> For the complete documentation index, see [llms.txt](https://getsquish.gitbook.io/squish/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://getsquish.gitbook.io/squish/recipes/prompt-library.md). # Prompt library Every prompt in this library follows one pattern: **tell the model what the image is before asking anything of it.** A Squish contact sheet is a grid of video frames, each cell stamped with its timecode — but a vision model that receives it cold just sees "an image with small pictures in it". So each prompt opens with the same primer sentence: > This image is a video contact sheet — frames sampled in order across a video clip, each labeled with its timecode. With the primer in place, the model reads the grid as a timeline and cites the stamped timecodes back. Because the primer is baked into every prompt below, each one is self-contained: make a sheet (see the [CLI quickstart](/squish/getting-started/quickstart-cli.md) or [getsquish.app](https://getsquish.app)), paste the image into ChatGPT, Claude, Gemini — or a local vision model in LM Studio or Ollama — add one prompt, and go. The same library with copy buttons lives at [getsquish.app/prompts](https://getsquish.app/prompts). ## Summarize the video ``` This image is a video contact sheet — frames sampled in order across a video clip, each labeled with its timecode. Summarize what happens across the whole clip, then list each major moment as "timecode — what happens". ``` Use when you want the whole clip understood in one pass — a summary first, then an indexed list of moments you can jump to. ## Find a specific moment ``` This image is a video contact sheet — frames sampled in order across a video clip, each labeled with its timecode. Find where [describe the moment you're looking for] happens and give me its timecode. If it appears more than once, list them all; if it never appears, say so. ``` Use when you're hunting for one event in a clip — fill in the bracket. The "say so" clause keeps the model from inventing a timecode when the moment isn't on the sheet. ## Play-by-play timeline ``` This image is a video contact sheet — frames sampled in order across a video clip, each labeled with its timecode. Give me a chronological play-by-play of the clip as a list, one line per event, in the format "timecode — what happens". ``` Use when you want a complete event-by-event log of the clip rather than a summary — the strict one-line format makes the output easy to scan or paste into notes. ## Steps from a tutorial or screen recording ``` This image is a video contact sheet — frames sampled in order across a video clip, each labeled with its timecode. This is a screen recording of someone completing a task. List the steps they take in order, each with the timecode where it happens. ``` Use to turn a screen recording or tutorial video into a numbered, timestamped how-to. ## What changes over time ``` This image is a video contact sheet — frames sampled in order across a video clip, each labeled with its timecode. Describe how things change from the first frame to the last: what is different, and around which timecode each change happens. ``` Use for before/after questions — progress on a build, a scene evolving, a UI drifting between states — where the answer is the *difference* between frames, not any single frame. ## Debug a screen recording ``` This image is a video contact sheet — frames sampled in order across a video clip, each labeled with its timecode. This screen recording shows a bug. Tell me what goes wrong, the timecode where it first appears, and the steps that led to it. ``` Use on a bug-repro recording: the model pinpoints where the failure first shows up and reconstructs the steps that preceded it — a timestamped bug report from one paste. ## Read on-screen text ``` This image is a video contact sheet — frames sampled in order across a video clip, each labeled with its timecode. Read any on-screen text, captions, or UI labels visible in the frames, grouped by timecode. ``` Use to extract captions, terminal output, UI labels, or slide text from a clip — grouped by timecode so you know when each piece of text was on screen. ## Going deeper If a prompt's answer lands *between* two cells, don't re-prompt harder — re-run Squish with `--start`/`--end` set to the timecodes around the gap and paste the denser sheet. That's the [navigation loop](/squish/the-primitive/the-navigation-loop.md); for agents that should run the whole loop themselves, install the [agent skill](/squish/recipes/agent-skill.md). --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter: ``` GET https://getsquish.gitbook.io/squish/recipes/prompt-library.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.