meshy-generate — Generate a 3D asset

One asset per run. Never in batch. Every generation call costs 20-30 real credits.

Input

• <slug> (required): kebab-case, no dates or task_ids. It is the folder name. No argument → show usage and stop.
• --from-image <path> (optional): uses image_to_3d instead of text_to_3d. With several images, multi_image_to_3d.
• --spec <path> (optional): an asset-spec spec the prompt comes from.

---

Phase 1: Dedupe — before spending

grep -rl "<term>" assets/3d/*/meshy.json 2>/dev/null

Search by slug and by the prompt's keywords. If there is an equivalent asset, stop: show it to the user, with its credits_total, and ask whether they want to reuse it, derive a new target with /meshy-derive, or generate anyway.

If assets/3d/<slug>/ already exists, don't overwrite it. Stop and ask.

---

Phase 2: Prompt

• With --spec: the prompt comes from the spec, verbatim.
• Without a spec: if an art bible exists (design/art/art-bible.md or equivalent), derive the prompt from it — style, palette, level of detail. If it doesn't exist, draft it with the user.
• Never invent a style that contradicts the art bible.

Show the final prompt and ask for approval. The approved prompt goes into the ledger verbatim: that is what makes the next session's dedupe work.

---

Phase 3: Credit gate

Before the first call that spends:

mcp__meshy__check_balance

Present to the user, in a single message:

Operation	text_to_3d preview
Estimated cost	20-30 credits (preview + refine)
Current balance	(from check_balance)

Wait for explicit confirmation. Don't generate on a vague "go ahead" about something else — the confirmation has to be about the cost shown.

---

Phase 4: Preview

Call mcp__meshy__text_to_3d (or image_to_3d / multi_image_to_3d) in preview mode. Poll with mcp__meshy__get_task_status — it is free, there is no reason to guess.

Download the preview with mcp__meshy__download_model and show it to the user. Record the task_id and the preview's credits in the ledger at this point already: if the refine is canceled, the preview spend still happened and has to be recorded.

---

Phase 5: Refine — only with the preview approved

If the preview isn't convincing: don't refine. Options, in order of cost:

1. Adjust the prompt and preview again.
2. retexture (10) if the geometry works and the problem is the texture.
3. Discard.

With the preview approved, call mcp__meshy__text_to_3d_refine. The resulting GLB is the master and is immutable — nothing edits it afterwards.

---

Phase 6: Write to disk

assets/3d/<slug>/
  meshy.json
  master.glb
  thumb.png

Show the contents of meshy.json and ask for approval before writing. It is written in this same turn, with the verbatim prompt, ai_model, one entry per task with its credits, and credits_total. The exact format is in the rule.

Check that master.glb is covered by Git LFS (git check-attr filter assets/3d/<slug>/master.glb). If it isn't, warn — committing a GLB without LFS pollutes the repo in a way that is hard to revert.

---

Phase 7: Summary

Report:

• Slug, the prompt used and where it came from (spec / art bible / drafted).
• Task ids and real credits spent (run check_balance again and report the delta against Phase 3).
• Files written.
• LFS status.

Verdict:

• COMPLETE — master on disk, ledger written, LFS ok.
• PARTIAL — credits were spent on the preview but there was no refine. The ledger has to reflect it anyway. State how much was spent.
• ABORTED — nothing was spent. State why.

Next Steps

• /meshy-derive <slug> for the targets (FBX, web, blend) — the master is not touched
• /meshy-print <slug> if it is going to print
• /asset-audit to validate against the project's standards

---

Full credit table, spending rules, ledger format and axis/scale conventions: .claude/rules/meshy-assets.md and .claude/rules/meshy-files.md. Read them before Phase 3.