Index of every page on this site: llms.txt
All skills
Run it yourself

Collects references from the user's own saves and every site they pick onto one linked board before a design is built.

Use this Skill: https://skilld.dev/gh/raphaelsalaja/skills/gather

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

sources.md

≈3.1k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Sources

How to reach each source in SKILL.md, the files each search agent writes and the board settings.

Which sources

Before searching, ask which sources to use, with each name below as its own option, in this order. Connect any picked source that is missing, then search only the picked ones.

Source The user's saves The site
Pond Saves from every platform it imports None, it is a local library
marked.dev Saved refs and collections Explore, browsed by what moves
Are.na Channels, through Pond when connected Search across public channels
Cosmos Clusters, through Pond when connected Visual similarity, colour and motion search
Pinterest Boards, through Pond when connected Search, and a pin's More Ideas
Savee Saves, through Pond when connected Graphic, brand, editorial and web work, searched by colour and image
Dribbble Likes and collections, through Pond when connected Shots posted by designers, mostly concepts
Behance Appreciations and moodboards, through Pond when connected Case studies posted by designers, mostly concepts
Layers Checked when signed in Work posted by designers, mostly concepts
Mobbin Saved collections Screens, flows and marketing sections
Recent Checked when signed in High-end sites, branding, motion and type, picked daily. The Web filter shows sites only
Refero Checked when signed in Real product screens, flows, patterns and elements, mostly web
Siteinspire Checked when signed in Editorial, typographic sites, filtered by style, type and subject
Minimal Gallery Checked when signed in Restrained, functional sites
Land-book Checked when signed in Landing pages, filtered by page type and style
Saaspo Checked when signed in Software landing, pricing and product pages
Page Flows Checked when signed in Recorded flows from live web and mobile apps
Nicely Done Favourites and collections Screens, components and flows from SaaS apps, searchable by the text inside them
Design Spells Checked when signed in Micro-interactions, easter eggs and small details from shipped products
60fps Checked when signed in Recorded motion from mobile apps, tagged by gesture, pattern and effect
Awwwards Checked when signed in Award-winning sites, judged from paid submissions
One Page Love Checked when signed in Single-page sites and page sections, picked by hand
Lapa Ninja Checked when signed in Landing pages with full-page screenshots and recordings
Fonts In Use Checked when signed in Type in real work, indexed by typeface
Typewolf Checked when signed in Type on current sites, named by typeface
CodePen Likes and collections Live front-end experiments, mostly concepts

One option covers both columns. A picked source gets a search of the user's saves there and a search of the site itself. Mark Are.na, Cosmos, Pinterest, Savee, Dribbble and Behance "through Pond" when Pond is connected.

For a site with accounts, open it in the browser and look for the user's likes, boards, collections or bookmarks before searching. When it shows signed out, ask once whether to sign in or search it signed out. A site with no account, or none signed in, gets the site search only.

Queries

Give each search agent the brief, its parts with their synonyms, the platform and the urls already on the board. Per part, run three to five short queries built from the synonyms, plus one image search. A long query can return nothing while its short parts match.

Brief part Queries
Top bar "top bar", "app header", "toolbar", "navigation bar"
Empty state "empty state", "blank slate", "zero state", "no results"
Row density "dense table", "compact list", "data grid"

Candidates

Each agent writes one file, .gather/<brief>/candidates/<source>.json, holding an array:

[
  {
    "title": "Linear",
    "source": "Mobbin",
    "url": "https://mobbin.com/screens/4c03063e",
    "media": "https://bytescale.mobbin.com/…/4c03063e.webp",
    "kind": "image",
    "group": "Grouping",
    "note": "Take the grouped list, not the colours.",
    "details": { "Kind": "Shipped" }
  }
]
  • url is the original and identifies the item, so each one is unique.
  • media is a full URL, an absolute local path or a path inside .gather/<brief>/. The collector downloads or copies it into img/.
  • kind is image or video. Without it, the file extension decides.
  • poster is optional, a still for a video.
  • source puts › between a library and the board inside it, such as Pond › Are.na.
  • saved is true for an item from the user's own saves on any source. Saved items lead each group.
  • details adds rows to the viewer, such as { "Creator": "Kevin", "Style": ["Minimal", "Clean"] }.
  • A detail value of { "text": "Kevin", "url": "https://dribbble.com/kevin" } shows as a link.

Board

.gather/<brief>/board.json sets what the candidates cannot:

{
  "title": "Settings Page",
  "brief": "Desktop settings page, looking for grouping and density. Not: card grids.",
  "groups": [{ "title": "Grouping", "description": "How long lists of options are split up." }],
  "cut": ["https://mobbin.com/screens/9f12ab07"]
}

The board title takes Title Case. Groups appear in this order, and any group named only by a candidate follows them. A url under cut stays off the board.

Run the collector from anywhere:

Command Does
node collect.mjs .gather/<brief> Builds the board once
node collect.mjs .gather/<brief> --serve Serves it, rebuilds on every change and keeps the board reloading
node collect.mjs .gather/<brief> --done Builds it a last time and stops the reloading

Downloads run 8 at a time and are cached under a hash of the media url, so a second run fetches only what is new. A download that fails is named in the output and left off the board.

What the user gave

Input Reach it with
Open browser tabs List the tabs and read each page
Screenshots in the thread Read the images
Design files The Figma MCP get_screenshot on the linked node
Live sites Open the page. For exact values, hand over to match

Live capture

When the best reference is a shipped product, open its page in the browser and screenshot the part the brief names. Use the screenshot's path as media, the live page as url and Live as source. Stop at a login and say so.

Pond

The user's own library of saves from X, Instagram, Pinterest, Are.na, Cosmos, Savee, Dribbble, Behance, Threads, Bluesky, Tumblr, Reddit, TikTok, YouTube, Vimeo, Unsplash, GitHub, Substack, the web and local files. The Pond choice downloads from https://pond.raphaelsalaja.com and connects under Settings, Model Context Protocol. Connecting an account needs the Pond browser extension, from Browser Extension in the library menu. Pond runs as a local MCP server over stdio:

{
  "mcpServers": {
    "pond": {
      "command": "/Applications/Pond.app/Contents/MacOS/Pond",
      "args": ["--mcp"]
    }
  }
}

Copy the exact command from Pond, under Settings, Model Context Protocol. The Search Saves switch under Access must be on.

Tool Use
search_library Search by meaning, such as { "query": "website design" }
search_library with source Only one platform: arena, cosmos, pinterest, savee, dribbble, behance, twitter and the rest of the list above
search_library with visual: { "kind": "saveId", "id" } Saves that look like one the user likes
search_library with media: ["video"] Motion references only
search_library with folderId Inside one of the user's folders, from list_folders
query_library Exact filters: keywords limited by keywordScopes, and scope by source, folder, tag or savedSince, with a sort and cursor paging
list_folders Find the user's own groupings before searching
read_save Caption, text in images, note, link and metadata, or one part with field: caption, ocr, transcript, notes or description
grep_library Exact words in query inside captions, transcripts, notes, titles and image text
open_save, apply_filter Show a result in the Pond window

compress_images replaces the originals in the library, so gathering never calls it.

For an image search, find the save closest to what the user gave, then pass its id as visual. Run it again from the best two or three results.

A save's files sit in <Library>.pond/items/<id>.info/, so media is the absolute path to the original file. The Local API also saves URLs with POST /enqueue, and created: false means the URL is already saved. Its token stays in Pond's settings and is never pasted into chat.

marked.dev

The user's saved interfaces and motion from X, Pinterest and the web, each credited to its maker. It also holds Explore, a public catalogue browsed by what moves: hover, scroll and drag. It runs as a remote MCP server with no API key. Copy the config from https://marked.dev/mcp, then approve the connection in the browser.

Tool Use
list_collections, get_collection The user's own collections, read before searching
search_refs The user's saved refs, with unsorted: true for refs in no collection
get_ref A saved ref's media, source and author
search_explore, get_explore_ref Public refs, with category: "motion" for what moves
search_refs or search_explore with filters categories (app, saas, software, motion), platforms (x, pinterest, other), tags and authors
save_ref Save a pick into a collection, only when the user asks

Search it by interaction when the brief is about motion, such as "hover card lift" or "scroll reveal". A ref's video is the media, with kind: "video".

Sites in the browser

Every source but Pond, marked.dev and Mobbin opens in the browser, or with computer use where the browser cannot reach it. Open only the sites the user picked.

With Pond connected, read the user's saves on Are.na, Cosmos, Pinterest, Savee, Dribbble and Behance through Pond. Set source to arena, cosmos, pinterest, savee, dribbble or behance. Open their own pages for a channel Pond hasn't imported and for each site's search.

Site Image search
Cosmos Open an element and use its similar results. AI images hide on request
Pinterest A pin's More Ideas
Savee Search by colour or by an image

Read the user's existing channel and board names before creating one. Prefer sites picked by hand to ones that take paid submissions. Dribbble, Behance, Layers and CodePen are mostly concepts, so mark what comes from them.

Mobbin

Tool Use
Mobbin search_screens Single screens. Set platform to web for desktop and web work
Mobbin search_flows Multi-step flows, such as onboarding or checkout
Mobbin search_sections Sections of marketing sites
Mobbin exclude_screen_ids Leave out screens already shown when the user asks for more
Mobbin list_collections, get_collection_contents The user's own saved collections, searched first. id: "all-saved" reads everything saved, one content_type and platform per call
Mobbin create_collection, save_content Save picks into a new or existing collection, only when the user asks
Mobbin add_comment Note on a saved item why it was picked, only when the user asks

A Mobbin result's image_url is the media and its mobbin_url is the url. The image link expires after 30 days, which the download covers.

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 alerttoday3 checks · Risk MEDIUM
  • Gen Agent Trust Hubtoday

    The skill collects design inspiration but includes a local script ('collect.mjs') capable of copying and serving arbitrary local files, including sensitive system files, if malicious paths are provided via search results. It also performs downloads from external design websites.

  • Sockettoday

    No alerts

  • Snyktoday

    Risk: CRITICAL · 4 issues

Signed by skilld at b2e4283. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 12 hours ago.

Activeupdated yesterday

README badge

README badge for raphaelsalaja/skills/gather