--- name: find-references description: >- Find real references for a video before anything is made, then turn them into ten visual ideas. The agent goes through seven sources in order (varg how to, X from the last 7 days, Refs, FrameThrower, the film-still libraries, Pinterest and Are.na, YouTube) and comes back with four kinds of reference: style (palette, light, texture, type), film (framing, lens, camera move), pop culture (moments people recognize) and trends (what people saved this week, with the numbers). It writes refs.md, trends.md, links.md and ideas.md and builds one board to pick from. Use when someone says "find references", "reference board", "what should this look like", "give me visual ideas", "moodboard references", "make it not look like AI", or starts a video with no look yet. Not for writing the look bible from chosen references (varg-video-prompting) or for downloading one known link (media-download). license: MIT metadata: author: vargHQ version: "1.0.0" openclaw: requires: anyBins: - curl - python3 homepage: https://github.com/vargHQ/skills/tree/main/find-references compatibility: >- Works in any agent with a shell and web access. Refs search is free with no key: connect it as MCP or call it with scripts/mcp_call.py (Python 3, standard library); its frame sheets and blueprints need Refs Pro. FrameThrower needs a FrameThrower account (OAuth sign-in, $2 free credit, then pay per use), so connect it as MCP. X needs an X MCP or API; without one the agent falls back to web search and says so. Pinterest images need gallery-dl (via the media-download skill), YouTube clips need yt-dlp. No varg credits are used. allowed-tools: Bash(curl:*) Bash(python3:*) Bash(yt-dlp:*) Bash(gallery-dl:*) Bash(ffprobe:*) Read Write Edit --- # Find references Without a reference the agent falls back to its default look: "centered text, gradient background, everything fading in" (https://x.com/rexan_wong/status/2103707054108299437, 18K bookmarks). "Most AI videos look cheap because nobody gives the AI a real film reference" (https://x.com/ScaleWthAI/status/2111821935021863254). A search that stays in the middle finds the same average. This skill sends the agent to seven places, in a fixed order, for four kinds of reference, and makes it show what it found before anyone writes a prompt. ## What you come back with Four kinds of reference ("atoms"). Each kept reference is one of them: | Atom | What it lends | Minimum | |---|---|---| | **Style** | palette, light quality, texture and grain, type, grade | 3 | | **Film** | framing, lens, camera move, cut, staging | 3 | | **Pop culture** | a moment people recognize: a music video, an anime opening, a game cinematic, an iconic costume, an era, an ad everyone remembers | 2 | | **Trend** | what people saved this week in this format, with likes, bookmarks, views and date | 3 | Then **ten visual ideas**, each built from at least two references of different kinds. Files, all in the project folder: ``` brief.md the brief in 6 lines refs/ every frame or clip the agent actually opened, named NN-atom-source.ext refs.md one entry per kept reference (template in references/templates.md) trends.md the trend posts with their numbers links.md search links for the sites only a person may open ideas.md the ten visual ideas board.html everything numbered on one page, to pick from ``` ## Rules 1. **Open every reference you keep.** Never describe a frame you did not see, and never invent a film, a scene or a post. If a source is out of reach, write what you could not open. 2. **Respect the sites that block bots.** Film-Grab, Shot Cafe, Eyecannndy, ShotDeck and Film Vibes forbid crawling and automated use (robots.txt, Cloudflare, terms). Never fetch them. Build search links for the person in `links.md`; frames the person drops into `refs/` count as references. 3. **Borrow the look, not the content.** Palette, framing, lens, light, texture and motion. Never actors, characters, logos, recognizable people or "in the style of" a living filmmaker. 4. **Found frames are for words and for judging, not for feeding a generator.** Whatever sits in a reference image comes back in the shot, including someone else's actor, set or logo. Feed a model only frames the person owns or licensed, and later your own approved stills. 5. **Skip the first, obvious pick.** For every idea, drop what comes to mind first and look one step further (for a sci-fi look, not Blade Runner). The middle of a search is the average you are trying to avoid. 6. **Give every reference one job** and say what it must not copy: "The reference image defines WHO the person is, not the pose, location, expression, clothing, lighting, or composition" (https://x.com/swiat_ai/status/2100482637341012201). 7. **Numbers come from the source.** Copy likes, bookmarks, views, installs and dates as the source shows them; never round them up or guess them. Bookmarks show usefulness better than likes. ## Step by step ### 0. Brief Write `brief.md` in six lines: what the video is, who it is for, where it plays (platform, aspect ratio), length, the brand or product, and taste (one or two things the person loves, one thing they hate). If the person already gave references, they go first in `refs/` and every search starts from them. If the brief is missing a line, ask once, then go on with a stated assumption. Turn the brief into search words: name the feeling, not the subject ("lonely, sodium-lit, wet" finds a look, "night market" finds stock photos), and add attributes (color, light direction, lens, decade, film stock, shot size, mood). ### 1. varg how to: what already worked Search the catalog for the format and the look. It returns style cards, tutorials and prompt cards taken from viral posts, with their bookmarks and views. ```bash curl -s "https://varg.ai/how-to/api/search?q=" ``` Or `search_how_to` on the varg how to MCP, or `vargai search ""`. Run 3 to 5 searches, not 50. Search results carry no numbers: open the 2 or 3 cards closest to the brief (add `.md` to the card's link) and copy their installs, GitHub stars or, for prompt cards, the bookmarks and views of the post they came from. Keep them as style or trend atoms. ### 2. X, the last 7 days: what people save now Search for the format, the look and the tools, newest week only, and rank by bookmarks: ``` since: min_faves:100 -filter:retweets prompt since: min_faves:50 ``` X has no sort by bookmarks: filter with `min_faves:` and rank by bookmarks yourself. Ask for 10 to 20 posts per search, not 100: posts often carry whole prompts and long results get cut off. Read the top posts; the prompt is in the post, its link or the author's first reply. Watch the video before you keep a post: the text and the clip do not always match. For each kept post write a line in `trends.md`: link, author and followers, date, likes, bookmarks, views, and the look or format in one line. Posts with far more bookmarks than the author's following suggests, and bait hooks, are suspect: keep them only when nothing else shows the same trend. Use an X MCP or the X API if one is connected. Without one, search the web for `site:x.com `, keep only posts you could open, and write in `trends.md` that X search was not available. Extra trend sources when they fit: the Civitai MCP (what AI looks are popular this week, with prompts; needs a Civitai key), TikTok Creative Center and the Meta Ad Library for ads (see `varg-ad-library-remix`). ### 3. Refs: films like this one, measured Refs (https://refs.video) holds 4,000+ launch films, product demos, explainers, motion pieces, teasers and UGC ads, each taken apart shot by shot (cut map, pacing, palette, type, voice, music). Searching is free with no key and returns short cards: title, maker, length, aspect, summary, thumbnail and a link that plays the film. - `search_videos` with the brief in plain words (filters: category, aspect, seconds, platform, brand). - Look at the thumbnails and open the 2 or 3 best on refs.video or their original post. - With Refs Pro: `get_frames` for labelled frame sheets and `get_blueprint` for the beats, shot list and style guide of the one you keep. Without Pro, say which films you could only see as cards. Connect it as MCP (`claude mcp add --transport http refs https://refs.video/mcp`) or call it with the script: ```bash python3 scripts/mcp_call.py https://refs.video/mcp search_videos '{"query":"founder-led launch film under 60 s","limit":8}' ``` Films play from their original posts and stay their makers'. Do not download or bulk-export them. ### 4. FrameThrower: film frames, ads and animation FrameThrower (https://framethrower.ai) is a cinematography library indexed frame by frame, with credits. It needs a FrameThrower account: connect it as MCP and sign in once (`claude mcp add --transport http framethrower https://framethrower.ai/api/mcp`, then `/mcp` to authenticate); new accounts get $2 of credit, then it is pay per use. It covers three of the four atoms: - **Film:** `search_frames` by concept, mood or color; `find_by_craft` by lens, shot size, visual style, setting, time of day, director or years; `find_similar` for "more like #2". - **Ads:** `search_commercials` and `search_commercial_shots` by idea or by what is in the frame. - **Pop culture:** `search_animation_frames` and `find_animation_titles` (anime and Western animation). Example call: `find_by_craft` with `{"lens": "anamorphic", "time_of_day": "night", "limit": 8}`. Use the frames to see and name the look. They are someone else's films: rule 4 applies. If FrameThrower is not connected, say so and lean on step 5 for film atoms. ### 5. Film-still libraries: links for the person These are the best human libraries and they forbid bots. Write searches into `links.md`, grouped by what to look for, and ask the person to open them and drop 1 to 3 frames each into `refs/`: | Site | Link pattern | Good for | |---|---|---| | Film-Grab | `https://film-grab.com/?s=` | full sets of stills from one film; lens, stock and process words | | Shot Cafe | `https://shot.cafe/tag/` (spaces as `%20`) | shot type, light, day or night, interior or exterior | | Eyecannndy | `https://eyecannndy.com/technique/` | the name of a technique you can see but cannot name | | ShotDeck | `https://shotdeck.com` (search needs a login) | color, light, camera angle, location; shot cards list palette, lens, lighting and film stock | Never fetch these pages, not even to check a link. ### 6. Pinterest and Are.na: curated boards People have already curated costumes, grades, posters, sets, cosplay and eras. Search by attribute, not subject: ` still`, ` color palette`, ` film grain`, ` cosplay`, ` lighting`. - Pinterest images: the `media-download` skill (`scripts/pinterest.sh -n 8 -o refs "" ""`, gallery-dl underneath, run through uvx if it is not installed). Its `references/pinterest-presets.md` has ready query sets. Saved files are not named by pin, so link the search page as the source. - Are.na: the Are.na MCP (`https://mcp.are.na/mcp`) or its API to search channels and read the person's boards; it can hold the final board. Its image files refuse direct downloads: put the channels in `links.md` and do not work around the block. ### 7. YouTube: motion and pop culture Music videos, trailers, title sequences, game cinematics, anime openings and ads people remember. Search, watch the 2 or 3 best, and note the timestamps of the moments you would borrow. Pull a 5 to 15 second clip with yt-dlp (`media-download`'s `video.sh`) only where motion or pacing matters, and keep it for reference. ### 8. Board and picks Put every candidate on one page, numbered, with its atom, its source link and one line on what it lends: use the `varg-mockups` board (`templates/board.html`) or a plain HTML contact sheet. Aim for 20 to 30 candidates. Keep the 12 or more that are strongest and most different from each other, covering every atom at its minimum. When the person is present, let them pick; otherwise pick yourself and say why. ### 9. refs.md One entry per kept reference, in the format in `references/templates.md`: number, atom, source and link, what it lends (palette with named colors, framing, light direction and hardness, lens, texture, motion), its one job and what it must not copy. ### 10. Ten visual ideas Write `ideas.md`. Each idea: - a title and a one-line pitch - the references it combines, by number, from at least two different atoms - the look line: medium, palette with the accent named, lens, light, grain, motion - the first shot in one sentence - why now: the trend post it rides, with its numbers; use a different trend for each idea where you can (wild ideas may have none) - the risk: what could make it slop, and any rights question Spread them: 3 close to what is working this week, 4 stretches, 3 wild. No two ideas share the same main reference. None may fall back to centered text, a gradient background and everything fading in. Add the ideas to the board and stop there: the person picks one or two, and the look bible comes next (`varg-video-prompting`, look line plus a Never list). ## Done when - `refs.md` has 12 or more references the agent or the person opened: 3+ style, 3+ film, 2+ pop culture and 3+ trends from the last 7 days with their numbers. - Every entry has a source link, what it lends, one job and a do-not-copy line. - `links.md` lists the searches only a person may open. - `ideas.md` has 10 ideas, each from 2+ references of different atoms, spread 3 / 4 / 3. - `board.html` shows every reference and idea, numbered, with sources. - Nothing in any file was invented, and no blocked site was fetched. ## The same thing as one prompt `references/goal-prompt.md` holds this skill as one prompt to paste after `/goal` in Claude Code (or as a task in any agent). `references/sources.md` has the seven sources in detail: tools, queries and what each one may be used for. ## Related skills - `media-download`: Pinterest images and YouTube or X clips for steps 6 and 7. - `varg-mockups`: the review board for steps 8 and 10. - `varg-video-prompting`: the look bible and Never list once the person picks an idea; its `references/look-references.md` is the short version of this skill. - `how-to-make-ai-video`: the whole ten-step film; this skill is its step 1.