop.com/skills