Houdini Architect Analysis

Houdini Architect Analysis

Purpose

Pull architect-level understanding out of any procedural modeling project (Houdini-first, but the methodology is tool-agnostic). The skill is built on one core conviction: every wrangle is a classical CS algorithm in geometric disguise — once you can name the algorithm, you know its complexity, failure modes, and refactor options.

Quick start (3-step workflow)

1. Extract the project (no Houdini needed):

   python scripts/extract_hip.py path/to/file.hip
   python scripts/parse_hip.py   path/to/file_extracted  out_dir/00_topology.txt
   python scripts/extract_vex.py path/to/file_extracted  out_dir/
   

   Produces 00_topology.txt (SOP graph) + 01_all_wrangles.md (every VEX) + 02_all_node_key_params.txt.

2. Walk the project layer by layer using the 4-segment template:

   • Functional contract — what capability does this layer/subnet provide?
   • What's implemented — concrete decisions, node choices, why each
   • What's deliberately not done — pushed downstream to which layer
   • What's missing entirely + how to extend — concrete refactor sketches

3. For each wrangle, name the algorithm before explaining the code. See ALGORITHMS.md for the catalog.

Anti-patterns to avoid (hard-learned)

• ❌ Don't write an encyclopedia document. Walk layer-by-layer interactively.
• ❌ Don't explain rint vs floor — that's function manuals, not architecture.
• ❌ Don't ask the user "do you want to continue?" after each layer. Auto-continue. Only stop on truly mutually-exclusive forks.
• ❌ Don't go abstract (Conway's Law, Category Theory) at the expense of concrete code. Math/algo lens must point at specific wrangles.
• ✅ Do combine topology view + code in every layer.
• ✅ Do name the classical algorithm + complexity + failure mode for every wrangle.

Reference index

When the question matches, load the relevant file:

• METHODOLOGY.md — the layer-by-layer process + 4-segment template details
• ALGORITHMS.md — wrangle algorithm catalog (longest path / Markov / KDE / level-set / spatial join / multi-ray decision tree)
• PATTERNS.md — loop block math (feedback=fixed-point, piece=parallel-map, count=truncated), 5-stage pipeline template, cross-subnet bus
• COLLABORATION.md — 6-layer collaboration analysis (L1 cross-wrangle through L6 cross-project) with the metrics dashboard
• DEBT_CHECKLIST.md — common architectural debts to flag, with concrete refactor templates

When to write deliverables

By default the skill outputs interactive layer-by-layer analysis in the chat, not files. Only write a partN.md summary file when the user explicitly says "write it up" or "save the analysis."