Roblox Build Retrospective

Roblox Build Retrospective

Closes the learning loop for roblox-object-builder. Extracts structured knowledge from build sessions and writes validated findings into the builder skill's docs.

Trigger Conditions

Automatic — trigger without user request if ANY apply

• Same operation required >= 2 MCP retries in one session
• Any pcall returned false
• User expressed negative feedback: "違う", "おかしい", "wrong", "not right"
• Build produced >= 100 parts and completed

Manual — always trigger when user says

"振り返り", "retrospective", "what went wrong", "analyze the result"

Target Files

roblox-object-builder/
├── SKILL.md              ← Anti-patterns, Roblox-specific gotchas summary
├── docs/build-log.md     ← Specific incidents (PRIMARY output target)
├── docs/gotchas.md       ← Universal technical pitfalls (promoted from build-log)
├── docs/design-wisdom.md ← Roblox-specific rendering/material behavior
└── docs/csg-guide.md     ← CSG decision flow and safe practices

Path: .claude/skills/roblox-object-builder/

Output flow

Finding → build-log.md (always)
                ↓
        Abstraction Check passes + HIGH confidence?
                ↓ YES
        gotchas.md or SKILL.md (promote abstract rule only)

Specific incidents, object-specific details, and orientation tables stay in build-log.md permanently. Only abstract rules that pass the Abstraction Check get promoted to the AI-facing docs.

Retrospective Process

Phase 1: Context Collection

First, scan session-log.md for Open questions relevant to this build type. Note any matches for evaluation in Phase 3.

Then reconstruct from conversation:

1. Goal: Object type, scale, style
2. MCP calls: Summary of each run_code invocation
3. Corrections & retries: Where the agent backtracked (richest signal)
4. Final state: Complete? User satisfied?

Explicit request → present summary, ask for confirmation. Proactive trigger → concise summary, ask only targeted questions.

Phase 2: User Feedback

Short checklist (proactive default):

1. Positioning/scale issues?
2. Color/material/neon issues?
3. CSG artifacts or broken geometry?

Explicit request → add: 4. Scale feel? 5. Anything else?

If user provides screenshot, analyze carefully for visual issues.

Phase 3: Structured Analysis

Per finding:

### [Title] (DOMAIN)
Signal: RED / GREEN
Symptom: [observed]
Root cause: [why]
Fix: [what to do — process-level, not object-specific]
Prevention: [rule to add — must pass Abstraction Check below]
Confidence: HIGH / MEDIUM / LOW

Abstraction Check (MANDATORY for every finding before writing to skill)

Ask: "Does this finding still make sense as a rule if I remove all specific object names?"

• YES → OK to write. It is general process knowledge.
• NO → Raise the abstraction level. Describe the structural condition, not the object.

Example:

BAD:  "Use SubtractAsync to carve a basin into a kitchen sink"
      → Remove "kitchen sink" and the rule collapses. It is object-dependent.

GOOD: "Any surface that is concave in the real world should be carved with
       SubtractAsync. Placing a block on top of a flat surface does not read
       as a recess."
      → Works regardless of object — sink, bathtub, sofa seat, fountain basin.

Only HIGH/MEDIUM → write to skill. LOW → session log only. Evaluate any Open questions from Phase 1 against this session's evidence.

Phase 4: Knowledge Integration

Step 1: Write ALL findings to build-log.md

Every finding goes to docs/build-log.md, regardless of confidence level. Include: date, build subject, incident, category tag, what was learned. Object-specific details (orientation tables, exact dimensions, code snippets for that specific object) belong here permanently.

Step 2: Promote abstract rules (HIGH confidence only)

For each HIGH confidence finding, apply the Abstraction Check:

"Would an AI reading this rule benefit from it when building a completely different object?"

YES → promote to gotchas.md or SKILL.md. NO → stays in build-log.md only.

When promoting:

1. Read target file, check for duplicates
2. Duplicate exists → update/strengthen. New → add in appropriate section.
3. Write using Edit tool, show diff to user

What goes into AI-facing docs (gotchas.md, SKILL.md):

• Structural conditions: under what circumstances does the problem occur?
• Root mechanisms: why does it fail?
• Process-level remedies: what prevents or fixes it?

What stays in build-log.md only:

• Object-specific code snippets ("sink CSG code", "stove recess code")
• Rules that depend on object names ("Use CSG for kitchen sinks")
• Hardcoded dimensions ("cutter should be 3.5 x 2 x 4.2")
• Prescriptive process rules ("always use 3-phase", "always use data tables")

Guard against over-constraining the builder skill:

The builder SKILL.md uses anti-patterns ("don't do X") rather than prescriptions ("do X"). When writing a new rule, frame it as what to AVOID. If it can only be stated as "always do X", question whether the AI needs that instruction or can figure it out on its own. If the AI can figure it out → don't write it.

Phase 5: Session Log

Append to .claude/skills/roblox-build-retrospective/knowledge/session-log.md:

## [Date] - [Object Name]
Goal: [what] | Outcome: SUCCESS / PARTIAL / FAILURE
Key findings: [finding - domain - signal - confidence]
Skill updates: [file: what changed]
Open question resolution: [VALIDATED / REFUTED / STILL OPEN + evidence]
Open questions: [LOW confidence items for future validation]

Signal Detection

Red Flags (knowledge gaps)

• Multiple MCP retries → missing/incorrect guidance
• Agent self-correcting → approach doesn't work
• User "that's not right" → verification gap
• pcall false → unhandled edge case
• Parts at Y=0 or overlapping → positioning formula error

Green Flags (preserve patterns)

• Clean first-try success → document approach
• Efficient part count → note decomposition
• User praise → capture design choices

Multi-Session Meta-Analysis (after 3+ sessions)

1. Frequency: Which domains dominate? Focus skill improvements there.
2. Recurring LOW: Same hypothesis 2+ times → elevate to MEDIUM, add to skill.
3. Contradictions: Flag for user review.
4. Bloat check: docs > 300 lines → consider reorganizing or extracting.
5. Abstraction audit: Scan all skill files for object-specific snippets or hardcoded dimensions that have crept in. Rewrite to process-level or delete.
6. Anti-pattern audit: Check if any rule in SKILL.md is prescriptive ("do X") rather than avoidance-oriented ("don't do X"). Prescriptive rules constrain AI reasoning — flag for user review or reframe as anti-patterns.
7. Staleness: last-validated > 3 sessions old → report to user (do NOT auto-delete).

Principles

• Write anti-patterns, not prescriptions: Skills should tell AI what to AVOID, not what to DO. Teaching "how to build" makes the AI follow steps instead of think. Teaching "what fails and why" preserves AI judgment while preventing known bugs. When in doubt, don't write it.
• Don't over-generalize: "CSG fails when overlap < 2 studs" not "CSG always fails"
• Don't over-specify: "Use CSG for kitchen sinks" only fires in kitchens. "Use CSG for concave surfaces" fires everywhere. Always prefer the latter.
• Preserve what works: Document WHY smooth builds went smoothly
• User drives priority: Focus on what the user flags
• Confidence gating: LOW speculation degrades skill quality — session log holds hypotheses