agentleFS
garfiecgarfiec/Librechat-Mobile/.claude/skills/marketing-refresh/SKILL.md

marketing-refresh

Skill160 starsChanged 6 days agoApache-2.0View on GitHub

Refresh Switchboard's marketing images and copy end to end: the README screenshots and recordings, the F-Droid/IzzyOnDroid store screenshots in fastlane/, and the Google Play listing (captioned screenshots, feature graphic, listing text). Indexes what the app can do, ranks it by what a newcomer would feel, picks a running joke for a demo account, seeds that account with real model replies, drives the emulator and simulators through framed captures and recordings, composes the store images, and stops for the user's review before anything is committed. Use when the screenshots predate a visible feature or a theme change, before a store launch, or when asked to refresh the README images, the store listing or the Play graphics.

  • Reads credentials

The file

Show its rules and text

We split this skill into the rules it gives an agent

22 rules. About 0 are common, 14 are distinctive, and 8 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈14

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Real app, real replies: Every message on screen is a real model reply from a real server, through the API the app uses.

    line 43
  2. Distinctive:

    Never ship a shot of something broken: If a feature renders wrong during capture, leave it out of the lineup, tell the user, and record the …

    line 47
  3. Distinctive:

    A demo account only: No real names, emails, servers or personal data on screen.

    line 53
  4. Distinctive:

    Platforms don't share an image: Never an Apple device beside an Android one (Apple's marketing guidelines).

    line 55
  5. Distinctive:

    Every server or device change made for capture is temporary: Keep a list as you go and undo all of it in Phase 10.

    line 67
  6. Distinctive:

    Look at the current images and the README Screenshots section.

    line 79
  7. Very specific:

    Every PR merged since the last refresh: gh pr list --state merged --base develop --search "merged:>=<date>" --limit 300 --json number,title.

    line 88
  8. Very specific:

    Read the bodies of the feat ones.

    line 90
  9. Distinctive:

    Is current when the refresh happens: Search for what is trending in AI and in food, internet culture, etc.

    line 133
  10. Distinctive:

    Ask the user for the persona's display name before the first seed: the server can't rename an account afterwards.

    line 147
  11. Very specific:

    Follow scripts/demo-seed/README.md.

    line 152
  12. Distinctive:

    Never print an API key. Pipe it from the user's secrets file into the server's env file.

    line 172
  13. Very specific:

    Every capture comes out framed.

    line 207
  14. Very specific:

    Ask for --raw copies of every store still.

    line 213
  15. Very specific:

    Delete old files that no longer have a slot, and renumber the prefixes.

    line 219
  16. Distinctive:

    Keep the title to one line: about 26 characters at the phone size.

    line 232
  17. Distinctive:

    Keep the subtitle under ten words, and don't repeat the title's words in it.

    line 233
  18. Distinctive:

    Use one spelling convention across captions and listing text.

    line 234
  19. Very specific:

    Write it for the Play build: check every claim about defaults against ChannelPolicy.

    line 245
  20. Distinctive:

    If the user may be away from the terminal, publish one private review page with the Artifact tool.

    line 254
  21. Very specific:

    Run the checks: frame.py --self-test, play.py --self-test, check-readme.py, scripts/check-fdroid-metadata.py, seed.py --self-test.

    line 267
  22. Distinctive:

    Commit in separate commits: seeder and scenario changes.

    line 269

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Marketing Refresh

Produces, from one demo account and one capture session:

OutputLives inWho reads it
Store screenshots (plain framed, half size)fastlane/metadata/android/en-US/images/{phone,tenInch}Screenshots/F-Droid, IzzyOnDroid, the README
Feature graphicfastlane/metadata/android/en-US/images/featureGraphic.pngF-Droid banner, uploaded to Play by hand
README recordings and README-only stillsdocs/screenshots/the README
README Screenshots section, Why and More features copyREADME.mdGitHub
Play screenshots (captioned canvases)built by scripts/screenshots/play.py set <dir>, never committedPlay Console, by hand
Play listing texthanded to the userPlay Console, by hand

The tools already exist: scripts/demo-seed/ (the account) and scripts/screenshots/ (frames, captures, recordings, Play canvases, contact sheets, README check). Read both READMEs before starting. This skill is the judgment around them: what to show, in what order, with what story, and how to get it captured without a person holding the device.

Three reference files go with it:

  • reference/capture-agent.md: the brief for the agent that drives the devices, the per-shot checklist, and every device gotcha found so far.
  • reference/store-rules.md: Play, F-Droid and Apple rules, sizes and limits.
  • reference/2026-10-refresh.md: the last refresh as a worked example, from the feature index to the captions, including what was cut and why. Read it before Phase 2; it is the precedent.

Invariants

  1. Real app, real replies. Every message on screen is a real model reply from a real server, through the API the app uses. Never write or edit a reply, never retouch a screenshot. The frame, the canvas and the caption are the only things added. A weak reply is regenerated (seed.py --only X --redo X), not fixed.
  2. Never ship a shot of something broken. If a feature renders wrong during capture, leave it out of the lineup, tell the user, and record the bug. (2026-10: inline math rendered as separate blocks, so there is no math shot.)
  3. Don't oversell table stakes. A store slot goes to something a newcomer would feel and the website or other clients don't give them. "Tables fit the screen" looked fine and was cut as overselling a basic feature.
  4. A demo account only. No real names, emails, servers or personal data on screen. The persona name is the user's choice. The status bar is in demo mode on every capture.
  5. Platforms don't share an image. Never an Apple device beside an Android one (Apple's marketing guidelines). The hero shows one platform in both looks. iOS frames from the Simulator's chrome are for the GitHub README only; App Store images use Apple's product bezels and are outside this skill.
  6. One copy of each image (fastlane/README.md). The README links straight into fastlane/; docs/screenshots/ holds only what the README shows and the store doesn't. Play canvases are generated from fastlane/ and never committed.
  7. The store reads the tag. F-Droid reads fastlane/ from the tagged release commit, so the refresh must merge before the release is cut. Play takes images and text only by hand.
  8. Store text is per channel. fastlane/ text describes the open-source build; Play text is written for the Play build's defaults (ChannelPolicy, see fastlane/CLAUDE.md). Never paste one into the other unchanged.
  9. Every server or device change made for capture is temporary. Keep a list as you go and undo all of it in Phase 10.
  10. The user approves before anything image-related is committed. Device work in this skill is the user's request for device work; it does not extend to filing bugs, pushing, or uploading.

Phase 0: Scope and baseline

git log -1 --format='%h %cs %s' -- fastlane/metadata/android/en-US/images/phoneScreenshots
scripts/screenshots/sheet.py /tmp/before.jpg fastlane/metadata/android/en-US/images/*Screenshots/*.png

Look at the current images and the README Screenshots section. Note what is visibly stale: an old theme, a removed screen, a clipped clock, a feature that has since changed. If the argument didn't say which outputs are in scope, ask once, with the stale list as the reason.

Phase 1: Index what the app can do

You can't rank features you haven't listed. Enumerate them from sources, not from memory:

  • README.md Why and More features.
  • Every PR merged since the last refresh: gh pr list --state merged --base develop --search "merged:>=<date>" --limit 300 --json number,title. Read the bodies of the feat ones; a title is a claim, the body says what a user sees.
  • Release notes since then (gh release list, the Highlights sections).
  • ls feature/, each feature/*/CLAUDE.md, and every Settings page: features hide in settings.
  • Channel differences: ChannelPolicy and the perChannel(…) defaults in AppPreferences.
  • What the official web app doesn't do. That is where the differentiators are.

Write one table:

FeatureWhat a user feels (one line, their words)New since last refreshPlatformsShows asDifferentiatorDemo content neededShips in the next release

"Shows as" is one of: a still, motion (a gesture or animation, so a README recording), or not visual (offline, speed, privacy: these belong in copy, not images). Mark anything that needs server configuration (search, image generation, tool approval, web search) or isn't merged yet. A feature that won't be in the build the listing describes doesn't go in the lineup.

Phase 2: Rank and lay out the lineup

Rank by, in order:

  1. Felt benefit to someone who has never heard of LibreChat: how it looks, finding things, working offline, images, help from agents.
  2. Difference from the website and other clients.
  3. Reads in one still. Store slots are stills; gestures go to README recordings.
  4. New since the last refresh, since that's what the current images are missing.
  5. Not table stakes (invariant 3).

Then lay out three lists:

  • README: a hero (one platform, both looks, the same chat) plus about 12 rows. Each row has a benefit title, a phone cell, a tablet cell, its look, the seeded chat it uses, and whether each cell is a still or a recording.
  • Phone store slots: 8, by impact. Slot 1 is the hero. Play shows about three before the viewer scrolls, so the first three carry the listing.
  • Tablet store slots: 8 on the unfolded Fold.

Vary the look across rows (accent, light or dark, Material or Liquid Glass) so the set itself shows the theming. Present the three lists to the user and get approval before seeding; the lineup decides which chats exist.

Phase 3: Pick the story

The demo account needs a running joke that ties the chats together and survives a skim. Propose two or three themes, one line each, and let the user pick. A good theme:

  • Is current when the refresh happens. Search for what is trending in AI and in food, internet culture, etc. that month. The 2026-10 theme was "an AI agent with a Dubai chocolate habit": agents doing chores crossed with a viral dessert. Rejected: pizza (dated), the World Cup (ages fast).
  • Gives search a hook. One word ("chocolate") recurs in titles, replies and reasoning across several chats, so a search for it lights up.
  • Has a recurring character for images, kept consistent across an album of 12 or more, so the media page fills (Mayor Whiskers, a cat who is the town's mayor).
  • Has a prompt two models disagree on, for the comparison row. Asked whether to walk or drive 50 m to a car wash, one model said drive and the other said walk.
  • Has filler chats with good titles, so the sidebar looks lived in.
  • Avoids real people, brands, news events that will age, and anything a store reviewer could read as harmful.

Ask the user for the persona's display name before the first seed: the server can't rename an account afterwards.

Phase 4: Seed the demo account

Follow scripts/demo-seed/README.md. Edit scenario.toml so every lineup row that needs content has a chat, with requires naming what the server must offer. Then:

scripts/demo-seed/seed.py --self-test && scripts/demo-seed/smoke.py
scripts/demo-seed/seed.py --env-file scripts/demo-seed/.state/<server>.env --profile scripts/demo-seed/profile.<server>.toml
scripts/demo-seed/review.py …        # read every reply before capturing

Lessons from the last run:

  • React artifacts may import what web's Sandpack offers (lucide-react, recharts, three, date-fns, shadcn/ui at /components/ui/*, ...), all bundled with the app. Anything else renders as a missing-package error, and dat.gui or the shadcn Drawer (vaul) stop at a consent prompt first, which a screenshot shouldn't show.
  • Seed artifact chats on the strongest model; a smaller one wrote a malformed :::artifact tag.
  • A web search chat needs a search provider on the server, or the model apologizes. The seeder skips it unless /api/config reports webSearch.
  • Image chats need a tool that GET /api/agents/tools reports authenticated: true.
  • Photo assets are generated once (the OpenAI image API, with the same key) and the avatar is a crop of an album image. They stay out of the repo (assets/ is gitignored).
  • Never print an API key. Pipe it from the user's secrets file into the server's env file.
  • The agent panels (clarifying questions, tool approval) can't be seeded. They're sent live while capturing, and tool approval needs a server policy that is switched on only after seeding.

Phase 5: Server and device prep

Server. Make the changes the lineup needs, and write each one down for Phase 10:

  • SEARCH=true (Meilisearch running, master key set), or the search screen shows a "server search is off" note.
  • The image tool's key (for example IMAGE_GEN_OAI_API_KEY).
  • The tool approval policy in librechat.yaml (keep a backup of the file), only after seeding.
  • Restart the container after each change and confirm it through the API, not by assumption.

Android (Pixel 10 Pro Fold emulator): install the foss debug build, clear its data if a newer database is installed, sign in as the persona (the emulator reaches the host at 10.0.2.2), turn on background prefetch so every chat is cached, and launch by component. Details are in reference/capture-agent.md.

iOS (only if the README has iOS rows): build Debug for the newest simulators that have a runtime, install, and sign in. Maestro drives the login. Shots are reached by tapping; iOS deep links don't navigate.

Phase 6: Capture

Hand the capture to a background agent, in batches, with the brief in reference/capture-agent.md. Driving the device takes hundreds of tool calls: the 2026-10 run needed three batches because the agent hit its turn limit. Keep each batch to one look or one posture where possible, since theme switches and fold changes are the slow part.

While a batch runs, don't touch the device yourself. When it reports:

  1. Build a contact sheet of the new files (scripts/screenshots/sheet.py, with --frames 6 for GIFs) and look at every tile. Check: demo clock, the right theme for the row, no keyboard, no toast or error banner, no "isn't responding" dialog, only the persona's name, nothing broken.
  2. Send the retakes back to the same agent (SendMessage), with what was wrong.

Every capture comes out framed. Android uses the Pixel 10 Pro Fold skin that Android Studio's "screenshot with frame" uses (~/Library/Android/sdk/skins/pixel_10_pro_fold/). Recordings get the same skin on every frame (ffmpeg overlays the screen video between the skin's back and its corner mask), and come out as GIFs with transparent corners. iOS uses the Simulator's own device chrome. scripts/screenshots/frame.py does both.

Ask for --raw copies of every store still. They let a canvas be rebuilt at full resolution, though half-size frames upscaled to a Play canvas were indistinguishable in testing.

Phase 7: Compose

  • Store screenshots: copy the framed shots into fastlane/ in slot order, replacing in place. Delete old files that no longer have a slot, and renumber the prefixes. The hero is two framed shots joined by play.py pair material.png glass.png 01-hero.png.
  • README: a full-width hero (the tablet pair), then a table of Feature | Phone (width="280") | Tablet / Foldable (width="380"). A 2×2 grid in one cell uses width="186". Recordings and README-only stills go in docs/screenshots/. Every image gets alt text. Run scripts/screenshots/check-readme.py.
  • Play: edit the captions in scripts/screenshots/play-set.toml, then scripts/screenshots/play.py set <dir>. It frames, captions and checks every image against Play's rules, and copies the feature graphic in use (fastlane/'s featureGraphic.png, the same on Play and F-Droid). play.py feature draws the wordmark alternative in design/.

Captions are marketing copy:

  • Name the benefit, not the feature.
  • Keep the title to one line: about 26 characters at the phone size.
  • Keep the subtitle under ten words, and don't repeat the title's words in it.
  • Use one spelling convention across captions and listing text.

Phase 8: Copy

  • README Why and More features: lead with what a user feels (looks, speed, finding things, offline), not with implementation detail. Why sells the reasons to install, in two groups: what a native app gives you, and what it does that the web app doesn't (each claim checked against the web client at the pinned tag). More features lists only what Why doesn't: never the same item in both. Borrow the captions' benefit names for Why titles so the sections read as one story. Leave out what the web app does just as well, however showy (image generation, steering).
  • Play listing: app name (30 characters), short description (80), full description (4000). Write it for the Play build: check every claim about defaults against ChannelPolicy.
  • F-Droid text (fastlane/…/*.txt) is a separate copy. If it changes, run scripts/check-fdroid-metadata.py.

Phase 9: Review

Nothing image-related is committed until the user has seen it. Show contact sheets for the store screenshots, the Play canvases and the README recordings.

If the user may be away from the terminal, publish one private review page with the Artifact tool. Put the decisions you need at the top, then:

  • the Play screenshots as sideways-scrolling strips in listing order, tap to enlarge;
  • the listing text with live character counts and copy buttons;
  • the README section as GitHub will lay it out, at its real widths, with the recordings playing;
  • the F-Droid order;
  • notes on what was left out and the bugs found.

Ship the images as supporting files (downsized JPEG and WebP), not data URIs, so the page stays small.

Phase 10: Ship

  1. Run the checks: frame.py --self-test, play.py --self-test, check-readme.py, scripts/check-fdroid-metadata.py, seed.py --self-test.
  2. Commit in separate commits: seeder and scenario changes; capture-script changes; README copy; images plus the README table; Play tooling and captions.
  3. Undo the Phase 5 list: the server policy, the backup files, demo mode, and any app settings changed for a shot.
  4. List the bugs found during capture for the user. File them only if asked.
  5. Open one PR with /create-pr. Say which devices and OS versions were used, what wasn't captured and why, and that the images must land before the next release is cut.
  6. Give the user the Play set: the play.py set output directory and the listing text, ready to paste. The directory is generated, so say how to rebuild it.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.