agentleFS
roselis-labroselis-lab/keel/.agents/skills/keel-design-system/SKILL.md

keel-design

Skill2 starsChanged 4 days agoApache-2.0View on GitHub

Use this skill to generate well-branded interfaces and assets for Keel, either for production or throwaway prototypes/mocks/etc. Contains essential design guidelines, colors, type, fonts, assets, and UI kit components for protoyping.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the README.md file within this skill, and explore the other available files.

    line 7
  2. Distinctive:

    If creating visual artifacts (slides, mocks, throwaway prototypes, etc), copy assets out and create static HTML files for the user to view.

    line 8
  3. Distinctive:

    If the user invokes this skill without any other guidance, ask them what they want to build or design, ask some questions, and act as an …

    line 9
  4. Distinctive:

    Hairlines, not shadows: Only the toast and the saved-dialog cast one.

    line 24
  5. Distinctive:

    Review-first: Every entity reads before it edits, and both wear the same section band.

    line 26

≈ “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

Read the README.md file within this skill, and explore the other available files. If creating visual artifacts (slides, mocks, throwaway prototypes, etc), copy assets out and create static HTML files for the user to view. If working on production code, you can copy assets and read the rules here to become an expert in designing with this brand. If the user invokes this skill without any other guidance, ask them what they want to build or design, ask some questions, and act as an expert designer who outputs HTML artifacts or production code, depending on the need.

Two front ends now

The reading pages (Overview, Reports, Compliance, Threats, Controls, Coverage) are a React app in web/, styled "Signal in Graphite": its rules are in web/DESIGN.md, and for those pages they replace this kit. Threats, controls, reports and the style guide are all edited in the app too (/threats/:id/edit, /controls/:id/edit, /reports/:system/:date/edit, /style-guide/:entity/:field); old /edit/… links redirect there, and nothing in the app links back. Everything below describes the single file keel/static/index.html, still served at bare /edit for anyone who opens it by hand; it is no longer where anything is edited.

Quick orientation

Keel is a framework for building and running your own living GenAI/LLM threat model. Its one UI product is a six-screen authoring app (Overview, Threats, Mitigations, Style guide, Coverage, Reports) served as a single static HTML file.

Non-negotiables, all documented in readme.md:

  • Two hues. Navy is structure and text; crimson is the only accent and means exactly selected / primary action / danger. Amber advises, green passes. Never decorative colour.
  • No webfonts. OS UI stack for prose, OS mono for anything that is an identifier. Mono is a semantic, not a style.
  • No icon set, no emoji, no SVG icons. A handful of Unicode glyphs (« » ▾ ▸ + × ⓘ ·), plus letters-in-tiles and CSS dots. The full inventory is in readme.md → Iconography.
  • Hairlines, not shadows. Only the toast and the saved-dialog cast one. No blur, no backdrop-filter, no translucency.
  • Almost no motion. .12s hover on background and border-color, one .2s toast rise. No custom easing, no press states.
  • Review-first. Every entity reads before it edits, and both wear the same section band. Empty fields are omitted and re-surfaced as gap chips. Red errors block a save; amber advice never does.
  • Sentence case everywhere. Buttons are verbs. Toasts are past tense. Ids and enum values are quoted verbatim and lowercase.
  • No logo exists. Render "Keel" as plain type wherever a mark would go. Never draw one.

Where the product has moved past this system

The kit was extracted from keel/static/index.html and describes it as it stood. Five things have since changed in the product; when they disagree, index.html wins and the note below says why.

  • Section headings are a real tier, not the eyebrow. readme.md says every section wears an 11px uppercase eyebrow. That put section headings a step below the 14px prose they introduce, so the screens read as one flat block of text. Section headings (h3.slabel, .detail fieldset > legend.slabel) are now 15px / 600, sentence case, navy-900. The eyebrow survives only where it labels a single field or group: facet keys, sub-labels inside a repeatable card, the rail list label, the finding chain's <dt>.
  • Severity has three bands, not four. critical is gone from keel/schemas/report.py and from the tokens. An organisation's risk policy is written against three bands, and a fourth level only ever meant "high, but really". high now takes the solid crimson fill that critical had, and it is the only solid fill on a finding, likelihood is graded on the same ramp but drawn as an outline (.sev.quiet), so one card never shows two solid crimson blocks.
  • Primary navigation is a top bar, not a strip in the rail. Five text tabs stretched across a 320px rail read as a cramped segmented control and pushed the rail's own header into a second squeezed row. The tabs also now survive collapsing the rail.
  • navy-400 is not a text colour. It measures 2.62:1 on white. It stays in the ramp for rules, dots and the low severity spine; text that used it moved to navy-500, which was itself darkened to #636e88 so the quiet tier clears 4.5:1 on all four Keel surfaces. --green-700 was added as green's text-only step, mirroring --amber-700. Badges are 11px, not 10.
  • The Reports screen has no "assessor dialogue" section. The kit describes one; it is now meta, "How this assessment ran", carrying what the agent asked, what the specialist volunteered unprompted, where it was corrected, and how long the run took. Folded at the foot beside "Ruled out", and never exported.
  • reachability no longer opens with "NOT applicable if". readme.md says those three words always start the value. They were the field's own label written into every value, which pushed the actual condition to the end of the line and read as a heading baked into the data. The field now carries the plain condition ("The model cannot influence the arguments…"), and the UI supplies the framing.
  • The Overview has a hard-error band, and colour in it is graded to importance, not to the subject. Full width, above both columns, absent entirely when clean. A record in it failed its schema at load and is not being served, so the catalog on disk and the catalog in the app have diverged, but "this is serious" is worth exactly one signal, not a red screen. The band is a normal white card with a 2px crimson left spine (the severity-spine idiom already in the kit); its heading wears the same 10px uppercase navy-600 eyebrow every other block wears; the file path is mono navy-500 like every other identifier; the message is navy-800 prose. The only other crimson is an 11px not loaded marker on the rows the app is actually missing. A crimson wash over the whole band put path, message and marker at one loudness and shouted before the reader knew what was wrong. Each row is a metadata line (file · field · marker) then the message, laid out as one wrapping row, a long message orphaned the marker onto a third line and one defect read as two. Nothing here clicks into the app, because the entry is not in the app to open; the file links to the repo editor when one is configured.
  • Links are navy with an underline; crimson is not a link colour. --text-link used to be crimson-600, so every entity name on every screen was red. Crimson is the only accent and it means selected, primary action, or danger, spending it on ordinary navigation left the Overview looking alarming and left nothing louder for the things that actually are. Links now use --text-link (navy-800) with a --line-link underline at rest, which also stops colour being the only thing marking them. Reserved crimson: the selected tab's underline, the selected rail row's wash, primary buttons, focus rings, error text, and the high-severity fill.
  • Views have real paths, there is no #. The API moved under /api, which freed /threats/T-X and the rest for the app, and the client router reads location.pathname. A single catch-all serves the app for anything the API did not claim, and that catch-all excludes /api explicitly: without the exclusion an unmatched API path answers 200 with a page of HTML, so a typo looks like it worked and a rejected path stops looking rejected.
  • Anything reported as broken is fixable where it is reported. A record that still loaded (one bad link, the rest intact) opens in its normal editor. A record that did not load has no entry and nothing that parses, so clicking its file name opens the raw YAML inline in the error band, and a save that would still not load is refused in place with the field and the reason, the same gate the loader applies. Errors are grouped by file: three bad fields in one record are one thing to go and fix.
  • Enum values read as English where you choose one, and as tokens where you read one. catalog/{harm,surface,source,components}.yaml carry a name and a one-line gloss per value; they existed and nothing loaded them, so every screen showed bare tokens. Selects and checkboxes in the editor now show "Downstream consumer, A system that executes or consumes the agent's output (CI/CD, DB, renderer)"; the read view keeps the verbatim lowercase token in its badge, because that is the value on disk and the badge is what you match against a YAML diff. validate_catalog requires those files and the schema's Literals to agree in both directions.
  • There is a sixth screen, Coverage, and nothing on it is crimson. It holds Keel's claim about the sources it tracks: a rail of sources with a covered/out-of-scope/gap tally bar, and a main pane of every entry in a pinned release, grouped where the source groups its own (an ATLAS tactic). The three states are graded by what they mean, not by how alarming they look, covered takes the pass green, gap takes the amber advice colour because it is work not yet done, and out_of_scope is navy, because a boundary is structure. It renders as "out of scope": the earlier word was "declined", which read as Keel disputing the source rather than describing its own edge. None of the three is an error, so the accent appears nowhere. The header states "N of M entries" whenever an import is unfinished; a matrix that shows twelve rows of a hundred-and-one-entry release reads as complete when it is not.
  • A card says who else names it, and says so when nobody does. The read view carries a "Corroborated by" section built from the coverage matrix's reverse index, never stored on the card, because the files are written source-first so that a gap has a row to be empty in. When no tracked source names an entry the section still renders, saying the entry is Keel's own. That is not an empty state to hide; to a sceptical reader it is worth as much as the corroboration.
  • A report has one verb, not three. Save, Correct and Reopen are gone. Editing is always allowed, and editing a final report drops it back to draft, the header swaps its final badge for the Finalize button the moment a save lands, so the badge never sits over a document that has changed since it was signed off. Three verbs asked the reader to classify their own edit ("is this a correction or a re-assessment?") before making it, which nobody wants to do before fixing a typo. Assessing the system again is a new report with today's date, started from the rail.
  • Prose has a measure. Body copy caps at var(--measure) (74ch). The editor pane is 1fr and runs to the full width of a large display.
  • The type ramp is referenced, not retyped. Every font-size and line-height in index.html is var(--fs-*) / var(--lh-*); there is not one raw value outside :root. A ramp only declared in :root is a suggestion, the app had 117 raw sizes against 6 token uses, and eight line heights against four tokens. Line heights are now exactly five: --lh-1, --lh-tight 1.35, --lh-base 1.55, --lh-prose 1.65. Form controls inherit type (button, input, select, textarea), because without that any control missing an explicit size falls through to the browser's 13.333px, which is not on the ramp and cannot be reached from it.

Where things are

  • styles.css, link this one file; it imports every token file.
  • tokens/semantic.css, author against these aliases (--text-body, --surface-section, --accent, --status-advice), not the raw ramps.
  • components/, 30 primitives in five groups (core, forms, structure, feedback, data), each with a .d.ts props contract and a .prompt.md usage note.
  • ui_kits/authoring-ui/, the five real screens, click-through. Read this before recreating any Keel view.
  • guidelines/, foundation specimen cards.
  • assets/, the three flat vector diagrams that exist. The app itself ships zero images.

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.