The Agent-Skills Bake-Off

gstack sprint — Think→Plan→Build→Review→Test→Ship→Reflect; the review gate catches real bugs before ship.

"Markdown Studio" — a live in-browser QA harness runs 26 assertions on every load, surfacing a tamper-evident "self-test 26/26 ✓" receipt. GFM tables + task lists, synced scroll, cursor position, autosave, light/dark. Its own review gate caught a </script> breakout + NUL sentinels.

Addy Osmani's "process over prose" — Google SDLC, verify gates, performance-conscious.

Fast + lean — a live "render 0.4ms" metric, O(n) rAF-coalesced rendering, a perf table (LCP/INP targets), and an in-file ?selftest=1 eval suite. Verification-first.

Vercel's frontend-composition skills — clean architecture, crisp modern design.

"Markdoc" — Geist-grade typography + spacing, 11 single-responsibility modules, Playwright E2E green. The best-designed of the light-theme builds.

Context-engineering — optimize the context window, persist everything else; lean, eval-driven, offline.

"Context Markdown" — dark, cyan, on-brand: content-addressed incremental render (FNV-1a hash → skip-if-unchanged; HUD shows render ms or "unchanged · skipped"), an evals drawer held to a visible spec (18/18 + 18/18 adversarial XSS), strict default-src 'none'. The measurement-first build.

Engineering-grade skills culture — brainstorm-first, strict TDD, systematic debugging, verify-before-done.

Test-first: a 103-assertion parser suite + 36/36 in-browser self-test, defensive XSS-safe correctness. Substance over flash — exactly its TDD doctrine.

ECC — plan → test → implement → review → verify → remember → improve, run literally every loop.

"ECC" — the loop made observable: named snapshot history with restore/delete + auto-snapshot before destructive ops, a live "30/30 tests pass" badge, escape-first parser + protocol allowlist. Its own verify step caught 2 real defects.

Anthropic's official Agent Skills reference — clean, correct, safe, progressive disclosure.

An official-reference-style build — a CommonMark-order two-pass parser, a 46-case suite incl. an XSS matrix, meticulously safe + documented. Understated, exemplary.

UX & product skills grounded in Norman/Cialdini — user-centered, highly usable.

"Markdown Studio" — UX-first: a 5-step onboarding coachmark, forgiving undo/redo, "what + why" tooltips, warm terracotta palette. The most human-friendly.

Waza's restraint doctrine — state the outcome + the red lines, build only what earns its place.

"Waza" — restraint: a 911-line hand-written line-walking parser (code-spans placeholdered first), pure (value,sel)→{text} transforms, native undo via execCommand. 59/59 behavioral + XSS assertions, strict CSP. The tightest, most considered build — fewest tokens, fewest lines.

planning-with-files — write PLAN.md + STATE.md to disk first, then execute against them, verifying each gate.

"planning-with-files" — PLAN.md (12-module architecture) + STATE.md written before any code; the script is organized into the exact 12 planned modules. Recursive blockquotes, indent-nested lists, auto-renumbering, code-span masking, 20 boot-time assertions.

Matt Pocock's composable, model-agnostic skills — shared vocabulary, tight feedback loops, anti-decay.

Composable + clear — pure single-purpose functions with a shared vocabulary and a live "parser: 25/25 tests pass" feedback loop. Teaching-example clarity.

65 full-stack skills (React/NestJS/Python/DevOps) — broad, production-ready.

The most feature-complete — a 20-button toolbar, document outline, in-editor find, five export formats, preferences, three view modes, tables + task lists.

Security-engineering skills from Trail of Bits — safe-by-default, hardened rendering.

"Sentinel" — security-first: a DOM-only renderer (never innerHTML on untrusted input), a scheme allowlist, 12 neutralized adversarial payloads + a live security-model dialog.