Creative Space
Build a Library of AI Demos You Have Tried

المحتوى

Provide: Post link; own TikTok/Xiaohongshu/etc. account for recent likes/bookmarks; or competitor/interest accounts to track.

Receive: Reproducibility findings, original/adapted prompt cards, user-asset Flova comparisons, inspiration-library entries, and optional periodic candidate lists.

Read how the author produced the work, distinguish claims from validation, try it with the user's content, compare, and retain lessons.

Three modes

  1. One post: capture/read, trial, library.
  2. Own account: inspect only a short recent period to identify current interests and candidate demos, not long-term profiling.
  3. Tracked accounts: identify worthwhile new demos; user selects before trials.

Use X Inspiration Studio for X's official MCP. This workflow covers browser-oriented platforms such as Xiaohongshu, Douyin, TikTok. Link-to-Prompts handles one link; this workflow adds discovery, trials, retention, tracking.

Tools/references

  • Platform MCP or computer use in a user-logged-in browser. Choosing access first matters: , steps 1–2, explains capabilities and recent-interest extraction.
  • Teardown — retrieve and read the full “Break Down a Reference Video into an Interactive Report” inspiration through Flova MCP by name: only the structure/visual/motion analysis needed for the chosen reproduction; host readers obtain post text.
  • Flova, required: current model capabilities, Quick trials/comparisons.
  • User Notion, Feishu Base, or local library, required.
  • , steps 3–4: post fields, demo types, difficulty, platform reading.
  • , steps 5–6: format, uploads/roles, reference wording, comparison.

Explain Flova quota use and any third-party plugin/API fees (paid directly to them) before use.

Boundaries

Read only: no likes/bookmarks/follows/comments; never request passwords; stop at CAPTCHA/risk controls. Attribute prompts to authors/original links. Do not bypass paywalls/DM/follow gates; users decide access. Do not republish others' media: library stores links/prompts/parameters and user's trial outputs; source images/video remain local comparison material. Curated author showcases may include post-production; matching outcomes are not guaranteed and author claims are not verified findings.

Workflow

Connection/confirm → optional recent interests → intake → assess/confirm → adapt/confirm → compare → library → optional tracking.

  1. Access — confirmation: For own/tracked accounts, check actual platform MCP, explain MCP versus computer-use differences, let the user choose. A supplied post link skips this and follows post-intake directly.
  2. Recent interests, optional: Inspect 7–14 days of likes/bookmarks; propose 3–5 current themes for confirmation. List reproducible saved demos for selection. Without account access, accept verbal interests or direct posts.
  3. Intake: Process links/screens/copied text individually, including batches. Use Teardown acquisition/post-intake. If unreadable, request screenshot/copied text; transcribe screenshot prompts verbatim for verification. Ask desired effect and replacement product/person/scene assets. “Save only” skips trials.
  4. Assess — confirmation: Record author/link, tools/models, original prompt (search body/comments/image text), parameters, prerequisites, showcased result, demo type. Query creation_generation_models, detail="full", for duration/ratio/reference counts/roles; otherwise unverified. Classify directly trialable, adapt then trial, other tools needed, or unreproducible with reason. If only artwork/no prompt, use Link-to-Prompts analysis. Confirm interpretation/selection.
  5. Adapt — confirmation: Preserve original unchanged as baseline. Create a separate user-asset version, adapted to available Flova models, removing author brands, real names, signature text. Use cards and distinguish original/adaptation. Confirm before trial.
  6. Trial: With agreement, Quick-generate original once when runnable, then adapted/user-asset version. Compare against the post, explaining matches/differences. Change one variable per further round, at most two more. If still failing, record unreproducible and known causes honestly.
  7. Library: Ask destination once (Notion/Feishu/local inspirations.md), reuse thereafter. Per row: title, link/author, type, tools/models, original prompt (inferred prompts labeled), adaptation, required assets, results/comparison, lessons, tags/status. Store only links/prompts/parameters/user outputs; others' media remain local. Deduplicate: update same URL; merge reposts of the same prompt.
  8. Track, optional: Confirm accounts/platform/access/focus/frequency/batch cap (suggest 5–10), save in library Tracking Settings or tracking.md. For competitors add one/two sentences of observed activity, not speculation. Run manually first; schedule after list approval. Every run is self-contained: read settings/new activity, select candidates, send list, update last-run time. Use host scheduling; without tools provide content for manual setup. MCP enables unattended runs; computer-use requires logged-in browser/computer on and may fail—disclose failure, do not bypass it. Tracking only returns candidates: no trials/library additions/Flova spend until selection, then resume step 3.

Deliver

Per-demo identification/reproducibility; original/adapted cards and comparisons; created/updated library links; recent-interest list if inspected; tracking setup/first candidate batch if enabled.

demo-<date>-<name>/
├── source/      Original screens/video (local comparison only), user assets
├── docs/        Interests, identification, cards
├── trial/       Outputs/comparisons
└── library/     Entries if not using Notion/Feishu

How to connect and learn about recent interests

Choose first: platform MCP or computer use

Before reading an account, decide how to read it. Sequence and judgment:

  1. See what tools are actually available in the current environment. List the connected plug-ins, connectors and MCP to see if they are available for this platform. Do not assume yes or no based on impressions.
  2. Select what you want to read:
What to readPrioritizeInstructions
The text, pictures, videos, and comments of a single public postPlatform MCP or public pageUse it if it cannot be read computer use
The public homepage and latest news of the designated accountPlatform MCP or public pagePublic content, no user login is required
User's own likes and collectionsAccount private content, usually computer useOfficial MCP generally does not provide this kind of data, and the actual tool list shall prevail; users can view it in the browser they have logged in
Periodic automatic readingPlatform MCPcomputer use requires the browser to remain logged in and in a human environment, and may not be able to run regularly unattended
  1. Tell the user the difference between the two methods and ask the user to choose:
    • MCP: Stable and can be run regularly, but the platforms and data covered are limited, and likes and collections are usually not available.
    • computer use: You can see the likes, favorite pages and domestic platforms after the user logs in, but it is slow and requires the user to log in by themselves. It may encounter verification codes or risk control, so it is not suitable for unattended use.
  2. If none of them are available, please take a screenshot of the collection page or send the link to the post you want to try.

The current situation (subject to the actual tools in the environment): TikTok, Instagram, and YouTube do not have official MCPs for ordinary users, and the official MCPs of TikTok and Meta only cover advertising; X official MCP can read posts, searches, timelines, bookmarks, and trends, and

When using computer use
  • Read only. Operate in the browser where the user is logged in, and do not ask for the account password.
  • Open the user's own likes page and favorites page, read back from the latest, and stop at the edge of the window.
  • No likes, no collections, no followings, no comments, and no operations that leave traces for users.
  • Do not turn pages quickly and continuously; stop when encountering verification code, login expiration, or risk control prompts, and ask the user to deal with it.

Learn about recent interests

When a user gives a link to his TikTok, Xiaohongshu, etc. account, he can first find out what he has been paying attention to recently from the likes and collections. The purpose is to serve immediate interests: to know what users want to try these days, not to build a long-term portrait of users.

How much to read
  • Time window: Default is the last 7–14 days, up to 30–50 items, user can adjust. Don’t read earlier.
  • Collection is more important than likes: Collection usually means "want to try it myself", while likes may just be passing by. When both are present, the collection weight is high.
  • Each note: link, author, title or text description, content type (tutorial, demo, case, inspiration picture, other), topic tag. Also note when interaction data is available.
Organized into "Recent Focus"
  • Summarize 3-5 themes, such as "Xiaohongshu product image style", "AI video transition" and "character consistency".
  • Each topic is attached with 2-3 examples collected by the user, and one sentence explains why they are grouped together.
  • Mark out which ones you want to do yourself (with prompts, tutorials, reproducible demos) and which ones you just want to look at.
  • Confirm and modify for users: which topic is right, which one is not, and what else they want to add.
How to use
  • Select a reproducible demo from the user's collection and enter the trial run process directly, without the user having to dig it out and send it again.
  • While continuing to follow, filter candidates based on recent topics of interest.
  • Only the topic list and post links are retained, with dates, and will be overwritten by the next refresh; the original text and media are not saved.
When the user does not want to give an account number

Asking users to dictate the types they want to try recently, or directly posting links, can also be used to filter candidates.

Read the post and judge whether it can be reproduced

What to take out of the post

ItemHow to find
Author and linkAccount name, homepage, post link, publishing time
PromptsText; the author's pinned reply and follow-up comments; text in the picture (screenshot prompts, verbatim transcription); external links in the post (document, Gist, Notion).
Tools and modelsName written by the author, hashtag, watermark in the screen or interface
Parametersaspect ratio, duration, resolution, seed, negative description, reference strength (noted until the author writes it)
Front-end assetsWhat reference pictures, reference videos, and audios were used? Did the author provide them
WorksPictures, videos, screenshots. Note down the duration and aspect ratio of the video, and take a few frames to see the key scenes
Information in commentsResults reproduced by others, problems encountered, and additional explanations from the author

If the item cannot be found, write "Post not provided" and do not speculate.

Several demo types

  1. Prompts are given directly: The complete prompts are written in the post, and the results are attached. The easiest to reproduce is to go directly to the test run.
  2. Prompt word plus reference assets: The prompt must be combined with the reference picture or reference video, such as "Use this picture to make the first frame." If the author of assets does not provide them, please use your own assets instead and explain that they will be different from the original post.
  3. Multi-step workflow: Several tools are used together, such as drawings, regenerating videos, and post-editing. Break it into steps and gradually map them to available tools.
  4. Only works: No prompts, only finished products. Use the method of "link copy prompt" to disassemble it. The obtained prompt is marked as "reverse prediction" and cannot be written as the author's original text.
  5. Display type: It is just a display of the effects of new functions or new models, and there is no reusable method. Note it down for future reference, don’t do a test run.
  6. Paid or privately available: Prompts are available behind a paywall, private message, or require attention. The source of the record is not bypassed and it is up to the user to decide whether to obtain it.

Reproduction difficulty judgment

DifficultyFeaturesHow to do
EasyOne-time generation, prompts are complete, the author's model Flova has or is similarTry it directly, once for the original text and once for rewriting
MediumMultiple reference pictures, specific parameters are needed, or the model capabilities do not fully correspondRewrite and post-test to explain the differences
DifficultyMultiple tools connected in series, requiring post-production or 3D, professional softwareDivide it into steps, only try the parts that can be done, and ask the user to confirm what to do with the rest
CannotThe model or tool has no corresponding capabilities, the assets are not available, and the author's results have obviously gone through extensive selection or post-productionRecord truthfully and do not force it

How to judge the credibility of a work

  • Are you showing a single result or a selection? There is only the best one and no failure cases, reminding users to lower their expectations. -Does the prompt match the work? No, the prompt may not be true.
  • Has anyone reproduced the success or failure in the comments? Write down the reasons.
  • Does the video have any obvious post-production (subtitles, transitions, color correction)? Those are not brought about by the prompts.

Several points on reading a single post

  • For the connection method, see .
  • X: For long posts and series posts, you need to read the entire series, and the prompts are often in the last few posts. If the text and media cannot be obtained, please provide screenshots.
  • Other platforms: Also look for prompts in subtitles, descriptions and pinned comments; if you cannot read them, please take a screenshot or copy the text.
  • Stop when encountering a login wall, verification code or risk control prompt, ask the user to handle it, and do not try to bypass it.

Prompt Cards and Flova Handoff

These inspirations deliver prompt cards, not images/videos: one per image or shot. Cards can be used directly for generation; this reference covers writing, Quick generation, and validation.

Card format

Include all fields; write “none” where absent:

FieldContent
IDE.g. P01, mapped to the requested image/source shot.
Model/capabilityE.g. Seedance 2.5 multimodal video; GPT Image 2 editing.
PurposeOne-sentence intended use.
PromptComplete copy-ready text in one code block.
ReferencesOne row each: filename, submission order, role below, responsibilities, exclusions.
ParametersDuration, aspect ratio, resolution, etc., from creation_generation_models; otherwise “unverified.”
InvariantsUnchanged people, products, text, composition.
RationaleTwo/three sentences linking choices to model-writing rules.
Acceptance checksE.g. identity, verbatim text, absence of teleporting motion.

+Write in the target language. Chinese-user requests default to Chinese unless the model guide/user calls for English. Preserve dialogue/on-screen text verbatim; never translate/polish it.

Quick generation

+Follow Flova Skill's Quick/reference guidance:

  1. Query the model through creation_generation_models, detail="full"; verify parameters, reference slots/roles.
  2. Prepare references: supported aspect ratio/padding, crops, formats.
  3. creation_upload_ticket → transfer bytes to its URL → creation_upload_complete → file_ref. It expires in roughly one hour; upload again before submission when necessary.
  4. Assign each inputs role: first_frame starts matching that image; last_frame ends on it; reference guides appearance/style/content without exact matching; driving_audio drives lips/action; base_video is edited/continued.
  5. Submit with client_request_id; retain card/reference/file_ref/task-ID mappings.

Referencing inputs in prompts

  • Use ordinal natural language, e.g. first image controls face/clothing, second controls product. Match inputs order within each media type; images/videos/audio each count from first.

Other tools

+Remove Flova-specific markers/role instructions and use the destination tool's upload mechanism and parameter controls. Explain that these prompts target a particular model; changing models may change outcomes.

Optional comparison

+Explain quota use before proceeding. Generate original and revised prompts once each with identical parameters/references; show side by side. Change one variable per round (e.g. reference roles or timeline), record edits/task IDs/results, and deliver the record. Generation is stochastic: one comparison shows a tendency; repeat important findings once or twice before concluding.

عملية الإبداع

  1. Flova

    Flova

    Try the reference approach

    Use your own assets to generate images or videos inspired by the reference.

  2. Notion

    Notion

    Save the inspiration

    Organize the original posts, prompts, and results of your experiments.