Source: raw/5_NEW_Vibe_Coding_Repos_You_Haven_t_Heard_of_Yet.md (YouTube roundup, video id UnzD_bwylWs). Repo metadata (URL / stars / license / maintenance) independently verified on GitHub 2026-06-17.

A YouTube “5 new repos” roundup surfacing under-the-radar vibe-coding tools, four of which are not yet covered in this wiki (the fifth, SkillSpector, already has its own article). Each is a small, mostly Claude-Code-ecosystem tool that targets a concrete pain point — speech-to-text, over-engineering, codebase auditing, and architecture diagrams. The repos were verified to exist with permissive licenses; star counts and ages are noted because three of the four are very new (cite the mechanisms, treat self-reported metrics as provisional).

Key Takeaways

  • Handy (cjpais/Handy) — free, fully-offline, cross-platform push-to-talk speech-to-text desktop app; a Wispr Flow / WhisperFlow alternative that saves ~$20/mo. Tauri 2 (Rust + React/TS), Silero VAD, transcribes with Whisper (Small→Large, GPU) or Parakeet V3 (CPU). Homebrew-installable, push-to-talk hotkey. ~24k★, MIT, 16-month history, actively maintained — the most mature of the five. Lacks Wispr’s AI-rewriting of dictated text.
  • Ponytail (DietrichGebert/ponytail) — anti-over-engineering plugin for Claude Code (+~10 other agents) that injects a 6-step necessity decision ladder (YAGNI → stdlib → native → deps → one-liner → minimal) and ships /ponytail-review and /ponytail-audit (dead-code / YAGNI scan) commands plus lite/full/ultra/off constraint levels. MIT. Exploded to ~32k★ but is only ~days old — the “replaces 50 lines with 1 / 47–77% cheaper” figures are self-reported. ^[ambiguous]
  • improve (shadcn/improve) — read-only codebase auditor Claude Code skill by the real shadcn (shadcn/ui author). Fans out parallel Explore subagents across ~9 categories (bugs, security, perf, tests, tech-debt, deps, DX, docs, direction) and writes a markdown remediation PLAN rated by effort / confidence / risk for cheaper models to execute. It plans; it never implements. MIT, ~5k★, new but credible-author.
  • draw.io architecture-diagram skill — Agent Skill that turns a natural-language prompt or an existing codebase’s import graph into auto-laid-out, editable .drawio architecture diagrams (presentation / state / service / database layers + edge functions), with vision self-check and multi-round refinement, exporting PNG/SVG/PDF. Several variants exist (e.g. Agents365-ai/drawio-skill, MIT, ~4k★; the official jgraph/drawio-mcp, Apache-2.0) — confirm which one before installing. ^[ambiguous]
  • SkillSpector (NVIDIA) — security scanner for skill libraries; already covered at SkillSpector (this roundup adds operator cost/usage detail, folded into that article).

Verified repo facts

ToolRepoStarsLicenseMaturity
Handycjpais/Handy~24kMITactive, 16-mo history
PonytailDietrichGebert/ponytail~32kMITbrand-new (<2 wk)
improveshadcn/improve~5kMITnew, credible author
draw.io skillAgents365-ai/drawio-skill (one of several)~4kMITactive

Try It

  1. Handybrew install --cask handy (or download from handy.computer), pick Parakeet V3 for CPU-only machines, bind a push-to-talk hotkey, and dictate into any text field offline. Free.
  2. improve — install the skill, point it at a repo, and let it write a PLAN.md of effort/confidence/risk-rated findings; feed the high-confidence/low-risk items to a cheaper model to execute. It won’t touch code itself.
  3. Ponytail — add the plugin, then run /ponytail-audit on an over-built module to get a dead-code / YAGNI report before refactoring. Treat its savings claims as provisional until you measure your own.
  4. draw.io skill — install one verified variant, prompt it to diagram your codebase, then open the resulting .drawio in diagrams.net and edit — the output is editable, not a flat image.
  5. Before installing any of these, run SkillSpector against the repo URL — these are young, single-maintainer projects.

Open Questions

  • The draw.io “skill” featured in the video isn’t unambiguously identified; several functionally similar repos exist. Confirm the exact one before recommending. ^[ambiguous]
  • Ponytail’s and improve’s headline efficiency/cost numbers are self-reported on <2-week-old repos — worth independent benchmarking before citing as fact. ^[ambiguous]