Agent skills are great. I wanted to share some of my favorites from Jesse Vincent's 'Superpowers' skill pack 𝚠𝚛𝚒𝚝𝚒𝚗𝚐-𝚙𝚕𝚊𝚗𝚜 skill…

Agent skills are great. I wanted to share some of my favorites from Jesse Vincent's 'Superpowers' skill pack 𝚠𝚛𝚒𝚝𝚒𝚗𝚐-𝚙𝚕𝚊𝚗𝚜 skill produces much better plans than any harness' built in plan mode that I've tried. Your agent will ask you a set of thoughtful questions one-by-one in the terminal (not in any specialized Q&A component). This pattern makes the agent thoughtful about its follow up questions, rather than generating a set of similar questions that warrant duplicate responses. Your agent then steps through each section of the plan with you, which is a great UX because it gives you breathing room to review each part and challenge assumptions (instead of throwing a huge doc at you)! It writes the final plan doc to file and lets you give further feedback, which is translated to very precise edits (versus Claude Code's built-in plan mode which rewrites the entire plan). Plans generally feel more polished. I think this is because they're more thoroughly researched, and success criteria is spec-driven rather than hand-wavy 𝚎𝚡𝚎𝚌𝚞𝚝𝚒𝚗𝚐-𝚙𝚕𝚊𝚗𝚜 skill makes agents much better at long-horizon coding work. This skill outlines two modes to implement a plan: 1) subagent driven, and 2) orchestrator-driven. I always choose subagent driven (Superpowers recommends this if your harness supports subagents). Selecting this mode will invoke the 𝚜𝚞𝚋𝚊𝚐𝚎𝚗𝚝-𝚍𝚛𝚒𝚟𝚎𝚗-𝚍𝚎𝚟𝚎𝚕𝚘𝚙𝚖𝚎𝚗𝚝 skill. Each task (a section of the plan your agent generated prior) is completed, assessed, and polished by a team of three subagents: an implementer, a spec reviewer, and a code quality reviewer. These agents iterate until the task satisfies the spec. If all looks good, your agent moves on to the next task (and spawns a new set of subagents) 𝚛𝚎𝚚𝚞𝚎𝚜𝚝𝚒𝚗𝚐-𝚌𝚘𝚍𝚎-𝚛𝚎𝚟𝚒𝚎𝚠 skill is invoked when I ask my agent to review a PR. To me, the review it produces matches Codex's /𝚛𝚎𝚟𝚒𝚎𝚠 mode in raw quality but feels more flexible. If you ask Codex to justify a review comment, it seems to lose all context from the review itself (outside of the final review output). With this skill, your agent retains the full context and does a good job backing up its comments 𝚋𝚛𝚊𝚒𝚗𝚜𝚝𝚘𝚛𝚖𝚒𝚗𝚐 skill has a really neat process called "visual companion" which the agent invokes when making frontend changes that are expensive to iterate on visually. When I was designing a 'thinking' loading indicator for a chat UI, my agent invoked this skill, then spun up localhost to show me variations of the spinner in isolation (so I didn't consume tokens while testing). This sounds gimmicky but it's the type of thoughtful polish that gets me really excited about using a product There's a lot of slop skills out there, but I believe well-written skills have the potential to drastically change the behavior of your agents

Source: Harrison Chase (X) | 2026-04-10

Related