Glb Mobile Performance Budget

glb-mobile-performance-budget

Purpose

Translate "make it work on mobile" intent into a concrete numeric budget table that downstream skills (scene planner, quality checker, render-export-evidence, glb-web-handoff) can validate against.

Quick start

• pick the device tier (low-end Android, mid-mobile, high-mobile, desktop-fallback)
• read the budget row for that tier
• pin the budget into the scene plan before mutation
• treat any later violation as a blocker, not a warning

When to use

• before scene plan is locked
• before any GLB export attempt
• when user says "for mobile" / "for the web" / "phone" / "low-end devices"
• when designing a scene that must run alongside other web content

When not to use

• desktop-only renders without web handoff
• offline pre-rendered video (no real-time budget needed)
• replacing engineering load testing (this skill sets a planning budget, not a real-world benchmark)

Trigger phrases

• "what's a good polycount for mobile"
• "how big can the GLB be"
• "will this load on iPhone"
• "what should I aim for on a phone"

Prerequisites / readiness

• intended device tier known (or user willing to pick one)
• expected scene scope known (single product / multi-product / environment)
• GLB target context known (hero card / e-commerce viewer / AR / configurator)

Input schema

Required inputs

| Input | Why it is required | |---|---| | Device tier | Different tiers permit very different budgets. | | Scene scope | Single mesh and full environment have different budget semantics. | | Target context | AR viewer is stricter than a static hero card. |

Optional inputs

| Input | Use | |---|---| | Existing competitor GLB references | Calibrates the budget against shipped peers. | | Network constraints | 3G/4G/5G affects file-size tolerance. | | Animation complexity | Skeletal + morph budgets differ. |

Assumptions to confirm

• Budget is a planning ceiling, not a guaranteed runtime performance result.
• Field-of-view, post-processing, and shader complexity affect real GPU cost beyond polycount.
• Mobile GPU memory varies wildly; budgets here favor the lower end of the tier.

Output schema

Primary output

Numeric budget table for the chosen tier covering: triangles, vertices, GLB file size, texture pixel budget, distinct material count, draw calls, animation channel count, bones, morph targets.

Secondary output

• a "violations would block handoff" rule list
• a fallback degradation plan (what to drop first if a violation appears)
• explicit assumptions about device tier, network, and FOV

Evidence / caveat output

Runtime status: Not Run | Attempted | Produced | Verified | Failed | Blocked / Not Run
Artifact status: Not Run | Not Produced | Produced | Verified | Failed
Evidence used: 
Limitations: 

Required laws

• ../../laws/evidence-before-done.md
• ../../laws/non-blender-user-language.md
• ../../laws/no-arbitrary-python-interface.md

Official runtime boundary

BlendOps runtime guidance uses 2 MCP execution paths plus a CLI fallback appendix (see ../../docs/runtime-stack-strategy.md):

1. Path 1 — Official Blender Lab MCP (Lab add-on + Lab server in Blender 5.1+, hosted from Anthropic Connector or any other MCP client manually).
2. Path 2 — Community ahujasid/blender-mcp (Blender 3.0+, third-party prior art).
3. CLI fallback (appendix) — blender --background --python .... Documented upstream as a first-class Blender CLI surface; no in-repo evidence file yet.

This skill is a planning skill. It sets numeric budgets but never installs Blender, runs Blender, or claims that an exported GLB satisfies the budget without measured evidence.

Operating procedure

1. Confirm device tier with the user; default to "mid-mobile" if not stated.
2. Confirm scene scope and target context.
3. Look up the budget row from the references budget tables.
4. Adjust ±10% if user constraints justify it; record adjustments.
5. Pin the budget table into the scene plan output.
6. List violations that would block GLB handoff if measured later.
7. Provide a degradation plan (what to drop first).
8. Hand off to the scene planner or quality checker with the pinned budget.

Decision tree

Tier unknown?
  → Default to "mid-mobile" + warn
Scene scope unknown?
  → Ask once, fall back to "single product"
Multiple competing budgets (AR + hero card)?
  → Pick the strictest, document the override
Existing GLB exceeds budget?
  → Hand off to quality checker for what to drop

Playbooks

Playbook A: Single product hero on mobile

Use mid-mobile baseline. Lock GLB ≤ 4 MB compressed (Draco), ≤ 50K triangles, ≤ 4 textures @ 1K. AR add-on: drop to 25K tris and 2K textures.

Playbook B: Configurator on mid-mobile

Allow 80K triangles (more material variants), but cap GLB at 6 MB and limit draw calls to ≤ 30.

Playbook C: Low-end Android

Hard cap: 25K triangles, 2 MB GLB, 2 textures @ 512 px, no morph targets, ≤ 1 skinned mesh.

Mode handling

Text-only mode

Produce the budget table and rules. Do not claim compliance.

Runtime-ready mode

The budget table is still the planning artifact; compliance is measured by render-export-evidence + quality-checker after a real export. Never upgrade to "Verified" without measured evidence.

Blocked runtime mode

Same budget output; mark compliance unknown.

Validation checklist

• [ ] Device tier chosen explicitly
• [ ] Scene scope chosen explicitly
• [ ] Target context chosen explicitly
• [ ] Budget table pinned with concrete numbers (no "low" / "high" without numbers)
• [ ] Violations rule list present
• [ ] Degradation plan present
• [ ] Assumptions section present
• [ ] No "guaranteed performance" claim made
• [ ] Handoff to next skill named

Pass / Warn / Fail rubric

| Verdict | Criteria | |---|---| | Pass | Budget table has concrete numbers for tier, scope, and context; degradation plan present; no compliance claim. | | Warn | Budget present but tier/scope ambiguous; degradation plan partial. | | Fail | "Will be fast on mobile" claim without numeric budget; missing degradation plan; treating budget as a measured result. |

Failure handling

• If user pushes for a "yes, it'll run on iPhone X" guarantee → restate that this skill produces a planning budget, not a benchmark.
• If competing constraints exist → pick the strictest and document the override in writing.
• If user has no device data → use mid-mobile as default and write the assumption.

Troubleshooting

| Problem | Response | |---|---| | User wants a single number ("just tell me the polycount") | Give the tier-appropriate triangle cap and immediately add the texture/file-size companions. | | Budget feels too tight | Confirm scene scope; the user may have over-stated complexity. | | Budget feels too loose | Confirm device tier; user may have under-stated low-end mobile expectations. |

Best practices

• Always pair triangle budget with texture and file-size budgets — one alone is misleading.
• Reserve 20% headroom against the hard cap for runtime overhead.
• Prefer Draco-compressed GLB file size in budgets unless explicitly excluded.
• Treat AR contexts as a separate stricter tier, not "mobile but a bit smaller".

Good examples

• "Mid-mobile, single product hero, web hero card. Budget: ≤ 50K tri / ≤ 4 MB GLB (Draco) / ≤ 4 textures @ 1024 / ≤ 8 draw calls. Compliance: Not Run."

Bad examples

• "Should be fine on mobile." — no numbers, no tier, no rules.
• "GLB will be fast." — no measurement, no caveat.

User-facing response template

Device tier: 
Scope: 
Context: 
Budget:
  Triangles: 
  Vertices: 
  GLB file size (Draco): 
  Texture pixel budget total: 
  Distinct materials: 
  Draw calls: 
  Animation channels: 
  Bones: 
  Morph targets: 
Violations would block handoff: yes / no
Degradation order if violated: 
Compliance: Not Run (planning budget only)

Anti-patterns

• Picking polycount in isolation.
• Treating the budget as a benchmark.
• Promising performance without measurement.
• Mixing tiers ("kind of mobile but desktop too").

Cross-skill handoff

• Scene planning → ../product-hero-scene-planner/SKILL.md, ../blender-composition-camera-planner/SKILL.md, ../blender-lighting-material-planner/SKILL.md
• Quality validation → ../blender-scene-quality-checker/SKILL.md
• Render/export evidence → ../render-export-evidence/SKILL.md
• GLB handoff → ../glb-web-handoff/SKILL.md
• Animation budget specific → ../glb-animation-handoff/SKILL.md

Non-goals

• Run a real-world benchmark.
• Install Blender.
• Run Blender.
• Choose a render engine.
• Write web-side optimization code.

References

• references/budget-tiers.md
• references/violation-rules.md
• references/degradation-plan.md
• ../../docs/skill-system.md
• ../../docs/runtime-stack-strategy.md

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ThanhNguyxnOrg
• Source: ThanhNguyxnOrg/blendops
• License: MIT
• Homepage: https://github.com/ThanhNguyxnOrg/blendops

Install and usage instructions live in the source repository linked above.