Make Claude Tell You the Truth at a Glance: The Quick Recap Skill

Coding agents bury the most important fact in a wall of output: did they actually do what you asked? This video walks through a small open-source skill called Quick Recap (shown as a /quick skill in the GitHub README) that fixes that. It forces the agent to end every message with a single color-coded line that says done, partially done, or blocked. The demo on screen shows the skill running inside a real session, with green, yellow, and red status dots appended to the agent's replies, plus the GitHub repo and the one-line install command.

The Problem It Solves

Claude is verbose. It produces long output and does not always make it clear when it knowingly skipped part of your request. If you are not explicit about how it should report back, you can read a long response and still not know where things stand. The on-screen demo makes this concrete: the user asks "did you do my third ask? github actions?" and the agent replies "You're right to push back. No, I didn't. I'll do it now." That admission was sitting inside a long block of text. Another frame shows the agent rambling about skills it read while a caption points to the buried line and reads "where's waldo," which is the whole point. The important signal gets lost.

What the Skill Does

Quick Recap adds a convention so every completed response ends with a clear status marker. The GitHub README on screen describes it as a "completion convention so every completed response ends with a clear green, yellow, or red work-status line." The three states map to:

• Green: it did everything and you are good.
• Yellow: it did part of your ask, plus exactly what it needs from you to continue.
• Red: it explicitly did not do what you asked, plus exactly what it needs in order to do it.

The README phrases the same three states as "done, pending a specific non-routine step, or blocked on the user." Once you see the marker, you can take the action, push back on it, or scroll up for more context on why it stopped.

The Demo

The video shows a real agent session running these recaps. Each turn ends with a colored dot and a one-line summary:

• Green dot: "All tasks completed"
• Yellow dot: "Listed 5 skills, need GitHub auth to complete"
• Red dot: "Need you to authenticate with GitHub to begin"

The session is about installing skills from a framework called agent-native. The agent finds the install command npx @agent-native/core@latest skills add <skill-name> and lists available skills like visual-plans, visual-recap, design-exploration, assets, and context-xray. At one point the user types the blunt prompt "You are an idiot," which is the creator demonstrating that you can push back directly and the recap still keeps the state legible. A later frame shows the agent reporting "Fixed test clobbering and verified Plan MCP tools" with a green dot and an edit to connect.spec.ts (+7 -0), running in what looks like a Codex session with "Full access" enabled.

Install and Defaults

The skill is open-sourced at github.com/BuilderIO/skills, titled "Skills for coding agents" and described as "small, composable skills for coding agents." The README pitches them for teams that want the agent to stay sharp where judgment matters: orchestration, review, planning, validation, and clear communication, without a giant process framework. You install the pieces you want.

The quick install shown is npx @agent-native/skills add. Beyond installing the skill, the CLI can append a note to your CLAUDE.md or AGENTS.md so the recap behavior runs by default on every session. Because it works through that standard instruction file, it is agent-agnostic: the creator says it works with Claude Code, Codex, the CLI, and the UI. The repo also lists adjacent skills in the same README, including an effort or budget skill (/eff...) that preserves an expensive frontier model for planning and final review while routing bounded heavy work to cheaper agents, and /visual-plan.

Key Takeaways

• Quick Recap makes a coding agent end each response with a green, yellow, or red status line so you know at a glance whether it finished, did part of the work, or is blocked.
• Yellow and red states must state exactly what the agent needs from you next, which removes the guesswork from a long response.
• Install with npx @agent-native/skills add; the source is github.com/BuilderIO/skills ("Skills for coding agents").
• The CLI can write the convention into your CLAUDE.md or AGENTS.md, making the recap the default behavior across sessions.
• It is agent-agnostic and works with Claude Code, Codex, and other agents because it rides on the standard instruction file.
• The same repo ships other composable skills, like an effort/budget router and /visual-plan, so you install only what you want.