Ultimate AI Starter Bundle v8.51.1

One installer that gives every AI coding app on Windows the same skills, tools and project knowledge — so a fresh chat already knows how you work and where your games, mods and tools live.

Workflow hints remain advisory. Creative work keeps its scope, tool discovery uses verified owners and bounded local roots, and detailed inventories are read when needed. Existing model choices and personal settings are preserved.

What it is not: another model, a hosted service, or a package manager for other people's tools. It is the foundation those tools plug into.

Install in three steps

1. Download …-Core.zip from the latest release and unzip it. On a slow or restricted network use Full-Offline instead.
2. Double-click START-HERE.bat. If Windows warns, choose Run; allow administrator rights if asked.
3. Wait for INSTALL COMPLETE, then close and reopen your AI apps.

Requires Windows 10/11, an internet connection and at least one supported app. The default install also adds a compiler suite (several GB); -SkipBuildTools skips it.

File	Pick it when
…-Core.zip	Normal connected install (~34 MiB). Payloads are fetched during install; same result as Full-Offline.
…-Full-Offline.zip	One file for an unreliable network (~128 MiB). Payload snapshots only — inference still needs your provider or a local model.
everything else	Checksums, build receipt, SBOM, attestation. Not needed to install.

Supported apps

Claude Code · Codex · Grok · Kimi · Hermes · Google Antigravity. Each gets the same skills and instructions in its own format; apps you do not have are skipped.

You want to	Do this
Update	Run START-HERE.bat again; it installs a newer release and keeps your settings.
Only some apps	START-HERE.bat -Providers Claude,Codex
Check the install	Run START-HERE.bat again; the last step prints PASS or the exact problem.
Verify by hand	powershell -File TOOLS/Test-Installed-State.ps1
Undo leftovers	powershell -File TOOLS/Clean-StaleState.ps1 -Apply (dry run without -Apply)

What you get

81 general skills per AI	planning, review, debugging, modding routes, publishing, research — loaded only when a task needs them
Base MCPs	context7, headroom, toolbox everywhere; code-intel in the code profile
Compilers	Go, Rust, GCC, .NET, CMake/Ninja, Xmake, MSVC and the Windows SDK
Verification	pack gate, installed-state doctor, RTK routing, evidence rules
Project state	PLAN.md/STATE.md templates plus a deterministic task controller
Free routes	local models, free tiers and local asset generators; no paid component is required

Add-ons stay independent

Forge families, Reverse THIS! and comfy-forge own their own code, releases and doctors. The bundle only discovers them, and reports a source checkout and its installed runtime as one owner:

python TOOLS/companion_registry.py --verify     # NOT_FOUND / DISCOVERED / INSTALLED_UNVERIFIED
python TOOLS/companion_registry.py --print-doctor skyrim-forge --json   # its doctor as argv + working directory
python TOOLS/graph_roots.py                     # code-graph projects whose folder has moved

The base packs are automatic when Skyrim or Forge is found and automatic when its kit is found; a missing owner is skipped. Nothing is installed twice, and turning the pack off leaves the owner intact. Skyrim Forge and Reverse THIS! install their own latest release through their own installers when detected; the bundle pins no version and never reimplements an owner's tools.

Projects, skills and private settings

• TOOLS/Initialize-Workspace.ps1 seeds a project with PLAN.md (task table with IDs, dependencies, verification, evidence), STATE.md (goal, current task, next action, blockers), MEMORY/, VISION.md and WORKSPACE_OWNERSHIP.md.
• With a task table present, a small controller answers "what next" without a model call: next, status, resume [--fix], handoff. Its absolute path is recorded in MACHINE-MAP.md, so it works from any project folder.
• Personal preferences live in ~/.uabs/profile/PROFILE.md; generated paths live in MACHINE-MAP.md. Neither is published.
• Your model choices, instructions and configs are preserved; files are backed up before any change.

Models

The starter ships one model: deepseek/deepseek-v4.1-flash-1017 at high reasoning effort. One strong worker finishes more work per dollar than a cheap aggregator plus a planner; the MoA presets are defined but disabled. Routing sorts by price with require_parameters: true and hardcodes no provider list.

python TOOLS/refresh-provider-routing.py          # today's cheapest cache-read providers
python TOOLS/refresh-provider-routing.py --write <config>   # pin them (backup + atomic write)

Existing model choices are never replaced on upgrade, and there is no fallback chain: a silent downgrade is worse than a clear error. Shipped aliases are kept and documented but are not recommendations: hermes model v4.1-flash, flash, v4-pro, muse, gemini-flash, glm, grok, sol, opus, nemotron-ultra, nemotron-ultra-nofree; local is whatever your own LM Studio serves, since the portable starter ships no machine-specific endpoint. Dated prices and a worked cost example: docs/HERMES-MODELS.md.

Verified, not claimed

A PASS means a command ran and its exit status was checked. The pack gate checks skill trees, PowerShell parsing, offline hashes and upgrade fixtures; the doctor re-hashes what each install receipt binds; the release pipeline builds twice and requires identical bytes before publishing. reviewer claims from a model are not evidence — PLAN.md evidence has to name a command, a revision and an exit code.

Documentation

Document	Contents
Reference guide	every install switch, provider wiring, what gets installed, troubleshooting
Hermes models	dated prices, aliases, compression, caching
Tool evaluations	what was evaluated, what was rejected, why
Decisions	the architecture rules this repository is held to
Changelog · history	release notes
Notices	licences and attribution

Known limitations

• A few skills exceed the body-size budget; TOOLS/audit_skills.py names them.
• The knowledge graph cannot read Papyrus; use source search for .psc.
• Compatibility holds are deliberate: newer upstream is not automatically tested.
• Free model availability and prices change; the live catalog outranks any snapshot.
• Full-Offline carries payload snapshots, not every runtime or dependency cache.
• Parsing, packaging and protocol handshakes do not prove gameplay or GPU performance.

Third-party components

This list includes redistributed code, fetched packages and catalog-only candidates. A listing does not mean installed or connected. Evaluated versions, source commits and compatibility holds belong in CATALOG.json.

Catalog ID	Upstream project	Integration
universal-modder	Universal Modder	Vendored CLI/source and routed guides; default install
uabs-toolbox	UABS Toolbox	Bundle-owned; always-on find/describe/call, starts project tools on demand
housecarl	houseCARL	Offline payload; game/profile scope
codebase-memory	codebase-memory-mcp	Offline payload; code profile
headroom	Headroom	Wheel/source; core MCP
superpowers	Superpowers	Vendored skills and native plugins
ponytail	Ponytail	Vendored skills and native plugins
codeburn	CodeBurn	Default on-demand CLI
impeccable	Impeccable	Vendored skill/detector; separately pinned CLI
code-review-skill	Code Review Skill	Vendored canonical skill
obsidian-skills	Obsidian Skills	Vendored canonical skills; no duplicate plugin needed
claude-mem	claude-mem	Explicit -WithClaudeMem only
playwright-mcp	Playwright MCP	Scoped browser MCP
firecrawl-mcp	Firecrawl MCP	Security-held; not installed/registered
perplexity-mcp	Perplexity MCP	Credential-gated MCP
context7	Context7	Core documentation MCP
github-mcp-server	GitHub MCP Server	Official offline binary; core MCP
serena	Serena	Code-intel / Hermes code profile; LSP navigation and symbol edits
chrome-devtools-mcp	Chrome DevTools MCP	Scoped browser inspection
shadcn-mcp	shadcn MCP	Matching project registry profile
blender-mcp	Blender MCP	Editor profile; Blender/addon prerequisite
godot-mcp	Godot MCP	Editor profile; Godot prerequisite
unity-mcp	Unity MCP	Project package/generated executable prerequisite
rtk	RTK	Default CLI; narrow routing
omni	OMNI, now Contexere	Catalog/evaluation only; not installed
windows-mcp	Windows MCP	Manual opt-in; OFF by default
super-mcp-router	Super MCP Router	Evaluated catalog entry; no default router wiring
playwright-cli	Playwright CLI	Default CLI; browser installation separate
comfy-cli	ComfyUI CLI	Optional user-managed tool; not installed by default
rea	REA	Optional pinned CLI; no global MCP or setup writes
reverse-this	Reverse THIS!	Optional standalone analysis workbench; reuses an existing discovered tool bench, no duplicate installation; source MIT, individual payload licenses apply
CLI-Anything	HKUDS/CLI-Anything	Companion hub of per-app agent CLIs (Blender, FreeCAD, GIMP, Audacity, Godot, ComfyUI, browsers...). pip install cli-anything-hub, install the harnesses you need. Apache-2.0; v0.4.0 checked 2026-10-16. Not redistributed or installed by the pack.
anti-slop	miqdadbadjuber/anti-slop	MIT rules filter against generic AI UI, filler copy and AI-shaped code (rules/antislop.md plus a CLI). Overlaps the bundled impeccable and humanizer skills; use it as an extra filter. v3.2.21 checked 2026-10-16.
jevgrep	dzhng/jevgrep	Hosted code-search CLI (jg) that answers repo questions through Jev and returns files plus verbatim excerpts; needs its own account. The bundled codebase-memory-mcp already answers that class of question locally. 0.8.0 checked 2026-10-16.
Caveman	JuliusBrussee/caveman	Voice/input compression with its own multi-agent installer and an optional routing layer. Conflicts with this bundle's wiring: its installer writes the same SOUL.md / AGENTS.md / skills / hooks / settings.json files this pack owns, and its routing layers repoint ANTHROPIC_BASE_URL. Evaluated 4.0.1 (Apache-2.0) 2026-10-16 — documented, not installed; see evaluations.
OmniRoute	diegosouzapw/OmniRoute	Local free-tier gateway for CLI agents (localhost:20128/v1, 13 setup-* commands, RTK + Caveman compression inside). Conflicts with this bundle's invariants: it takes over provider endpoints and depends on stacked third-party free tiers, which the bundled token-efficiency skill rules out as a default. Evaluated 3.8.51 (MIT) 2026-10-16 — documented, not installed.

Additional canonical payloads without separate tool-catalog entries:

Payload	Upstream	Integration
img2threejs	img2threejs/img2threejs	Pinned pointer skill (v1.5.1); the tool tree lives in its own repo
AssetForge	SenjuWoo/AssetForge	Owns 2D/3D asset production, validation and engine export (MIT). Detected by the companion registry; optional providers (Blender, glTF-Transform, comfy-forge) are discovered, never installed by the bundle
Ponytail	off by default through its own defaultMode; `python TOOLS/ponytail_mode.py lite	full
mattpocock/skills, addyosmani/agent-skills	curated, namespaced subset installed on request: `python TOOLS/recommended_skills.py install pocock	addy --provider --apply`
awesome-gamedev-agent-skills	gamedev-skills/awesome-gamedev-agent-skills	Optional engine pack (Apache-2.0, pinned c6ca6c4, 2026-11-12). Not vendored and never installed by the bundle: the detector reports it, routing picks the engine, and the review covers install, precedence and removal
skill-creator, build-mcp-server, build-mcpb, build-mcp-app	anthropics/claude-plugins-official	Selected cross-provider skills, not the entire marketplace
Hermes native Unity bridge	CoplayDev/unity-mcp	Existing Coplay editor backend, isolated in unity/creative; distinct from the IvanMurzak project route

Skyrim Forge is the bundle author's product, released from SenjuWoo/SkyrimForge; the installer takes its latest release, verified by digest. Tool notices and skill notices retain licenses, credits and adaptation notes. The pack's MIT license does not relicense these projects. Preserve upstream attribution when redistributing.

Version

The version is in VERSION.txt. Every release, with notes, checksums, SBOM and attestation, is on the releases page.

Licence

Original code and docs are MIT; third-party payloads keep their own terms — see the notices.

Starting a chat, planning a big job for another model to build, resuming after a break or a usage limit, or handing a project to a newer model for a full audit? Copy one of the four short cross-provider prompts and add your task.