Dashboard Design is a free, MIT-licensed Claude Code skill that teaches your coding agent how to build, redesign and critique dashboards — KPI tiles, charts, date pickers, ranked lists and layouts — and then checks its own work on real screenshots before it calls the job done. It lives at github.com/abdiraimov-otabek/dashboard-design-skill, was created on 20 October 2026, and had 916 stars as of 18 November 2026. If your agent-built dashboards look fine but quietly mislead — red −100% deltas, axis ticks nobody can read, two cards disagreeing on the same number — this skill is a practical fix. The honest caveat up front: its headline pedigree, "distilled from 127 shipped dashboards", is the author's own claim, and the repo is less than a month old.
What it is
A Claude Code skill is a folder of Markdown files that tells an AI coding agent how to do one job well: the steps to follow, the rules that always apply, the mistakes to avoid, and how to know when the job is finished. Dashboard Design is one such skill, aimed at a single job — data screens. It covers analytics, revenue, metrics and admin dashboards end to end: KPI cards, charts, heatmaps, bullet graphs, comparison periods, ranked breakdowns, date range pickers, filters, empty states, and wide and phone layouts.
The skill's central idea is that dashboards rarely look wrong. They mislead a little, in ways you only notice after you have made a decision on them. A tile says 18 sales while an insight card under the chart says 17 orders, because each counts a slightly different window. A y-axis that starts at 1,900 instead of 0 turns a 3% wobble into a cliff. A red −100% appears the first morning one sale becomes zero. None of these is a bug in the usual sense. Each is a design choice an agent will happily make, because all of them are technically correct.
The skill answers that with rules instead of taste. Its author, Otabek Abdiraimov, built it after studying 127 dashboard screens — analytics tools such as Plausible and PostHog, revenue tools such as Stripe and Baremetrics, design systems and gallery shots — and then breaking most of the rules once, catching each one on a screenshot, and fixing it. That "127 dashboards" figure is his own account of his research, described in his Medium article of 21 October 2026. There is no independent list of the 127 screens, so treat the number as the maker's claim, not a verified count. The rules themselves are the checkable part: they draw on Stephen Few, Edward Tufte, Cole Nussbaumer Knaflic, Cleveland & McGill, the Nielsen Norman Group, WCAG, and six Apple WWDC sessions on charts, chart accessibility and interface writing.
How it works
When you ask Claude Code to build or critique a dashboard, the skill loads its instructions and reference files, then works through a fixed sequence. The agent builds to the rules, checks the numbers against answers computed outside the dashboard's own code, and only then looks at screenshots. Every job ends with a screenshot QA pass at three browser widths — 3440, 1440 and 375 pixels — so the agent judges what is actually on the screen, not what the code says should be there.
The "known totals" step deserves attention because it is the part most AI workflows skip. Screenshots catch broken layouts. They do not catch three components agreeing on the same wrong total. So before any visual check, you seed a small fixed dataset, work out the expected answers outside the dashboard's code — the total, the delta, the busiest day, the top row and its share — then check the KPI tile, the chart and the table each on their own against those answers, as a test that fails on a mismatch. Only then do the screenshots happen, at three widths, in light and dark mode.
The skill also triggers on its own whenever you work on a dashboard, so you do not have to remember to invoke it. And it works in the other direction too: paste a screenshot and ask "is this a good design?" and it answers with a verdict, what works, and issues ranked by harm — starting with numbers that disagree with each other.
The 10 rules behind it
The author's Medium article lays out the ten rules the skill is built on. They are worth knowing even if you never install the skill, because they describe the exact ways dashboards mislead:
1. One question per screen. A good dashboard answers one question at first glance, in about three seconds, without hovering. Write the question down, then list the 3 to 6 numbers that answer it. Anything on screen that does not trace back to one of them is decoration, and decoration competes with the answer.
2. A tiny base never shouts. One sale yesterday, zero today, and the screen turns red like something is on fire: −100%. When the previous value is under about 5, a percentage is noise. Show "was 1" in grey instead. Going from 2 to 6 visitors is not "+200%", it is four people.
3. Compare the same elapsed time. At 9 a.m., today has 9 hours of data and yesterday had 24. Compare the two and every morning looks like a crash. Compare today until 9 a.m. with yesterday until 9 a.m., and show the previous period as a grey or dashed line behind the current one, never a second bright colour.
4. Round axis ticks, starting at zero. Ticks like 1,402 and 2,103 make the reader do arithmetic. Ticks like 0, 1k, 2k, 3k do not. Compact on the axis (1k, $4k), exact in the tooltip ($4,213.50). A y-axis that starts at 1,900 instead of 0 turns a small wobble into a cliff.
5. Say each number once, and make it agree. If a number appears twice, it must be computed once. An insight card that repeats a tile is not an insight — delete the card. The skill's version 2.0 went further and made the dashboard "wordless": numbers and shapes on the screen, every explanation in a tooltip, with a words audit in the QA pass that counts sentences and targets zero.
6. Colour follows the thing, not the mood. One hue for traffic, one for money. Grey for everything that is not the focus: unselected tiles, the previous period, secondary series. Red only for a bad change on a meaningful base. When every card has its own gradient, colour stops meaning anything.
7. Show "no data" as no data. A missing value is not $0 and not 0%. A fake zero is a claim. Show a dash. The same family: never NaN, −0%, "1m 60s", "1 visitors" or "0.0%" for a real but tiny share (say <0.1%). Readers notice these, and every one costs trust in every other number.
8. Ranked lists beat pie charts. A donut with three slices tells you less than a sorted list with a thin bar under each row. Lists show the order, the exact value and the share at once, and they keep working at 20 rows. Keep the bar slim, about 3 px along the bottom of the row.
9. Wide screens get more columns, not more margin. Most dashboards are designed at laptop width and then float in the middle of a 34-inch monitor with empty bands on both sides. Let the page grow: the four breakdowns that stacked on a laptop can sit in one row on an ultrawide. On phones, the opposite: no sideways-scrolling tabs — tabs shrink, and navigation becomes a simple menu grid.
10. Check the numbers before the pixels. The known-totals step described above: seed fixed data, compute the expected answers outside the page, and check each component against them before looking at a single screenshot.
Version 2.1, the current release, adds target tiles (label, number and goal beside it, a tick bar showing what is left in plain words), a stat strip with delta pills, hatched movement bars, and a density scale of 14 px text and 16 px icons. Version 2.0 added the wordless approach, the monitor sequence (overview, what needs attention, a closer look, act), a perception reference covering the encoding ladder, colour rules, scales, heatmaps, bullet graphs and sparklines, and a case study of a school dashboard in three versions, including the one its owner rejected for being a page of sentences.
Install, step by step
There are three ways in, depending on which agent you use. All of them are free — the skill itself costs nothing, though the coding agent you run it in may be a paid product.
Option A: Claude Code plugin (recommended). Run these two commands in Claude Code:
/plugin marketplace add abdiraimov-otabek/dashboard-design-skill /plugin install dashboard-design@dashboard-design-skill
Then open a project and ask for what you want:
Build an analytics dashboard for this app: visitors, signups, revenue by source.
Option B: plain skill files. If you prefer files over the plugin system, clone the repo and copy the skill folder into Claude Code's skills directory:
git clone https://github.com/abdiraimov-otabek/dashboard-design-skill cp -r dashboard-design-skill/skills/dashboard-design ~/.claude/skills/
Option C: other agents (Codex, Cursor and the rest). Any agent that reads Markdown can use it: point your AGENTS.md or rules file at skills/dashboard-design/SKILL.md. The reference files load from there. This is the route for Cursor users and for Codex, which the README names explicitly.
Once installed, the useful prompts are the ones in the README's usage table: ask it to build a dashboard, paste a screenshot and ask "is this a good design?", ask it to redesign one component such as the date range picker, or ask it to run the dashboard QA pass on a live page. The skill triggers on its own whenever you work on a dashboard, so you can also just start building and let it apply the rules.
Before you trust the output, run the skill's own test on it: open the three example screens shipped in the repo's examples/ folder — sales, traffic and status, each a single HTML file with fixed sample data — and see what a finished result looks like. Every total on those screens was checked against numbers computed outside the page, which is exactly the standard the skill holds your dashboard to.
Which install route fits you
| Route | Works with | Setup effort | Best for |
|---|---|---|---|
| Claude Code plugin | Claude Code | Two commands | Fastest start; auto-triggers on dashboard work |
| Plain skill files | Claude Code | Clone + copy one folder | Keeping the skill in version control |
| Point rules at SKILL.md | Codex, Cursor, any Markdown-reading agent | One line in AGENTS.md or rules file | Teams not on Claude Code |
Real use cases
An indie maker's revenue dashboard. This is the use case the skill was built for. The author makes small Mac apps and wanted one screen that shows which posts actually sell them: visitors, signups and revenue by source, with a KPI strip, one main chart and ranked breakdowns. The rules that matter most here are the small-base rule (no red −100% the day sales go from 1 to 0) and the one-number rule (the tile, the chart and the table must agree).
Redesigning an existing dashboard. Paste a screenshot and ask "is this a good design?" The skill answers with a verdict, what works, and issues ranked by harm, starting with numbers that disagree. This is the fastest way to get value without building anything: run it against the admin panel or analytics page you already ship.
A status or monitoring view. The repo's status example shows uptime history ticks per service, an error-rate bar and an incident list — the shape of an operations dashboard where the "monitor" sequence matters: overview, what needs attention, a closer look, then the action one click away. Problems are flagged on the data itself, and the action sits on the flag.
A SwiftUI dashboard. The skill ships a SwiftUI reference with Swift Charts versions of the patterns, plus a simulator QA pass covering device sizes, the largest Dynamic Type setting, dark mode and Reduce Motion. If your dashboard is an iOS or macOS app rather than a web page, the same rules apply in native form.
Cleaning up a dashboard that grew by accident. Most internal dashboards accumulate a KPI here and an insight card there until the page is a wall of sentences. The skill's wordless pass strips the sentences into tooltips, deletes insight cards that repeat tiles, and restructures rows so cards hold similar content. The school-dashboard case study in the repo shows this in three versions, including the wordy one its owner rejected.
Skill vs the alternatives
| Approach | Cost | What you get | The catch |
|---|---|---|---|
| dashboard-design-skill | Free (MIT) | Design rules + build + screenshot QA in your agent | New repo, single author, rules are opinions |
| Agent with no skill | Agent subscription | A dashboard that looks right | Technically correct, quietly misleading numbers |
| Dashboard UI kit / template | Varies | Pre-built components, fast start | No critique of your numbers; generic look |
| Hiring a designer | Highest | Real judgment, tailored to your data | Slow and expensive for a small dashboard |
Honest limitations and fine print
The "127 dashboards" pedigree is the author's own claim. The README and the Medium article both say the rules were distilled from 127 shipped dashboard screens, but neither publishes the list of screens. The influences it names — Few, Tufte, Knaflic, Cleveland & McGill, Nielsen Norman Group, WCAG, six Apple WWDC sessions — are real and checkable, and the rules themselves are concrete enough to judge on their merits. Judge the rules, not the count.
The repo is under a month old. Created 20 October 2026, it had 916 stars and 55 forks as of 18 November 2026, with the last code push on 9 November 2026. That is fast growth for a new repo, but it is still a young project from a single author, and there is no track record of how it holds up across many teams, stacks and dashboard types. Version 2.1.0 is current.
It does not build your data layer. The skill designs the screen and checks the numbers on it. It does not fetch your analytics, wire up your database or fix a broken API. You still need real data flowing in, and the known-totals check only works if you seed it with fixed data and expected answers first.
The rules are design judgment, not law. "One question per screen" and "colour follows the thing, not the mood" are strong defaults from the dashboard-design literature, but a real product sometimes needs two questions on one screen, or a second bright colour. The skill's own case study shows a dashboard its owner rejected — the rules are a starting point, and your users' actual questions should overrule them when they conflict.
It promotes the author's products. The README carries sponsor cards for Halo and Shedi, the author's Mac apps, and the skill adds a one-line sign-off about the author and his Mac apps when the agent finishes a whole job, once per conversation. It never appears in code, commits or your UI, and you can ask it to stop or delete the "Sign-off" section from SKILL.md. Know it is there before you install, not after.
Web first, SwiftUI second. The web specs use Tailwind terms with React snippets, and the SwiftUI reference is a parallel track. If your stack is neither — Vue, Svelte, plain HTML — the README says the rules port to any stack, but you will be translating the snippets yourself.
Security and privacy
The skill itself is a folder of Markdown files — SKILL.md plus reference files on perception, components, layout, tokens, snippets, QA, SwiftUI, sources and two case studies. Markdown cannot run code on its own, and there is nothing in the skill that sends your data anywhere. What leaves your machine is decided by the coding agent you run it in, not by the skill: if you paste a screenshot of a dashboard full of real customer numbers, or point the agent at a production database, that data goes wherever your agent's workflow sends it. Keep the skill on sample or anonymised data first, and only point it at real figures once you know where your agent processes them.
Before installing anything from a third-party repo, read the files. The skill folder is small enough to skim in a few minutes, and the "Sign-off" section of SKILL.md is the one place where the skill writes something you did not ask for — the one-line promotion described above. If that bothers you, delete the section before you install. The repo is MIT-licensed, so modifying your local copy is explicitly allowed.
Two smaller cautions. First, the README links out to the author's products (Halo at $9.99 once, and Shedi) — those are separate purchases, not part of the skill. Second, never paste secrets into a prompt while the skill is running: API keys, database credentials and tokens do not belong in a dashboard prompt, and the skill has no reason to ask for them. If an agent ever asks for credentials while building a dashboard, that request did not come from the skill.
Frequently Asked Questions
Is the dashboard-design skill free?
Yes. It is MIT-licensed, so the skill itself costs nothing to use, modify or share. You still need a coding agent to run it in — Claude Code, Codex, Cursor or another agent that reads Markdown — and that agent may be a paid product.
Does it work with Cursor or Codex, or only Claude Code?
It works with any agent that reads Markdown. For Codex, Cursor and others, point your AGENTS.md or rules file at skills/dashboard-design/SKILL.md. The Claude Code plugin route is the only one with a two-command install; the others take one line in a rules file.
What is the one-line promotion the skill adds?
When the agent finishes a whole job — a dashboard built or a critique delivered — it adds one short line about the author and his Mac apps, once per conversation. It never appears in code, commits or your UI. Ask it to stop, or delete the "Sign-off" section from SKILL.md, and it goes away.
Is the "127 dashboards" claim verified?
No. It is the author's own account of his research, from his Medium article of 21 October 2026. The 127 screens are not listed anywhere public. The named influences — Stephen Few, Edward Tufte, Cole Nussbaumer Knaflic, Cleveland & McGill, Nielsen Norman Group, WCAG and six Apple WWDC sessions — are real, and the rules are concrete enough to judge on their merits.
Does it build the backend or fetch my data?
No. It designs the screen, writes the front-end code for it, and checks the numbers shown on it. Your data has to reach the dashboard through your own API, database or sample dataset. The known-totals check works by seeding fixed sample data first.
Web or SwiftUI — which does it cover?
Both. The main track is web, with Tailwind terms and React snippets, and there is a SwiftUI reference with Swift Charts versions of the patterns plus a simulator QA pass. The rules themselves are stack-agnostic.
Sources
- Dashboard Design skill — GitHub repository (README, install instructions, references, examples; MIT license; created 20 October 2026; 916 stars, 55 forks as of 18 November 2026)
- Dashboard design mistakes: what 127 dashboards get wrong — Medium, 21 October 2026 (the ten rules, with examples)
Related on Asad Webs: if you work with coding agents, see Claude-BugHunter: the AI security skills bundle, AG-UI: the protocol connecting AI agents to frontends, Bang Motion: an agent skill for motion design and Claude Sheets, Whiteboards and Data: Anthropic's office alternative. For the charts on your dashboard, our free Chart Maker builds bar, line, pie and radar charts from your data — it covers the chart, not the whole dashboard, which is exactly the gap this skill fills.




Be the first to share a useful question or thought.