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.

SKILL.md

≈32 tokens for metadata: the name and description. ≈3.6k when used: this file. ≈3.1k more on demand in 2 files.

Description uses 1.6% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈127 characters, or 1.6%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

Gather

This skill collects inspiration before a design is built or reworked. It searches every source at once, fills one board as results land and links every reference back with the part to take.

Evidence for a choice, not a style to copy

A reference shows how one part could work. Taking a whole app's look when the user asked for its top bar is a finding. With no number given, gather at least 20 references from at least three sources.

Use it when the user asks for inspiration, references or "inspo", or says a screen looks bad and asks how to improve it. Building variants from the board belongs to prototype. Rebuilding one reference value for value belongs to match.

1. Name the brief

Write one line naming the surface, its platform and what to look for: layout, navigation, density, type, colour, motion or overall taste. A search with no target is gather-name-the-brief.

Split the brief into its parts, such as "top bar", "empty state" and "row density", and give each part two or three synonyms. Queries come from the parts, two or three words each. One long query carrying the whole brief is gather-split-the-brief.

Search on the same platform as the work. A desktop or web app takes web and desktop references. Mobile screens for a desktop app is gather-same-platform.

2. Start with what the user gave

Open the tabs, screenshots, links and design files already in the thread before searching anywhere. They seed the image searches in step 4. Searching Mobbin while the user's screenshots sit open in the browser is gather-start-with-given.

When the thread holds no reference, ask for them in the same round as the source question in step 3. Ask for 1–3 products that do the part well, one the user dislikes and why for each. Offer "None, search wide" so the question never blocks.

Read each answer for the part it does well, so "like Linear" becomes "Linear's dense grouped rows". Add each dislike to the brief after Not:, so every search agent avoids it. Searching with no reference and no question asked is gather-ask-for-references.

3. Ask which sources to use

Ask one multi-select question with every source in sources.md as its own option, marking each one connected or not. Search only the sources picked.

Where the question tool caps its options, split the list across questions in the order sources.md gives. A source missing from the question, a searched source left out or no question at all is gather-ask-which-sources.

Connect a picked source that is missing before searching it. How to connect each one is in sources.md. Wait for the answer before searching anywhere. Ask once per brief, so More Like This and "more" keep the same picks.

Pond imports the user's saves from Are.na, Cosmos, Pinterest, Savee, Dribbble and Behance, so with Pond connected read those saves through Pond's source filter. Calling a board's API, or asking to sign in only to read saves Pond holds, is gather-boards-through-pond.

A source the user left out stays out of Pond too. Pond's source filter cannot exclude, so drop every Pond result whose source was left out.

4. Search every picked source at once

Start one agent per picked source, or run their searches as parallel calls where agents are unavailable. Each writes candidates/<source>.json in the shape sources.md shows. Searching one source after another is gather-parallel-sources.

When Mobbin answers that monthly usage is high, stop calling it, tell the user and carry on with the other sources. Search Mobbin again only when the user asks for it by name. Another Mobbin call after that warning is gather-stop-at-usage-limit.

Every picked site the user can sign into gets two searches: their saves there and the site's own catalogue. Searching only one side while the other holds matches is gather-saves-and-site.

The user's own saves lead each group, from any source, each marked saved: true. A board with none of the user's saves while a picked source holds matches is gather-own-library-first.

Pull from at least three of the picked sources, or from every one when fewer are picked. A board drawn from one app is gather-go-wide.

Seed image searches from what the user gave and from their closest saves, through Pond visual and Cosmos. Text queries alone when the user shared screenshots is gather-search-by-image.

When the best reference is a shipped product, open its live page and screenshot the part. A gallery thumbnail of a site that is one click away is gather-capture-live.

Pull about twice the number asked, then cut. When the user asks for more, leave out every url already on the board. Repeats under "50 more" is gather-enough-to-choose.

5. Collect into one board

Run the collector once the first candidates land and open the board straight away. It copies the board page, downloads every image and video and fills the board live as agents finish.

node <skill>/collect.mjs .gather/<brief> --serve

.gather/ stays out of version control. Write the title, brief and group order to .gather/<brief>/board.json as sources.md shows. Links pasted in chat instead of a board is gather-one-board.

Gathering only reads the user's library. The collector copies each save into img/ and the originals stay as they are. Compressing or rewriting saves in place is gather-read-only-library.

The collector and the board ship with the skill, so a run writes data only. A hand-written board page, or a per-run script that downloads or builds, is gather-shipped-collector.

6. Look at every reference

Open each downloaded file and look at the part the brief names. A reference kept on its title, metadata or a preview is gather-look-at-every-image.

Cut only clear misses, by adding their url to cut in board.json. When two references sit close, both stay and the user decides. A close call settled by the agent is gather-cut-clear-misses.

Each reference gets one note naming the part to take, such as "the top bar". A reference with no note is gather-name-the-part. Each keeps its source link. An image with no link is gather-link-every-reference.

The same shot often lands from Pinterest, Pond and the maker's own page. Keep the copy closest to its maker, following a repin's link or an image search to find it. Cut the rest. Two tiles of one image, or a repin kept over the maker's shot, is gather-one-copy-per-shot.

Mark each reference Shipped or Concept under Kind in its details. A concept shot shown as a shipped product is gather-concept-or-shipped.

Name the creator under Creator, with their profile as a url when the source shows it.

Name each group in two or three words of Title Case, such as "Agent Activity", and put the detail in its description. A group named like a sentence is gather-short-group-names.

When the cut is done, run the collector once more with --done so the board stops reloading.

7. Act on what comes back

The board copies two prompts. Copy Picks returns the picks with the user's notes, which become the brief for the next step. More Like This returns one reference.

Sum the picks up as one line of what to take and one of what to avoid, then ask the user to confirm it. Picks handed on as a list of links is gather-sum-up-the-picks.

A lukewarm reply, such as "these are fine" or "nothing really pops", is not a pick. Ask which one or two come closest and what each lacks, naming the brief's parts. Building from a shrug is gather-lukewarm-is-not-a-pick.

When the user rejects the whole board, ask which reference came closest and which came furthest. Add what the furthest does to Not: and seed image searches from the closest. A second board on the same queries after "none of these" is gather-no-blind-retry.

For More Like This, seed an image search from that reference and keep to its group, with every url on the board left out. A text search on its title is gather-more-from-the-pick.

8. Save what the user keeps

Send picks where the user says: their library, a channel or a board. Read the destination's naming first and match it. A new channel named "FW 26 🔥" beside fashion / inspiration / 2026 / fall winter is gather-match-the-naming.

Check for each item before adding it. A second copy of a saved item is gather-no-duplicates. When a source needs a login or a CAPTCHA, stop and say which one. Working around a login is gather-ask-at-login.

The pair

Every check is a pair. The broken line sits beside the line that replaces it.

Rule Broken Replacement
gather-name-the-brief "Searching settings pages" "Desktop settings page, looking for grouping and density"
gather-split-the-brief query: "desktop settings page with grouped dense options" "settings groups", "preferences list", "dense settings"
gather-same-platform iOS screens for a desktop app platform: "web" and desktop apps
gather-start-with-given A Mobbin search while the user's tabs are open The open tabs read first
gather-ask-for-references A Mobbin search started on "settings page" alone "Which apps do settings well, and which do you dislike?", asked with the sources
gather-ask-which-sources Six sources offered while Savee and Refero go unlisted Every source in sources.md as its own option, marked connected or not
gather-boards-through-pond "Are.na's API blocks scripted requests, so I'll search it in the browser" while Pond is connected Pond searched with source: "arena", "cosmos" and "pinterest"
gather-parallel-sources Pond, then Mobbin, then Cosmos, one after another One agent per source, started together
gather-stop-at-usage-limit Mobbin searched again after "your monthly usage is high" "Mobbin says monthly usage is high, so I've stopped searching it"
gather-saves-and-site Mobbin searched while the user's Mobbin collections go unread get_collection_contents with id: "all-saved", then search_screens
gather-own-library-first A board of Mobbin screens while Pond holds matches The user's saves first in each group
gather-go-wide Twenty screens from one app The user's saves and three sites
gather-search-by-image Text queries while the user's screenshot sits in the thread search_library with visual seeded from it
gather-capture-live A Land-book thumbnail of linear.app A screenshot of the live top bar, linked to linear.app
gather-enough-to-choose "50 more" returns the first 20 again 50 new references, earlier urls left out
gather-one-board Twenty links pasted in chat .gather/settings/ served and opened, grouped by idea
gather-read-only-library Saves compressed in place for the board Copies in img/, originals untouched
gather-shipped-collector A build.mjs written for this run node collect.mjs .gather/settings --serve
gather-look-at-every-image A screen kept on its title or a preview The downloaded file opened and the top bar looked at
gather-cut-clear-misses Two close references, one cut by the agent Both on the board for the user
gather-name-the-part A screenshot with no note "Take the grouped list, not the colours"
gather-link-every-reference An image with no source The image linked to its original
gather-one-copy-per-shot One shot twice, as a Pinterest repin and a Dribbble post The Dribbble post kept, linked to its maker
gather-concept-or-shipped A Dribbble concept shown as a product details: { Kind: "Concept" }
gather-short-group-names "Activity: thinking, tool calls and steps" "Agent Activity"
gather-sum-up-the-picks Six links passed to prototype "Take: grouped rows and quiet colour. Avoid: cards. Right?"
gather-lukewarm-is-not-a-pick "These are fine" read as approval "Which two come closest, and what does each lack?"
gather-no-blind-retry "None of these" answered with 20 more from the same queries The closest seeding an image search, the furthest's trait under Not:
gather-more-from-the-pick More Like This searched as "Linear settings" An image search seeded from the picked screen
gather-match-the-naming A new channel named "FW 26 🔥" fashion / inspiration / 2026 / fall winter
gather-no-duplicates The same pin added twice Existing items checked and skipped
gather-ask-at-login A login worked around "Pinterest needs you to sign in"

Reporting

Severity. HIGH is a whole look copied when one part was asked for, items saved twice or under the wrong name, or saves rewritten in the user's library.

MEDIUM is a search with no brief, the wrong platform, references in the thread left unread or no references asked for. It is also a skipped source question, a source searched that the user left out or a board reached around Pond while Pond holds its saves.

MEDIUM also covers one source only, sources searched in turn or the user's own saves missing. So is a site searched without the user's saves there, or the reverse. So is a Mobbin call after its usage warning, or a login worked around.

Links pasted in chat, a board with no links and a hand-written board or build script are MEDIUM too. So is a reference kept unseen, a close call settled by the agent or a concept shown as shipped. So is a lukewarm reply built on, or a second board after "none of these" on the same queries.

LOW is one long query, text search where an image search fits or a gallery thumbnail of a live site. So is a reference with no note, a long group name, repeats under "more" or More Like This searched by title. So is one shot shown twice, a repin kept over its maker or picks handed on without a summary.

Format. One finding per root cause, citing the request when no file applies:

file:line - [rule-id] what is wrong
request 2 - [rule-id] what is wrong

Then a Rule | Count | Severity summary. Block when any HIGH remains, Approve otherwise. Mark every source you could not reach as Not verified.

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 warning2d3 checks · Risk SAFE
  • Gen Agent Trust Hub2d

    The skill is a design inspiration tool that collects references from various web sources and local libraries. It is generally functional but possesses an indirect prompt injection surface due to its ingestion of third-party content and includes a local file access capability that could be misused to expose sensitive files.

  • Socket11d

    No alerts

  • Snyk2d

    Risk: MEDIUM · 2 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