Delegate repo work to parallel external coding agents (Codex CLI, Cursor CLI, Gemini or Antigravity CLI, or others) and calibrate how hard to review what they return. Use whenever the user says an external agent can pitch in, hands off a doc/task to one,…
Audit a codebase for AI-generated quality decay and fix it, using deterministic analyzers first and reading second. Use this whenever the user wants to review, audit, or clean up code that AI agents wrote; complains that a codebase is bloating, sprawling,…
Explore a codebase and generate its first SYSTEM.md, a single tool-agnostic living reference any coding agent (Claude Code, Codex, Cursor, Aider, Gemini CLI) reads before working on the repo, or upgrade an existing SYSTEM.md to the current contract. Use when…
Find open-source code that already solves part of a goal (new build, refactor or fix), vet it, take only the needed pieces, and stitch them in with provenance recorded.
Manual online-intelligence pass for coding work, in two modes. Search first: before planning, cheap scout models search the web for how people do the job today, working examples, and known pitfalls, and the main model turns that into a short brief to plan…
Layer skill that propagates the user's GOAL (not just the plan) into every plan, design doc, and executor handoff, and grades how much freedom the executing agent gets. Use whenever a plan/spec/handoff doc is being written for another agent or session to…
Audit a frontend repo against core UI/UX basics (states, accessibility, tokens, spacing, type, color, depth, hierarchy) and report fixes with file:line. Use for UI reviews, design polish, or UX quality checks.
Build a researched, layered study guide on any topic and deliver it as a polished PDF. The notes go big-picture first, then zoom into the details needed to fully understand the concept or flow, with diagrams (fetched from good sources or drawn), memorable…
Investigate what is REALLY happening behind any topic, headline, claim, snippet, trend, or incident: research it online, collect the competing stories and suspected root causes, test each against evidence, and tell the verdict as a plain-language story. Use…
Use when the user asks to collect, research, scrape, or save RAW source files, one file per source, web research notes, or structurally cleaned documents about any topic. Use when they say save raw, keep the original text, don't change the wording, one file…
Use after a RAW research pass (or when the user asks to synthesize, interpret, find patterns, or learn what actually works from collected sources). Reads sources/<topic>/ without modifying them and writes a plain-language insights/<topic>.md — easy…