Moonshot

English · 简体中文

A lot of people can't write prompts and still want one sentence to produce something good. This skill is for them. You hand it a sentence — "build me a booking site for my coffee shop" — and it turns the sentence into an actual project: it sets its own bar, has the work torn apart by a separate agent, keeps improving until it stops improving, and writes everything to disk and git as it goes, so a run that takes hours survives being interrupted.

npx skills add Likely7/Moonshot-skill --skill=moonshot

Where the idea comes from

A prompt made the rounds a while ago. Someone wrote Claude a single paragraph and let it run for seven hours. Out of nothing it produced a Call of Duty-grade browser shooter (Claude-of-Duty). No art assets, no plan — the architecture, visuals and sound were all improvised by the agent itself.

The paragraph contained no technical instructions at all. It said three things: the bar is absurdly high, finished work gets criticized by a different agent, and don't stop early. That's the part that worked: set the posture, not the method. How to do the work is left to the agent doing it.

Moonshot extracts that idea so it works for any project, then adds what a one-off run doesn't need but a real project does.

What it adds

1. Any domain. Websites, mini-programs, games, slide decks, financial models, even "write me a video-generation prompt." What counts as good, and how work gets criticized, is decided by the agent per task. The skill doesn't decide for it.

2. Project management. Context won't survive a long run, and even a short one deserves a tomorrow. So every run leaves the project's state on disk: the goal contract, the plan, verified progress, decisions with their reasons, the evidence from every review round. One file for a small task, a full .moonshot/ set for a big one, with git commits only when checks pass. Close the window, switch models, come back tomorrow — it picks up where it stopped.

3. A way to stop. "Loop until perfect" never ends. Moonshot watches the improvement curve: when a few rounds in a row barely move it, the choice goes back to you. Here's the curve, here's the budget left, keep going or wrap up? Ask for one more round and you get one more round. Exactly one, so "just once more" can't go on forever.

And one rule holds throughout: the working agent is never told the goal might be unreachable. The bar is high so it runs hard. How close it gets is found out by running, and the closing report states it as fact — what made it, what fell short, and what the reviews actually relied on.

Two modes

Just build (default). The goal is clear: size the work, set a checkable bar, start.

Interview first. For when you're serious about the thing. Before building, it digs: 5–10 questions per round, each with a recommended answer so you can clear a batch in one line ("the rest, go with yours"), each round going deeper based on your answers. It looks up what it can instead of asking. It keeps going until it knows the project inside-out — then tells you it's ready. Say "just start" at any point and it starts. The questioning style is borrowed from mattpocock's grill-me, with its real flaw fixed: one question per round-trip, hundreds of turns deep.

Why not just paste an ambitious prompt yourself

The original paragraph is public and you can. What you'd be missing:

• The wording is harder than it looks. "Get your work adversarially reviewed" and "size it before you start" do nothing if phrased loosely.
• One run is easy; the next one starts from zero. Without state on disk and git checkpoints, a long task that dies is gone.
• Nothing tells you when to stop. An agent saying "done" means nothing. Someone has to check the evidence, and someone has to watch the curve to see whether it really topped out.

Moonshot has these built in.

What it won't do

• Pretend to have capabilities it doesn't. If the model can't see images, it switches to checks that don't need sight and says so at the end.
• Spend your money. Checks that cost money (like an actual video render) stay out of the loop: they run once, at the end, with a price quote, after you approve.
• Lock you to a platform. Loops, parallelism, sub-agents — whatever your environment provides (Claude Code, opencode, Codex, others). No vendor-specific commands live in the skill itself.

Install

npx skills add Likely7/Moonshot-skill --skill=moonshot

Or manually:

git clone https://github.com/Likely7/Moonshot-skill.git
cp -r Moonshot-skill/skills/moonshot ~/.claude/skills/moonshot

Then say what you want built, or type /moonshot.

Layout

Moonshot-skill/
  skills/
    moonshot/
      SKILL.md                  # the constitution: posture not method, two modes, six steps
      references/
        grilling.md             # how interview mode runs
        state-templates.md      # project state templates — a starting point, not rules
  LICENSE
  README.md
  README.zh-CN.md

Credits

The idea comes from the seven-hour run behind mshumer/Claude-of-Duty. Interview mode borrows from mattpocock/skills' grill-me.

MIT License