Exploring /grill-with-docs: A New Rival to Superpowers in AI Coding Skills

The author, a long‑time AI developer, tried the recently popular "Skills" collection on X, created by Matt Pocock—known for Total TypeScript and now focusing on AI coding agents. His GitHub repo, named skills, has 154 k stars and offers 21 skills split into engineering and productivity groups.

The core engineering flow consists of five slash commands: /grill-with-docs to clarify ideas, /to-prd to generate a product‑requirements document, /to-issues to split the PRD into independent issues, /implement to develop each issue, and /code-review to finish. /grill-with-docs implements a "grilling" approach: the agent asks one question at a time, provides a recommended answer, and only asks the developer for decisions that cannot be inferred, continuing until the entire design tree is resolved. The underlying implementation is only 12 lines of code.

Compared with Superpowers, which contains 14 skills totaling over 3 300 lines (e.g., the test‑driven‑development skill alone is 371 lines), Matt's skills assume the model is usually correct and only intervene at critical anchors. Superpowers, by contrast, assumes the model will slack and blocks every possible escape, injecting hooks that force automatic skill execution from the start of each session.

Installation is a single command: npx skills@latest add mattpocock/skills Afterward, running /setup-matt-pocock-skills in each project configures where issues are stored (GitHub, GitLab, or local markdown) and sets triage labels required by the main flow.

To validate the workflow, the author applied it to an iOS English‑learning app. The five‑step chain processed a single idea into a compiled app, with /grill-with-docs prompting 16 decisions, each requiring manual approval. The skill also generated a CONTEXT.md dictionary containing 17 project‑specific terms and four ADRs, reducing token usage and enforcing a ubiquitous language.

When the author asked the agent to perform research, three parallel research agents fetched and verified thirty papers, producing a report that identified CarPlay as lacking a voice‑dialogue category and suggesting a shift to audio playback. The decision was recorded in an ADR after the author replaced the suggested model with DeepSeek.

The /to-prd skill is disabled for automatic model invocation, requiring the user to type the slash command manually, which the author found a useful reminder of the design’s intentional restraint.

Creating the PRD generated 36 user stories; /to-issues split them into seven linked issues using GitHub's native sub‑issue and blocking relationships. The subsequent TDD skill ran 22 tests (13 Swift, 9 contract tests), compiled the app with the iOS 26.5 SDK, and achieved sub‑second first‑byte latency on Vercel.

Finally, the author notes that both Superpowers and Matt's skills can coexist: Superpowers aggressively pushes its 14 skills via hooks, while Matt's skills stay mostly manual. Depending on whether a project prefers the agent to make every decision or the developer to retain control, each approach offers a distinct experience, and the author plans to experiment with swapping the two pipelines to compare delivery quality.