/mm-publish

The /mm-publish skill drives the full creator publish flow inside your AI agent. You only confirm decisions — the agent runs every CLI command. The user creates the draft on the web first; the skill resumes from a pack-id and handles auth, signing, source discovery, pack assembly, and publish.

This page mirrors the canonical packages/cli/skills/mm-publish/SKILL.md file shipped in the memorymarket npm package. The SKILL.md is what your agent actually reads at invocation time — if this page drifts from it, the SKILL.md wins.

Usage

Arguments

Phase 1 — Auth gate

Before doing anything else, the agent confirms you are signed in and have a registered signing key.

1. Run mm whoami. If output is “Not authenticated” or the command exits non-zero, the agent runs mm login, which opens a browser window for OAuth. Wait for the CLI to print “Logged in as @<username>”.
2. Probe for ~/.memorymarket/keys/signing.pub. If absent, run mm keygen (writes a fresh Ed25519 keypair to ~/.memorymarket/keys/).
3. Run mm register-key to upload the public key to the server. Key already registered is fine — proceed.

If any of these steps error (network, rate-limit, refused), the skill stops and surfaces the error to you. It does not proceed to Phase 2.

Phase 2 — Memory source discovery

Your memory layout is unknown. The agent scans, proposes a list, and asks for confirmation.

Auto-scan locations from CWD and home, in order:

The agent prints a numbered table (index, path, line count, one-line shape guess), then asks: “Which of these should I include? Reply with numbers (e.g. 1,3,5-8), all, or paste your own paths.”

Banned: the skill never invents memory content. Every line in the resulting pack traces back to a file you confirmed.

Phase 3 — Pack assembly

The agent does the extraction in-context — there is no mm pack or amp packCLI. The full AMP extraction pipeline runs inline as Phase 3 sub-phases (3a shape detection → 3b unit extraction → 3c primitive classification + capability routing → 3d draft pack assembly → 3e validation → 3f confirm + sign). Output: a directory containing manifest.json, agents.md, and memory/<capability>.md files. There is no separate skill for the extraction step.

1. Asks where to write the assembled pack. Default suggestion: ./.memorymarket-drafts/<slug>/. Confirms the directory does not already contain a manifest.json (asks before overwriting).
2. Runs the AMP extraction pipeline (in-agent, not a CLI call) against the confirmed sources. Produces a manifest.json on the AMP v0.4 schema (primitive_format: "inline-tag-v0.4"), agents.md routing index, and one memory/<capability>.md per detected capability. The manifest the skill writes follows this exact shape — every field below is required by mm publish validation:
   manifest.jsonCOPY

   {

   "name": "[pack-slug]",

   "title": "[Pack Title]",

   "version": { "number": 1 },

   "layout": "semantic-cluster",

   "primitive_format": "inline-tag-v0.4",

   "agents_file": "agents.md",

   "agents": ["claude-code"],

   "capability_files": ["operating-principles.md", "copy-voice.md"],

   "signed": false,

   "signature": ""

   }
   signed and signature start empty; mm sign writes them in-place during Phase 3f.
3. Shows you the resulting file tree (paths + line counts). Asks for confirmation.
4. Signs the pack: mm sign <pack-dir>. Then runs mm verify <pack-dir> to confirm the signature is valid. mm verify is a signature check (it expects a signed pack); pre-sign structural validation happens implicitly during assembly.

Phase 4 — Draft binding

Every published pack needs a server-side draft row, and the web is the only place that creates drafts. The skill never creates a draft.

1. If --pack-id <uuid> was passed: validate it is a valid UUID (8-4-4-4-12 hex shape). On valid input, confirm to you — “Publishing into draft <uuid>.” — and proceed to Phase 5.
2. If --pack-id was NOT passed: stop and redirect. Print this exact message:

Wait for you to paste a UUID. Validate the shape. On valid input, continue to Phase 5. On invalid input, ask again — never retry silently.

Never run mm create. Never call POST /api/packs/create. Drafts are created on the web only.

Phase 5 — Publish

1. From the pack directory, run mm publish --pack-id <uuid>.
2. The CLI signs the pack with your local key, uploads files, and flips the row to published.
3. On success, the skill prints the live URL https://memorymarket.co/pack/<slug> and the marketing reminder:

Error handling

• MANIFEST_NOT_FOUND / FILE_NOT_FOUND — wrong directory; the skill asks you to cd into the assembled pack dir and retries.
• NO_SIGNING_KEY — re-runs Phase 1 step 2 (key registration). Never loops silently — it tells you what happened.
• INVALID_SIGNATURE / PACK_VALIDATION_FAILED— surfaces the API's error details verbatim and asks how to proceed.

Banned behaviors

• Never run mm create. Drafts are created on the web only. If --pack-id is missing, redirect — never silently create a draft.
• Never run mm publish before you have confirmed the assembled pack files in Phase 3.
• Never skip the auth gate in Phase 1, even if you “look logged in.”
• Never invent memory content, capability files, or manifest fields not derived from confirmed source paths.
• Never paste long shell pipelines. The skill runs one CLI subcommand at a time so you can see and trust each step.
• Never skip the price-confirm phase — the skill must ask you for a price before invoking mm publish.
• Never skip the post-publish edit reminder — tell the creator where they can change marketing copy and price.

Out of scope

• Setting screenshots or post-publish marketing copy edits. Those happen at https://memorymarket.co/sell/<uuid>(Marketing step). Price IS in scope here — it's set via --price when you run mm publish.
• Stripe Connect onboarding for paid packs. The web UI handles that.
• Editing or refactoring source memory files. Sources are read-only inputs.

Related

• /mm-add — buyer skill
• Publish a Pack — full creator walkthrough
• MM-AMP Integration — raw CLI reference