Hitem3D

Use this skill to turn images into downloadable 3D models via scripts/hitem3d.sh.

Use the script

Resolve all relative paths against this skill directory.

• - Script: INLINECODE1
• API reference: INLINECODE2
• Product guide / positioning: INLINECODE3
• ClawHub-style description: INLINECODE4
• Release checklist: INLINECODE5
• Live validation notes: INLINECODE6

If the user asks for a normal one-off generation, prefer the full pipeline:

1. 1. submit task
2. wait for completion
3. download result
4. report output path, format, model, and estimated credit cost

Do not leave the user with only a task ID unless they explicitly want async handling.

Operator standard

Act like a production 3D operator, not a thin API wrapper.

• - Infer likely intent from the request instead of asking obvious questions.
• Distinguish single-image, portrait, multi-view, and batch every time.
• Prefer finishing the workflow over handing back raw JSON.
• Surface cost, output path, format, and caveats in plain language.
• Stop early on ambiguous expensive jobs, invalid inputs, or unsafe assumptions.
• Never echo secrets, bearer tokens, or raw Authorization headers.

Required setup

Runtime requirements

• - Binaries: curl, python3, INLINECODE9
• Required env vars: HITEM3D_AK, INLINECODE11
• Optional env var: INLINECODE12

The user must provide credentials through environment variables:

CODEBLOCK0

If credentials are missing, stop and tell the user exactly which variables are not set.

Default operating policy

When the user says “turn this image into 3D” without more detail, use these defaults:

• - model: INLINECODE13
• request type: 3 (geometry + texture)
• resolution: INLINECODE15
• format: 2 (GLB)
• face count: unset

Why: this is the best general-purpose quality/output tradeoff.

Choose the right mode

Single-image general object

Use when the input is one product shot, concept image, illustration, object render, furniture photo, toy, packaging, or other non-portrait asset.

Default command pattern:

CODEBLOCK1

Portrait / bust generation

Use portrait models when the user explicitly wants a person, face, bust, avatar head, or half-body likeness.

Preferred defaults:

• - model: INLINECODE17
• resolution: INLINECODE18
• format: GLB unless user asks otherwise

Command pattern:

CODEBLOCK2

3D printing

If the user mentions 3D printing, printer, slicer, or STL, prefer:

• - format: 3 (STL)
• request type: 1 for mesh-only when texture is irrelevant
• optionally set higher face count for smoother geometry

Command pattern:

CODEBLOCK3

AR / Apple preview

If the user mentions Apple AR, Quick Look, Vision Pro, AR preview, or iPhone preview, prefer format 5 (USDZ).

Fast / cheap draft

If the user wants a rough draft, cheap test, or low-cost preview, prefer:

• - model: INLINECODE22
• resolution: INLINECODE23
• type: 1 or 3 depending on whether texture matters

Multi-view reconstruction

Use multi-view when the user provides 2-4 images of the same subject from different angles.

Use:

• - generate-multi or INLINECODE27
• pass a bitmap describing which views are present

Bitmap order is:

• - front
• back
• left
• right

Examples:

• - front + left = INLINECODE28
• front + back + left + right = INLINECODE29
• front + right = INLINECODE30

Do not use multi-view mode for unrelated images.

Batch processing

Use batch mode only when the user wants to process many separate images independently. Do not confuse batch with multi-view.

Batch means:

• - one task per image
• submit all
• wait/download each
• return a concise summary of successes/failures and output paths

Cost policy

Before batch jobs or obviously expensive jobs, tell the user the expected cost and ask for confirmation.

Must confirm before proceeding when either is true:

• - batch job with more than 5 images
• likely cost exceeds 100 credits total
• the request implies repeated retries, alternative variants, or multiple output formats that would multiply spend

For one-off standard jobs, confirmation is not required unless the user has asked to be cost-cautious.

Reference pricing in references/api.md.

Practical heuristics:

• - cheapest draft: hitem3dv1.5 + 512 + mesh-only = 5 credits
• default general result: hitem3dv2.0 + 1536 + textured = 50 credits
• premium portrait / 1536pro textured = about 70 credits

When estimating cost, say whether the number is per task or total batch cost.

Output policy

Prefer saving downloads into a user-relevant folder under the current workspace unless the user provides a destination.

Suggested defaults:

• - one-off tasks: INLINECODE36
• batch tasks: INLINECODE37

When reporting completion, include:

• - task ID
• final status
• saved file path
• output format
• model used
• resolution
• estimated credit cost

If the API returns a preview/cover URL, include it when useful.

Common commands

Check balance

CODEBLOCK4

Submit only

CODEBLOCK5

Full one-shot pipeline

CODEBLOCK6

For long jobs, explicitly raise the timeout instead of assuming 10 minutes is enough:

CODEBLOCK7

Full one-shot pipeline with custom options

CODEBLOCK8

Multi-view submit

CODEBLOCK9

Multi-view full pipeline

CODEBLOCK10

Long-running jobs

If a job is likely to take a while, set explicit polling and timeout values:

CODEBLOCK11

Wait for an existing task and download

CODEBLOCK12

For slow tasks:

CODEBLOCK13

Batch folder processing

CODEBLOCK14

Decision rules

If the user does not specify format:

• - use GLB by default
• use STL for 3D printing intent
• use USDZ for Apple AR intent

If the user does not specify model:

• - use scene-portraitv2.1 for portraits
• otherwise use INLINECODE39

If the user does not specify resolution:

• - use 1536 for general work
• use 1536pro for portrait premium output
• use 512 for cheap drafts

If the user says “highest quality”:

• - general object: hitem3dv2.0 + INLINECODE44
• portrait: scene-portraitv2.1 + INLINECODE46
• mention expected credit usage before running

Failure handling

Missing credentials

Stop and ask the user to set HITEM3D_AK and HITEM3D_SK.

Bad input images

Reject unsupported formats and files larger than 20MB before calling the API.

Auth failure

Tell the user credentials are invalid or expired. Do not keep retrying.

Generation failure

If API error is 50010001, tell the user the image likely could not be parsed and credits should be refunded automatically.

Token expiry

The script auto-fetches tokens. Retry once if a token-related request fails.

URL expiry

If the returned download URL has expired, re-query task status to fetch a fresh URL before declaring failure.

Multi-view mismatch

If the user labels images as multi-view but the set is missing the front view, warn that Hitem3D expects the front view and that results may fail or degrade.

Security rules

• - Never print or store AK/SK in user-visible output.
• Treat callback URLs as advanced usage; do not invent or auto-fill webhooks.
• Do not claim full production validation unless live API runs have been completed with real credentials.
• If credentials are unavailable, say the skill is design-validated but not fully live-validated.

Download URL expiry

If the result URL expired, re-query task status to obtain a fresh result URL if available, then download again.

Unsupported combination

Reject obviously bad combos before calling the API, especially:

• - request type 2 with INLINECODE51
• multi-view without a valid 4-bit view bitmap
• face count outside 100000-2000000

Notes for good judgment

• - A user saying “folder of 30 product renders” means batch, not multi-view.
• A user sending front/back/left/right of the same object means multi-view, not batch.
• For printing intent, geometry matters more than texture.
• For showcase/demo intent, textured GLB is usually the right default.
• Prefer completing the full task over dumping raw API responses.
• Keep user-visible replies concise and outcome-focused.