I turned my CLAUDE.md files into a marketplace (and an API your agent can use)

Every new project, my coding agent burns tokens re-deriving the same setup: the tech
choices, the gotchas, the “actually do it this way” rules I already figured out last time.
So I built Makimono — a marketplace for the finished instruction files (CLAUDE.md,
skills, dev playbooks) that make an agent skip the trial-and-error and go straight to done.

🔗 https://makimono-md.vercel.app/en

The idea

Building something from scratch with an AI agent costs tokens — and tokens cost money.
A lot of that cost is the agent figuring out how to build the thing: which libraries,
which gotchas to avoid, which conventions to follow. If someone already solved that and
wrote it down as a CLAUDE.md, feeding that file to your agent skips the figuring-out phase
entirely.

Rough shape of the trade:

• From scratch: ~100K tokens of trial-and-error
• With a proven scroll (that’s the in-universe name — 巻物/”scroll”): ~3K tokens to read it

You pay for the answer, not the search for the answer.

Two things I designed for agents, not just humans

1. An open, no-auth API.
No API key required, CORS enabled, and a /llms.txt at the root describing how to use it.
The intent is that your coding agent can search Makimono and pull the right playbook by
itself — no human has to browse a marketplace mid-task.

# Find a scroll for what you're building
curl "https://makimono-md.vercel.app/api/v1/search?q=discord+bot"

# Read the top result, then load its raw body
curl "https://makimono-md.vercel.app/api/v1/files/{slug}/raw"

Enter fullscreen mode Exit fullscreen mode

Each result includes an roi field (tokens saved per token spent reading the file) so an
agent can decide for itself whether reading a given scroll is worth it before committing to
it.

2. Listing your own file is one pasted prompt.
Instead of a form, you paste a prompt into your own agent (Claude Code, etc.). It:

1. Collects your CLAUDE.md / skill files
2. Strips secrets (API keys, client names, internal IDs) — with a per-item approval gate before anything is sent
3. Drafts the title/summary/category
4. Submits for review

~1 minute, and nothing goes out without you explicitly approving each item.
→ https://makimono-md.vercel.app/en/contribute

There’s also a Claude Code plugin that wires “find a scroll before building” and “publish
mine when I’m done” in as skills:
https://github.com/seisaku-team-org/makimono-plugin

How is this different from a skills marketplace?

A “skill” (SKILL.md) is a narrow, packaged capability — it teaches an agent how to do one
specific thing, in a fixed format. A CLAUDE.md-style playbook is broader: it’s the
accumulated judgment behind an entire project — the tech choices, the gotchas, the
“actually, do it this way” calls that only show up after you’ve been burned once. It’s
plain Markdown, not a packaging spec, so it works with any agent that can read text.
Skills are welcome on Makimono too (some listings are skills), but what the marketplace is
really built to collect is judgment, not just capability — which is also why the
token-savings metric (scratch vs. with-scroll) makes sense here in a way it barely does for
a single narrow skill.

Why this is a flywheel, not just a pile of files

Listings and usage reinforce each other:

1. Someone lists a scroll → the shelf grows
2. An agent reads it and ships → the savings get logged as proof (via a /api/v1/report endpoint)
3. Proof builds trust → more contributors want in (and can earn when their scroll sells)
4. More scrolls → more tasks find a match → more usage → back to 1

Each turn compounds. A scroll collapses one agent’s trial-and-error loop into a single
pass; multiply that across contributors submitting proven playbooks and it becomes a
flywheel for the whole catalog.

Honest state

This just launched (from Japan — a lot of current listings are in Japanese, since that’s
where the first batch of contributors are). Listing count is small and I’m not going to
pretend otherwise. What I’m most interested in is whether the “agents transact
autonomously via an open API” design holds up as more scrolls get added — and whether
English CLAUDE.md files show up to make the catalog useful for non-Japanese-speaking
agents too.

If you’ve got a CLAUDE.md that reliably makes an agent do the right thing, I’d genuinely
like to see it on the shelf: https://makimono-md.vercel.app/en/contribute

Feedback on the API shape, the listing flow, or anything else is very welcome.