Inspiration and what it does

P100 brings back broadcast teletext — the 1980s TV text service that died with analogue television — and puts it to work as a live status desk for the services a small software agency depends on.

During an incident we open six vendor status pages, each with its own layout, its own clock and its own idea of what "degraded" means. P100 gives them one shelf: a numbered index (page 100), one page per vendor (101–106), and page 199 for the limits. Every page is a strict 40×24 character teletext grid, navigated by page number with a keyboard or the on-screen remote, exactly like a 1980s set.

  • Six public Atlassian Statuspage summary.json feeds fetched server-side in parallel, 5-second timeout per source, 60-second cache.
  • Affected components are listed before operational ones; a vendor with many affected components spans subpages without ever taking another vendor's page number.
  • A failed or slow source renders NO SIGNAL; incidents = [] is never shown as an all-clear. Age is time since retrieval, not time since the vendor changed.
  • Authentic front end: the eight teletext colours, double-height header with a live clock, blocky mosaic logo, CRT scanlines, flashing status. Scales from a 360 px phone to a desktop and respects prefers-reduced-motion.

It is a vendor-status viewer: not active monitoring, not an SLA calculator, and not a claim about any customer's availability. P199 says so on the page itself.

How we built it: how Kiro was used

  • Spec-driven development. .kiro/specs/p100/ holds requirements in EARS form (WHEN/WHILE/IF … THE SYSTEM SHALL …), a design that fixes the 40×24 grid rules and the fetch/cache contract, and a task plan worked through to completion. The EARS criteria map one-to-one onto the tests, so the spec is checkable rather than aspirational.
  • Steering. .kiro/steering/product.md, tech.md and structure.md pin the product boundary (vendor viewer, no keys, honest status), the stack (Worker, TypeScript, no framework, pure buildPages) and the module layout. Keeping the boundaries in steering stopped the build from drifting into a monitoring service.
  • Agent hook. run-tests-on-save.kiro.hook runs the suite whenever a src/*.ts file is saved, so the grid-bounds and status rules never went stale during edits.
  • Red before green. The eight acceptance cases in fixtures/cases.json were encoded as tests first. Mutating the page builder (forcing one subpage, treating unknown incidents as green) turned the suite red — four failures — until the correct implementation was restored.
  • Property-based testing (fast-check). For any generated feed input: every page is exactly 24 lines of 40 characters, vendor page numbers are unique and stable for the same service order, and unknown or failed feeds render NO SIGNAL without throwing. The 360 px clipping bug was found in a real screenshot, not by the PBT — the properties guard the grid, they do not replace reading the screen.
  • What would have been slower without the spec. The subpage rule ("never take another vendor's page number") is the part a vibe-coded version gets wrong: without it written down as an acceptance criterion, pagination silently renumbers the neighbours.

Links

The demo narration is synthetic (text-to-speech); every screen in the video is a real capture of the running app.

Built With

Share this project:

Updates

Submission history