Skills that make your agent grow

Reads your PRD, invents the people who'd actually use the thing, then drives your running app through each of their flows in a real browser. Most UX bugs are invisible in code review — a stale closure, a caret that lands one character off, a menu clipped at the viewport edge — and only surface when someone walks the path.

• Personas from your PRD, each with a concrete flow to run
• Nielsen's 10 heuristics, scored 0–4 with evidence
• Accessibility audit — WCAG checks plus a keyboard-only walkthrough
• Edge states — empty, extreme, loading, multi-tab, refresh

Grounded in 20+ peer-reviewed papers · needs browser automation · MIT

A practitioner's guide to designing AI research, distilled from 25+ published Anthropic papers (2024–2026). It's deliberately about how to do the work — the frameworks, experimental designs, and measurement approaches — rather than what any one paper concluded. Findings age; method travels.

• Safety & alignment — red-teaming, constitutional classifiers, sabotage arenas
• Interpretability — concept injection, persona vectors, circuit tracing
• Behavioral evaluation — automated pipelines, model-as-judge calibration
• Six cross-cutting patterns that recur across all seven domains

25+ papers · 7 domains · SKILL.md plus deep references · MIT

A meta-skill — it works on your other skills. Most skills are frozen the day they're written, while you have a thousand quiet preferences you never spell out: dates as YYYY-MM-DD, summary first, and please edit the one sentence rather than regenerating the whole document. Bloom catches those leaks, acts on them next time, and eventually folds the durable ones back into the skill files themselves.

• Capture — mines each run for the assumptions you never said out loud
• Recall — applies what it learned by default, and tells you so you can override
• Evolve — injects durable lessons into your skill files, at the point of action
• Seed it — declare your own pain points and skip the wait for evidence

Four modes · keeps its memory in plain markdown you can read · MIT