DEV Community

Leon
Leon

Posted on

Programs Beat Prompts: How Tap Turns AI Understanding Into Deterministic Programs

The Problem

AI agents need to operate interfaces — read data, click buttons, fill forms. But AI is too slow ($0.10/call, 2-5s latency) and too unreliable to drive every interaction.

The Insight

Operating an interface is a solved problem the moment you figure out how. The hard part is understanding the page. That's what AI is good at. The easy part is executing the same steps again. That doesn't need AI.

Tap: Forge Once, Run Forever

Tap separates understanding (AI, once) from execution (deterministic, forever):

forge_inspect → forge_verify → forge_save → tap.run
   AI analyzes     AI tests      AI saves    runs forever, $0
Enter fullscreen mode Exit fullscreen mode

The Protocol

8 core operations — the irreducible atoms of interface interaction:

eval · pointer · keyboard · nav · wait · screenshot · run · capabilities
Enter fullscreen mode Exit fullscreen mode

17 built-in operations composed from core. One program runs on Chrome Extension, Playwright, and macOS native apps.

Unix Pipes

Taps compose like Unix commands — auto TTY detection, no flags needed:

tap github trending | tap tap/filter --field stars --gt 1000
tap watch github trending --every 5m | tap tap/filter --field stars --gt 500
Enter fullscreen mode Exit fullscreen mode

Self-Healing

tap doctor
  ✔ github/trending     score=1.0  25 rows (180ms)
  ✘ douyin/hot           score=0.0  0 rows
Enter fullscreen mode Exit fullscreen mode

Health contracts + declared examples = taps verify themselves.

Security: 3 Layers

  1. Sandbox — Deno Worker with zero permissions
  2. Static analysis — 7 CI checks on every community PR
  3. Data isolation — .gitignore blocks secrets, no git remote on user taps

v0.5.0 Highlights

  • 199 constraint tests (safety + quality + principle)
  • tap doctor — self-healing health checks
  • tap contribute — PR workflow to community skills
  • tap watch — outputs changes only, time dimension
  • Forge captures real API traffic for accurate strategy
  • Sandbox for untrusted community taps

106 skills across 50+ sites

GitHub, Reddit, HN, X/Twitter, YouTube, Bilibili, Zhihu, Xiaohongshu, Weibo, Medium, arXiv, and more.

GitHub: github.com/LeonTing1010/tap
Skills: github.com/LeonTing1010/tap-skills

This post was published using Tap itself (devto/post tap).

Top comments (0)