GitHub - RudVlad473/iluvatar: Classify a raw app idea into CS terms and emit a frozen, machine-parseable architecture contract — before any technology is named. · GitHub
/" data-turbo-transient="true" />
Skip to content
Type / to search
Sign in<br>Sign upAppearance settings
You signed in with another tab or window. Reload to refresh your session.<br>You signed out in another tab or window. Reload to refresh your session.<br>You switched accounts on another tab or window. Reload to refresh your session.
Dismiss alert
{{ message }}
RudVlad473
iluvatar
Public
Notifications<br>You must be signed in to change notification settings
Fork
Star
master
BranchesTags
Go to file
CodeOpen more actions menu
Folders and files<br>NameNameLast commit message<br>Last commit date<br>Latest commit
History<br>13 Commits<br>13 Commits
.claude-plugin
.claude-plugin
skills/iluvatar
skills/iluvatar
.gitignore
.gitignore
COMPARISON.md
COMPARISON.md
LICENSE
LICENSE
README.md
README.md
View all files
Repository files navigation
✨ iluvatar
Conceives the world before anything is built.
Ask an AI agent to architect your app and it names a tech stack before<br>it's named the problem — Postgres, Redis, and a queue appear before<br>anyone's asked what happens when two writes race, or what the app does<br>when the network is down.
iluvatar is a Claude Code skill that forces the boring questions first.<br>Give it a one-sentence app idea and it runs a 15-point classification<br>pass — state, failure, concurrency, compliance, the stuff that gets<br>skipped — and emits HIGH-LEVEL-ARCHITECTURE.md: a frozen,<br>schema-validated contract that names archetypes and cross-cutting<br>concerns but never a single technology. No tool gets picked until the<br>shape is settled.
In Tolkien's legendarium, Ilúvatar conceives Middle-earth in thought,<br>complete, before the Ainur ever sing it into being. That's the whole<br>pitch: architecture before creation, not architecture as an afterthought<br>to "which framework."
📦 Install
As a Claude Code plugin:
/plugin marketplace add RudVlad473/iluvatar<br>/plugin install iluvatar@RudVlad473
Or copy it directly — works anywhere, no plugin system required.<br>Requires Python 3.9+ on PATH (stdlib only, nothing to pip install).
git clone https://github.com/RudVlad473/iluvatar<br>mkdir -p ~/.claude/skills<br>cp -r iluvatar/skills/iluvatar ~/.claude/skills/ # available in every project
Or scoped to one project only:
git clone https://github.com/RudVlad473/iluvatar<br>mkdir -p /path/to/your/project/.claude/skills<br>cp -r iluvatar/skills/iluvatar /path/to/your/project/.claude/skills/
On Windows (PowerShell):
git clone https://github.com/RudVlad473/iluvatar<br>New-Item -ItemType Directory -Force -Path "$HOME\.claude\skills"<br>Copy-Item -Recurse iluvatar\skills\iluvatar $HOME\.claude\skills\
Then just describe your app idea, or say "run iluvatar."
🤝 Pairs well with
iluvatar classifies whatever idea you hand it — it doesn't sharpen a<br>vague one first. Best served after Matt Pocock's<br>grill-with-docs: a relentless<br>interview that pressure-tests the idea and produces docs (ADRs, a<br>glossary) as it goes, so iluvatar has something more complete to<br>classify by the time it runs.
/plugin marketplace add mattpocock/skills<br>/plugin install mattpocock-skills@mattpocock
What it produces
Give it: "Paste a URL, get a clean readable version saved for later,<br>with a weekly email digest of what I haven't read yet. I want to<br>self-host it on one small server. No browser extension or mobile app for<br>now."
iluvatar returns HIGH-LEVEL-ARCHITECTURE.md, classifying it as<br>etl-pipeline + crud-service + background-jobs, deriving a 10-stage<br>flow spine across write, read, and scheduled paths, and running real<br>numbers instead of adjectives:
archetypes: [etl-pipeline, crud-service, background-jobs]<br>primary_artifact: a saved article (URL in, extracted text stored)<br>budget_notes: "~7 saves/day, ~639MB total storage over a 5-year<br>retention window (botec.py), single self-hosted server..."
...and a flow spine that starts:
1. submit-url — user-entered URL → validated, canonicalized URL<br>2. fetch — canonicalized URL → raw HTML (or a recorded fetch failure)<br>3. extract — raw HTML → title + readable text + metadata<br>...
Every number is tagged with where it came from — a script run, something<br>the user stated, or an explicit assumption — so nothing in the document<br>reads as more certain than it is. The full worked example, all six<br>phases, is at<br>skills/iluvatar/references/example-output.md.
Where this fits
Tool<br>What it does<br>How iluvatar differs
BMAD Method (Analyst→PM→Architect)<br>Narrative architecture.md with tech-stack decisions already baked in, for a coding agent to read<br>Never names a technology; output is a schema-validated classification a tool parses by section name, not prose for an agent to interpret
GitHub Spec Kit<br>Captures requirements (what/why); deliberately excludes architecture<br>Sits downstream of what Spec Kit deliberately skips — the architecture classification, not another...