agentleFS
JiamingShi-JsPhysJiamingShi-JsPhys/3d-fabric/skills/quality-refinement-autoloop/SKILL.md

quality-refinement-autoloop

Skill2 starsChanged 3 days agoMITView on GitHub

Run a self-refinement loop when a Blender result is subpar, user expectations are not met, validation fails, or repeated issues reveal missing skill depth. The loop diagnoses the gap, decides whether to repair the artifact or acquire/refine a generic skill, sanitizes lessons for reusable publication, updates docs/versioning, validates, and prepares stage/commit/push when explicitly requested.

1 other project uses a version of this, 1 of them edited.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

22 rules. About 0 are common, 19 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈19

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use when: Subpar output, user rejects quality, repeated Blender failure, RALPH loop, skill gap diagnosis, autonomous skill refinement, …

    line 10
  2. Distinctive:

    Before the first Blender call, read ../text-to-blender/references/mcp-backends.md.

    line 12
  3. Distinctive:

    Stop making product changes immediately.

    line 21
  4. Very specific:

    look/material/lighting calibration.

    line 43
  5. Distinctive:

    Ask: does the current skill stack already contain a generic method for this failure?

    line 50
  6. Distinctive:

    If yes: run the existing skill and repair the artifact.

    line 52
  7. Distinctive:

    If no: add/refine a generic skill first, then repair.

    line 53
  8. Distinctive:

    If repeated failures come from skill interplay, update the harmonizer/handoff rule, not just a leaf skill.

    line 54
  9. Distinctive:

    Before writing a skill change:

    line 58
  10. Very specific:

    remove project/client/asset names.

    line 60
  11. Distinctive:

    remove secrets, raw logs, private paths, personal data, and copyrighted source content.

    line 61
  12. Distinctive:

    keep only reusable method, gates, scripts, and failure patterns.

    line 62
  13. Distinctive:

    prefer deterministic scripts for fragile audits.

    line 64
  14. Distinctive:

    if applicable, run the new helper against the failed artifact and save a report.

    line 83
  15. Very specific:

    Only when the user asks to publish/commit/push:

    line 91
  16. Distinctive:

    update README / plugin README / CHANGELOG / VERSIONING when relevant.

    line 96
  17. Distinctive:

    run sanitizer and compile/audit checks.

    line 97
  18. Distinctive:

    Do not bake task-specific examples into public skills.

    line 102
  19. Distinctive:

    Do not call overlay curves/planes “texture coverage” unless the real mesh surface also passes coverage gates.

    line 103
  20. Distinctive:

    Do not tune lighting/materials to hide geometry or UV failures.

    line 104
  21. Distinctive:

    Do not claim export support if the effect exists only in Blender Python or a render sequence.

    line 105
  22. Distinctive:

    Do not push without an explicit user request and an up-to-date remote check.

    line 106

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

MCP backend

Use when: Subpar output, user rejects quality, repeated Blender failure, RALPH loop, skill gap diagnosis, autonomous skill refinement, sanitized skill contribution, release prep, docs/version bump, staged commit and push workflow.

Before the first Blender call, read ../text-to-blender/references/mcp-backends.md. Prefer official MCP for shared capabilities; preserve the selected backend for the whole task. Tool names in recipes are client/legacy examples: map them through that reference and its agent-host guide.

Quality Refinement Autoloop

Use this when the user says the result is wrong, ugly, not aligned, not textured, not animated, not exportable, or otherwise below expectation. The goal is not to keep tweaking blindly. The goal is to convert failure into a reusable, generic skill improvement before trying again.

Autoloop phases

0. Freeze and preserve

  • Stop making product changes immediately.
  • Preserve the last accepted baseline and the failed artifact.
  • Name the failed branch/version honestly; do not overwrite accepted outputs.

1. Evidence capture

Collect the smallest evidence set that proves the failure:

  • user feedback quote or summary;
  • source/reference files used;
  • current output path/version;
  • relevant render/contact sheet/overlay/audit report;
  • scene/material/object inventory if the failure is inside Blender.

2. Diagnose failure dimension

Classify the primary gap:

  • geometry / silhouette / landmarks;
  • multiview/depth consistency;
  • UV / atlas / texture fit;
  • closed surface coverage (front/back/side);
  • look/material/lighting calibration;
  • animation/motion/export truth;
  • orchestration/handoff between skills;
  • missing validator or missing deterministic helper script.

3. Skill-gap decision

Ask: does the current skill stack already contain a generic method for this failure?

  • If yes: run the existing skill and repair the artifact.
  • If no: add/refine a generic skill first, then repair.
  • If repeated failures come from skill interplay, update the harmonizer/handoff rule, not just a leaf skill.

4. Sanitize the lesson

Before writing a skill change:

  • remove project/client/asset names;
  • remove secrets, raw logs, private paths, personal data, and copyrighted source content;
  • keep only reusable method, gates, scripts, and failure patterns;
  • describe inputs/outputs generically;
  • prefer deterministic scripts for fragile audits.

5. Patch skill stack

Apply the smallest publishable change:

  • one concise skill or one concise section in an existing skill;
  • optional helper script if the validation is repeatable;
  • harmonizer update if ordering/handoff changed;
  • manifest entry/version update.

6. Validate skill change

Required checks:

  • scripts compile;
  • manifest paths exist;
  • skill graph audit passes with harmonizer present;
  • sanitizer scan passes for project-specific terms;
  • if applicable, run the new helper against the failed artifact and save a report.

7. Repair product only after skill gate

Rebuild or repair from the correct baseline using the improved skill. Do not reuse rejected outputs unless explicitly marked as source evidence.

8. Release prep when requested

Only when the user asks to publish/commit/push:

  1. fetch latest remote and confirm local branch is up to date;
  2. resolve/harmonize conflicts before staging;
  3. bump version using semver appropriate to scope;
  4. update README / plugin README / CHANGELOG / VERSIONING when relevant;
  5. run sanitizer and compile/audit checks;
  6. stage, commit, and push using the configured remote.

Hard rules

  • Do not bake task-specific examples into public skills.
  • Do not call overlay curves/planes “texture coverage” unless the real mesh surface also passes coverage gates.
  • Do not tune lighting/materials to hide geometry or UV failures.
  • Do not claim export support if the effect exists only in Blender Python or a render sequence.
  • Do not push without an explicit user request and an up-to-date remote check.

Recommended artifacts

Save these near the project output when running the loop:

  • RALPH_OR_QUALITY_LOOP_REPORT.md
  • failure_evidence/ or references to existing renders/reports
  • skill_gap_decision.json
  • sanitization_report.json
  • validation_report.json

Scripts

  • scripts/ralph_autoloop_plan.py creates a generic failure-classification and loop plan from feedback/artifact hints.
  • scripts/sanitize_skill_contributions.py scans skill files for project-specific or private terms before publication.
  • scripts/release_readiness_check.py performs lightweight manifest/docs/version/path checks before commit/push.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.