Skip to content

Repository files navigation

Web App Completeness

Validate MIT License

Four portable agent skills that make operational web applications complete on purpose—not by dumping a generic feature checklist into every screen.

Kurz auf Deutsch: Dieses Repository hilft Coding-Agenten, bei Verwaltungs- und Daten-Apps die normalen, oft vergessenen Produktfähigkeiten bewusst mitzudenken: Anlegen und Bearbeiten, Auffindbarkeit, vollständige Datenreichweite, Zustände, sichere Aktionen, Berechtigungen und mehr. Die ausführliche Dokumentation und die Skills bleiben als wartbare Quelle auf Englisch.

The problem

AI-generated admin and SaaS interfaces often render the happy path while omitting the surrounding workflow. A resource list may show twenty rows but make the rest unreachable. A form may create records but never edit them. An error may look identical to an empty dataset.

Web App Completeness turns those omissions into explicit product decisions. It asks what is applicable, preserves valid project patterns, and distinguishes standards from design-system guidance and project heuristics.

It is not a component library, a visual-design skill, or a rule that every table needs every control.

Choose a skill

Skill Use it for Invocation
completeness-plan Plan a new or expanded resource-management workflow before implementation Automatic when the request matches; explicit invocation also works
completeness-build Implement a new or expanded workflow after approving its completeness contract Automatic when the request matches; explicit invocation also works
completeness-audit Audit an existing app, module, or feature for evidenced gaps Explicit $completeness-audit only
consistency-critique Compare analogous modules and define one target standard Explicit $consistency-critique only

The automatic skills deliberately stay out of marketing pages, styling-only work, isolated bug fixes, and fully specified one-off changes. Manual invocation always remains available.

Install

The skills CLI is the most portable installation path.

# Choose skills and detected agents interactively
npx skills add Pleko-script/web-app-completeness

# Install one skill globally
npx skills add Pleko-script/web-app-completeness --skill completeness-audit -g

# Install the complete pack globally
npx skills add Pleko-script/web-app-completeness --skill '*' -g

For automation, name the destination agents explicitly:

npx --yes skills add Pleko-script/web-app-completeness --skill '*' -g \
  -a codex -a claude-code -a cursor -a github-copilot -a opencode -y

Do not substitute --all: in the CLI it means every supported agent, not every skill for your chosen agents.

Claude Code plugin

claude plugin marketplace add Pleko-script/web-app-completeness
claude plugin install web-app-completeness@web-app-completeness

Claude plugin skills are namespaced: invoke the manual skills as /web-app-completeness:completeness-audit and /web-app-completeness:consistency-critique. A skill installed directly with the skills CLI uses /completeness-audit or /consistency-critique in Claude Code. The $skill-name form in the examples below is the portable prompt form used by Codex and the repository's cross-client datasets.

The repository also ships a portable Agent Plugins manifest and a Codex plugin manifest. See client compatibility for current installation and invocation caveats.

Examples

Plan an employee-management area for this application.

completeness-plan inspects the codebase, classifies relevant capabilities, and returns a scoped plan without changing code.

Build asset management here. Use the recommended completeness defaults and proceed.

completeness-build may skip its single approval stop because the prompt explicitly pre-authorizes the recommended contract.

$completeness-audit Review the forms module and report only evidenced gaps.

The audit remains read-only. If rendered behavior was not inspected, relevant claims are labeled manual-review.

$consistency-critique Compare users, customers, and assets. Define one target pattern.

The critique produces a capability matrix and migration batches; it does not refactor the modules.

What gets considered

  • Resource lifecycle: create, view, edit, archive/delete, restore when applicable
  • Dataset reachability, search, filters, sorting, pagination alternatives, and state persistence
  • Selection, row actions, bulk operations, import, and export
  • Forms, validation, relationship controls, and retained input
  • Loading, refresh, empty, no-results, permission, success, and failure states
  • Mutation safety, recovery, authorization, and long-running work
  • Keyboard access, focus, semantics, responsive behavior, and status announcements

These are decision areas, not universal requirements. A short fixed list may need no filtering or pagination. An irreversible deletion may need confirmation while a reversible action may be better served by undo.

One invariant is not optional:

displayed rows = paginate(sort(filter(complete reachable dataset)))

Filtering or sorting a server-truncated slice while leaving other records unreachable is a correctness defect.

Evidence model

Rules declare their authority and strength:

  • standard: normative accessibility, security, or platform requirements
  • official-pattern: conditional guidance from established design systems and frameworks
  • project-heuristic: explicit, testable product reasoning maintained by this project

Strength is one of must, must-decide, should, or may. Most product capabilities are must-decide: the agent must classify them, but it may conclude defer or not-applicable with a reason.

Read the methodology for statuses, evidence requirements, and severity rules.

Framework support

The core is framework-neutral. The first detailed implementation path covers TypeScript, React, Next.js, shadcn/ui, TanStack, and tRPC without requiring any of them. Refine and React Admin are pattern sources, not dependencies.

Quality

Every pull request runs deterministic, credential-free validation:

  • Agent Skills core fields, declared client extensions, and plugin manifest contracts
  • self-contained references and generated catalog drift
  • at least twenty positive/negative routing cases per skill
  • at least three synthetic scenario fixtures per skill
  • stable rule IDs and objective golden expectations

Trusted pre-release runs exercise Codex and Claude Code three times per case. Cursor fixtures and a provider canary ship in v0.1, but live Cursor routing is advisory until it has been calibrated in this environment. See evaluation design.

Contributing

Start with CONTRIBUTING.md. A rule change needs source provenance and affected scenario coverage; a trigger-description change needs routing cases on both sides of the boundary.

Sources and inspiration

The rule catalog cites primary sources including WCAG 2.2, WAI-ARIA APG, GOV.UK Design System, Carbon, Shopify App Home patterns, OWASP, React Admin, and Refine.

Repository architecture and writing were informed by mattpocock/skills, the Cursor pstack plugin, and andrej-karpathy-skills. No skill text is copied from those projects.

License

MIT © 2026 Nik Tykhomyrov

About

Portable agent skills for complete operational web-app workflows

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages