メニュー
Audit a High Level Design document for template compliance, source integrity, ADR alignment, architectural completeness, and context fidelity. Walks through each issue interactively with resolution options including parallel code + web research.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Interface with Blender (3D modeling, animation, rendering) through its Python API (bpy). Use whenever the user mentions Blender, .blend files, bpy, bmesh, geometry nodes, shader nodes, Cycles, EEVEE, Workbench, scene/object/material manipulation, mesh editing, armatures/rigging, keyframe animation, F-curves, drivers, NLA, importing/exporting 3D formats (GLTF/glb, FBX, OBJ, USD, STL, Alembic, Collada), batch rendering, headless render farms, CLI invocation (`blender --background --python`), authoring add-ons, custom operators/panels/menus, modal operators, PropertyGroups, Geometry Nodes trees, baking, simulation (cloth/fluid/rigid body), grease pencil, compositor nodes, or hooking Blender up to MCP / live AI control. Trigger even on vague asks like 'script Blender', 'automate Blender', 'render this from a script', or any 3D-graphics task where Blender is implied. Also trigger when the user asks to install bpy as a pip module or run Blender headless on a server.
Diagnose and fix Unraid (Lime Technology NAS OS) issues — array start/stop hangs, parity errors, disabled or emulated disks (red ball/red X), New Config and trust-parity, Docker macvlan/ipvlan call traces and custom networks, VM problems (KVM, libvirt, QEMU), GPU/USB/PCIe passthrough (IOMMU groups, VFIO, ACS override), BTRFS scrub/balance and ZFS pool degraded/corruption, mover and share allocation policies (high-water/most-free/fill-up), user shares vs disk shares, /mnt/user vs /mnt/cache vs /mnt/disk*, SMB/NFS permissions and newperms, Community Applications (CA) plugin and .plg install, /boot/config persistence, USB boot and GUID/license issues, syslog and diagnostics bundle interpretation, Fix Common Problems plugin, and version-aware release-notes lookup against docs.unraid.net. Use this whenever the user mentions Unraid, unRAID, Lime Tech NAS, or describes any symptom on an Unraid box — array won't stop, won't start, disk dropped, parity sync errors, container has no network, VM black screen, cache pool
Diagnose and fix Linux desktop issues — graphics drivers (Nvidia, AMD/Mesa, Intel), display servers (Wayland, X11, XWayland), audio (PipeWire, PulseAudio, Bluetooth), boot failures, systemd unit errors, suspend/hibernate, NetworkManager/Wi-Fi, kernel module problems, Flatpak/Snap sandbox breakage, and general desktop environment failures. Use this whenever the user reports ANY Linux desktop symptom — black screen, login loop, no audio, screen tearing, frozen cursor, GPU hangs, crashes after kernel update, broken external monitor, suspend doesn't work, app won't launch under Wayland, screen recording missing, Bluetooth pairing fails, or anything else where their Linux machine isn't behaving. Trigger even when the user just says 'my Linux is broken' or describes a vague desktop problem without naming the subsystem. Also use for proactive triage of dmesg/journalctl logs the user pastes, even if no explicit fix is requested.
Find every flaw in the overall approach and strategy of the code. Challenges whether the right problem is being solved, whether the algorithm is appropriate, whether assumptions are validated, and whether better alternatives exist. Produces severity-ranked issues with file:line references. Use when you want a dedicated approach review of code changes, files, or directories.
Find every flaw in the structural soundness, component design, and system organization of the code. Challenges whether boundaries are clean, coupling is minimized, dependencies flow correctly, and the module structure scales. Produces severity-ranked issues with file:line references. Use when you want a dedicated architecture review of code changes, files, or directories.
Find every violation of language and framework conventions, established patterns, and anti-patterns in the code. Challenges whether idioms are followed, APIs are designed correctly, and deprecated patterns are avoided. Produces severity-ranked issues with file:line references. Use when you want a dedicated best-practices review of code changes, files, or directories.
| name | audit-hld |
| description | Audit a High Level Design document for template compliance, source integrity, ADR alignment, architectural completeness, and context fidelity. Walks through each issue interactively with resolution options including parallel code + web research. |
| version | 2.0.0 |
| context | fork |
| argument-hint | [hld-path] [--context path] [--adr path] |
| allowed-tools | Read, Glob, Grep, Write, Edit, Task, AskUserQuestion, WebSearch, WebFetch |
| model | opus |
Audit a High Level Design document for quality, completeness, and correctness. Each issue is presented to the user one at a time with multiple resolution options, always including a research option that forks parallel code + web research agents.
$ARGUMENTS — path to the HLD document, and optionally:
--context <path> — context file from the gather phase (enables context fidelity checks)--adr <path> — predecessor ADR (enables ADR alignment checks)Extract from $ARGUMENTS:
--context <path> (optional)--adr <path> (optional)--context provided, read the context file.--adr provided, read the ADR.${CLAUDE_SKILL_DIR}/references/checklist.md.Run every check from the checklist. Build a prioritized issue queue:
## HLD Audit Summary: [HLD Title]
**Document:** [path]
**Context file:** [path or "not provided"]
**ADR:** [path or "not provided"]
Found **N issues** and **M open questions**:
- 🔴 CRITICAL: [count]
- 🟡 WARNING: [count]
- 🔵 INFO: [count]
- ❓ Open Questions: [count]
Starting sequential resolution...
If zero issues, skip to Phase 5 with PASS verdict.
For each issue, present using AskUserQuestion with:
When "🔍 Research code & web" is chosen:
Dispatch two parallel Tasks:
Task 1 — Code Research:
"Search the codebase to answer: [specific question about the HLD issue].
Focus on: [relevant files, patterns, interfaces].
Return findings with file:line citations."
Task 2 — Web Research:
"Search online to answer: [specific question about the HLD issue].
Focus on: [best practices, patterns, known issues].
Return findings with URLs."
Synthesize findings, then re-present the issue with refined options.
When a fix is chosen: Apply edit, show diff, log resolution. When Skip is chosen: Log as skipped, move to next.
# Audit Report: [HLD Title]
**Document:** [hld-path]
**Context File:** [context-path or "not provided"]
**ADR:** [adr-path or "not provided"]
**Date:** [today's date]
**Verdict:** [PASS | PASS WITH WARNINGS | FAIL]
## Summary
[2-3 sentence overview]
## Issues Resolved
| # | Severity | Category | Issue | Resolution | Research Used? |
|---|----------|----------|-------|------------|----------------|
## Issues Skipped
| # | Severity | Category | Issue | Reason Skipped |
|---|----------|----------|-------|----------------|
## Open Questions Resolved
| # | Question | Answer | Source |
|---|----------|--------|--------|
## Check Results
### Template Compliance
| Check | Status | Details |
|-------|--------|---------|
### Source Integrity
| Check | Status | Details |
|-------|--------|---------|
### ADR Alignment
| Check | Status | Details |
|-------|--------|---------|
### Architectural Completeness
| Check | Status | Details |
|-------|--------|---------|
### Context Fidelity
| Check | Status | Details |
|-------|--------|---------|
Verdict logic:
Save to same directory as HLD: docs/hld/<name>-AUDIT.md.
Report the audit verdict and file path.