Playlist Builder — Get a Verified Playlist You Can Actually Import

Your own playlist, made by your AI.

Everyone's library is full of playlists made by strangers and algorithms. Mixtape is an agent skill for Claude Code and OpenAI Codex that turns a described mood, a handful of seed songs, or your actual listening taste into a real playlist in your streaming account — delivered as a one-click Soundiiz import link. No platform API keys, no developer accounts.

"Playlist for cooking on a Friday night, city pop, ~20 tracks, YouTube Music" → a numbered tracklist with an arc, and a link. Open → confirm → it's in your library.

What it does

Mixtape turns a mood, a few seed songs, or the playlists you already love into a sequenced playlist you authored. Before it hands anything to Soundiiz, it checks every title and artist against the selected Apple Music storefront, rejects wrong covers/remixes/DJ-mix versions, and produces a catalog evidence report. A fingerprint check then prevents unverified or subsequently edited tracklists from being posted. Iterations keep one stable library identity: revision labels stay in the private history, while the streaming playlist keeps its human title and superseded Mixtape versions are reconciled away by exact platform ID.

That matching layer matters. A beautiful 33-track recommendation is still a broken product when only eight songs arrive in the user's library.

When it fires

• Describe a scene: “late-night coding, warm electronic, 90 minutes.”
• Give seed tracks: “these are the three kinds of rock I mean—expand them.”
• Share existing playlists: “work out what I actually like, then mix familiar songs with discoveries.”
• Report a failed import: Mixtape records misses, repairs the matcher, and updates one complete playlist instead of making you manage fragments or duplicate version-numbered playlists.

Design notes

A mixtape is the opposite of an algorithmic playlist: it's authored. You bring the taste; the agent does the hours of picking, matching, and sequencing you never have time for. The goal is the feeling of "this playlist is MINE."

Three modes

Mode	Say something like
Quick order	"Make me a rainy-Sunday jazz playlist, 25 tracks, Spotify"
Seed expansion	"I like two kinds of rock — these 2 songs are one kind, these 2 are the other. Expand to 30."
Taste profile	"Here are my existing playlists — figure out what I actually like, then curate for me."

The taste profile persists in ~/.mixtape/ (plain Markdown, stays on your machine). Over time the agent blends songs you already love with adjacent discoveries — and remembers what failed to match and what you vetoed.

Install

Install the standalone skill with the Skills CLI:

npx skills add m1nga/mixtape

Or clone it directly for a specific agent:

git clone https://github.com/m1nga/mixtape.git ~/.agents/skills/mixtape

Then just ask for a playlist in a session — no command to memorize.

Requirements: curl + python3 (present on macOS/Linux by default), and a free Soundiiz account with your streaming platform connected (that's where the import authorization lives).

Supported destinations

Platform	Import link preselect
Spotify, YouTube Music, YouTube, Deezer, Tidal	✅ preselected
Apple Music	one extra tap on the import page

How it works

The agent generates and sequences the tracklist (real, released recordings only — it verifies when unsure), then checks every title+artist pair against the requested Apple Music storefront. The canonicalizer writes the official display metadata, a catalog evidence report, and a fingerprint. Only that fingerprinted payload can reach Soundiiz's free, key-less Playlist Import API.

The Soundiiz contract is deliberately enforced at the last mile: its field is artists (plural), as an array or string. Sending artist appears to work because the endpoint accepts the JSON, but the artist is silently ignored and matching falls back to the title alone. Mixtape migrates legacy payloads but never sends the wrong field.

You get a link valid for 24 hours; the confirmation click doubles as your preview/veto step. Limits: ≤200 tracks per link and 10 requests/min.

For Apple Music on macOS, the included reconciliation helper inventories the local library, dry-runs an exact-ID plan, then (only with --apply) renames the retained playlist, deletes explicitly named predecessors, and reads the library back. It refuses stale names, unexpected track counts, and unrelated playlists that happen to share the final title. Optional protected IDs are also checked and returned as unchanged in the final receipt.

Field-tested

The bug that prompted the strict pipeline was a two-hour running mix where only 8/33 tracks transferred despite Apple Music carrying almost all of them. The first “fix” reported 0 unconfirmed but had actually changed The Less I Know the Better — Tame Impala into a similarly titled song by ALB and Raw Control — Discip into a DubMabs track. The revised matcher now rejects both cases, blocks the post, and shows the catalog evidence needed to choose a real replacement. The verified replacement imported all 33 tracks; the final read-back kept one 33-track playlist named 跑步 and removed the two superseded managed versions.

中文说明

Mixtape 是给 Claude Code / Codex 用的歌单 skill：一句话描述场景（「周五晚做饭的 City Pop，20首」）、给几首种子歌，或让它分析你已有歌单；它负责选歌、编排、逐首 核对 Apple Music 曲库，再给出 Soundiiz 一键导入链接。它会同时校验歌名和歌手， 找不到或疑似错版就停止发布，不再让“33 首只进去 8 首”到最后一步才暴露。口味档案 存在本机 ~/.mixtape/，越用越懂你。迭代时平台里的歌单名称保持不变，版本号只写进 内部历史；成功导入后按精确 ID 清理旧版，并回读确认只剩一个当前歌单。

Notes

• Not affiliated with Soundiiz; uses their public Playlist Import API.
• Song selection quality depends on the model running the skill.
• Local taste files are not uploaded by the bundled scripts. Catalog lookups send song queries to Apple; Soundiiz receives the playlist title, description, destination and {title, artists} entries. Your agent provider also processes what you put in the conversation.

MIT © m1nga

What “verified” covers

Apple catalog confirmation checks recording identity in the selected storefront. It does not prove Spotify, YouTube Music or another destination will match every track. The local fingerprint detects tracklist edits after that check; it is not a cryptographic signature or independent attestation. Final delivery still needs the destination import result and, for revisions, the library read-back.

The 15 bundled offline tests passed on 2026-09-07 (Asia/Shanghai), including wrong-artist rejection and exact-ID reconciliation. Run python3 -m unittest discover -s tests -v from the installed skill directory. This is fixture evidence, not a new live import.

Author

Built by Ming. The design notes above explain the real problem and tradeoffs that shaped this skill.