🍷 Taste Your Taste

Autonomous Developer Taste Curation & Vibe Stacking Engine for AI Agents

CLAUDE.md • .cursorrules • .cursor/rules/*.mdc • AGENTS.md • .agent/rules.md

"Code generation is cheap. Taste is rare."

# 🧙 Interactive 10-Second Setup Wizard:
npx taste-code init

# 🍸 Instant Vibe Stacking (Lego Bricks):
npx taste-code blend --style antfu --tone karpathy

# 💧 Token Compactor & De-slop Optimizer (Save 30-70% Context Tokens):
npx taste-code prune CLAUDE.md --write

# ⚖️ Engineering Philosophy Comparator (Antfu vs Karpathy):
npx taste-code diff antfu karpathy

# 🌶️ Savage Linus Torvalds Taste Roast Critic:
npx taste-code roast CLAUDE.md

# 🪪 Dynamic GitHub Profile Taste HUD Card Generator:
npx taste-code card --user yourname --theme cyber --output taste-card.svg

# 🛡️ Sleek Vector Micro-Badge Generator (Shields style):
npx taste-code badge --type score --output taste-badge.svg

---

Latest Radar Activity

Updated after each completed daily scan. Discoveries and rule changes are automated observations awaiting review; curated labels refer to project selection.

122 tracked projects | 26 curated | 96 pending review

Last completed scan: 2026-10-19 07:06:26 UTC.

Latest content activity: 2026-10-19 07:06:26 UTC.

Browse all projects in the gallery | Review discoveries

Updated (UTC)	Project	Activity	Review status	Rules
2026-10-19 07:06:26	percona/pmm	New discovery	Pending review	Archive
2026-10-19 07:06:20	kirodotdev/KiroCrew	New discovery	Pending review	Archive
2026-10-19 07:06:13	androidx/androidx	New discovery	Pending review	Archive
2026-10-19 07:06:08	huggingface/OpenEnv	New discovery	Pending review	Archive
2026-10-19 07:06:03	Comfy-Org/ComfyUI	New discovery	Pending review	Archive
2026-10-19 07:05:54	thedotmack/claude-mem	Rules updated	Pending review	Archive
2026-10-18 06:47:39	requarks/wiki	New discovery	Pending review	Archive
2026-10-18 06:47:34	PHPantom-dev/phpantom_lsp	New discovery	Pending review	Archive
2026-10-18 06:47:28	anomalyco/opencode	New discovery	Pending review	Archive
2026-10-18 06:47:23	anthropics/claude-code	New discovery	Pending review	Archive

---

📖 The Manifesto

In the pre-AI era, developers loved exploring the .dotfiles (.zshrc, .vimrc, tmux.conf) of legendary hackers to see how they tuned their tools.

In the Vibe Coding era, human programmers don't write every line of syntax by hand. Instead, developer taste is captured in the behavioral guardrails, engineering constraints, and architectural temperaments we impart to our AI agents:

• How do world-class teams prevent AI slop (verbose apologies, gratuitous mocks, boilerplate)?
• How do high-velocity solo developers force AI to produce concise, elegant diffs?
• How do infrastructure projects like NVIDIA, Stripe, and tidyverse instruct Claude Code and Cursor?

taste-your-taste is an end-to-end, LLM-first developer ecosystem:

1. 🛠️ taste CLI: A zero-dependency Lego-brick stacking tool to blend master developer tastes into your repository.
2. 🧙 taste init Wizard: Interactive 10-second TUI to configure CLAUDE.md, Cursor MDC, or Antigravity rules.
3. 💧 taste prune (De-slop Compactor): Strips out polite conversational filler and tautological fluff while preserving 100% of technical rules, reducing context window tax by 30-70%.
4. ⚖️ taste diff (Philosophy Comparator): Side-by-side comparative analysis contrasting conflicting engineering temperaments (e.g., Antfu vs Karpathy).
5. 🌶️ Taste Roast Critic: Linus Torvalds-inspired LLM critic that audits your agent instructions with brutal technical honesty.
6. 🪪 taste card & taste badge: Generates aesthetic cyberpunk HUD profile cards (495x195) and sleek vector micro-badges (28px pill) with custom themes and radial gauges.
7. 🏛️ Top-Level Open-Source Directories: Authentic, battle-tested configs from 26+ star projects organized directly as first-level directories.
8. 🧠 Curated Agent Skills Library (skills/): Modular SKILL.md playbooks (Andrej Karpathy guidelines, offensive AI security, de-AI writing, design tokens, etc.).
9. 📡 Autonomous LLM Radar: GitHub Action engine that monitors GitHub Trending and tracks Prompt Evolution via Git Blob SHAs.
10. 🌐 Interactive Web Gallery: Daily-updated explorer, Taste Blender, Pruner preview, and live Badge Studio.

---

🏛️ Repository Architecture

All star open-source project configs and modular skill playbooks are structured directly as top-level directories:

taste-your-taste/
├── 📋 Agent Directives & Specifications
│   ├── AGENTS.md                  # Autonomous agent operating directives & constraints
│   └── CLAUDE.md                  # Claude Code repository conventions & style guidelines
├── 🛠️ Core Engines & CLI
│   ├── bin/taste                  # Executable CLI entrypoint
│   ├── src/                       # Zero-dependency modular engines
│   │   ├── registry.js            # Style & tone brick loader
│   │   ├── stacker.js             # Marker-based non-destructive injector
│   │   ├── pruner.js              # Token compactor & de-slop optimizer
│   │   ├── differ.js              # Philosophy & constraint comparator
│   │   ├── roaster.js             # Linus Torvalds technical audit engine
│   │   └── card.js                # GitHub profile taste badge generator
│   ├── registry/                  # Atomic Lego taste bricks (styles, tones, presets)
│   └── api/                       # Serverless SVG Profile Card Badge API (api/card.js)
├── 🧠 Curated Agent Skills Library
│   └── skills/                    # 7 modular SKILL.md behavioral playbooks
│       ├── karpathy-guidelines/   # ⭐️ 216k — Simplicity & surgical diffs
│       ├── offensive-ai-security/ # ⭐️ 7.1k — AI red-teaming & vulnerability assessment
│       ├── agentic-design/        # ⭐️ 2.9k — Frontend Generative UI & design tokens
│       ├── sepia-hemingway/       # ⭐️ 2.9k — De-AI humanized writing engine
│       ├── claude-osint/          # ⭐️ 2.7k — Open-source reconnaissance & intelligence
│       ├── markit/                # ⭐️ 1.3k — Clean LLM markdown extraction
│       └── repo-task-proof-loop/  # ⭐️ 732  — Task proof loop with autonomous exit gates
├── 🌟 Star Open-Source Projects (Top-Level Repositories)
│   ├── openclaw/                  # ⭐️ 390.8k — AI personal assistant operating system
│   ├── andrej-karpathy-skills/    # ⭐️ 216.0k — Andrej Karpathy's surgical guidelines
│   ├── ponytail/                  # ⭐️ 148.7k — Minimalist web framework
│   ├── VoiceStudio/               # ⭐️ 49.8k  — Production voice synthesis pipeline
│   ├── agentsmd/                  # ⭐️ 24.7k  — Universal agent guidelines
│   ├── open-saas/                 # ⭐️ 16.0k  — React & Node SaaS template
│   ├── cbmc/                      # ⭐️ 12.8k  — C Bounded Model Checker
│   ├── nvidia-openshell/          # ⭐️ 11.6k  — NVIDIA AI shell automation
│   ├── securego-gosec/            # ⭐️ 8.9k   — Go security AST checker
│   ├── rust-on-nails/             # ⭐️ 8.4k   — Rust full-stack production architecture
│   ├── rtags/                     # ⭐️ 7.6k   — C/C++ indexing engine
│   ├── claude-red/                # ⭐️ 7.1k   — Offensive AI red-teaming playbook
│   ├── claude-token-efficient/    # ⭐️ 6.1k   — Token-optimized Claude Code config
│   ├── devin-cursorrules/         # ⭐️ 5.9k   — Devin-styled cursor rules
│   ├── stripe-java/               # ⭐️ 4.7k   — Enterprise Java SDK by Stripe
│   ├── gravity/                   # ⭐️ 4.6k   — Antigravity agentic workflow engine
│   ├── tidyverse-readr/           # ⭐️ 3.9k   — R data ingest library
│   ├── awesome-design-skills/     # ⭐️ 2.9k   — Frontend agent design tokens
│   ├── sepia-skills/              # ⭐️ 2.9k   — Hemingway de-AI writing engine
│   ├── claude-osint/              # ⭐️ 2.7k   — Cloud OSINT & reconnaissance
│   ├── graphframes/               # ⭐️ 2.3k   — Graph processing for Apache Spark
│   ├── keras-cv-attention-models/ # ⭐️ 1.8k   — Deep learning CV architectures
│   ├── prismer/                   # ⭐️ 1.6k   — Vision-language reasoning model
│   ├── markit/                    # ⭐️ 1.3k   — Intelligent web content extractor
│   ├── browser-operator/          # ⭐️ 1.2k   — Autonomous browser automation agent
│   └── repo-task-proof-loop/      # ⭐️ 732    — Spec-driven task proof loop
└── 📡 Automation, Data & Gallery
    ├── scripts/radar.py           # Autonomous LLM Radar & PR Stager
    ├── data/seen_repos.json       # Discovery & blob hash tracking
    ├── docs/index.html            # Real-time Web Gallery & Badge Studio
    └── test/                      # Native Node.js test suites

---

🚀 Key Features

1. 🧙 Interactive Setup Wizard (taste init)

Configure your project in 10 seconds through an interactive terminal interface:

npx taste-code init

The wizard guides you through selecting:

• Target Format: CLAUDE.md, .cursor/rules/*.mdc, .cursorrules, .agent/rules.md, or AGENTS.md.
• Engineering Style: antfu, karpathy, stripe, minimalist, defensive, hacker.
• Agent Tone: karpathy (anti-slop diff-first), linus (anti-overengineering), terse (ultra-compact), teacher (edge-case pedagogy).

---

2. 🍸 Taste Stacking Engine (taste add & taste blend)

Copy-pasting someone else's 300-line prompt is clumsy and brittle. taste breaks master developer styles and agent tones into atomic, composable Lego bricks.

# View all available styles, tones, and presets
npx taste-code list

# Stack minimalist zero-dependency rules into CLAUDE.md
npx taste-code add minimalist

# Stack Antfu TypeScript rules into Modern Cursor MDC (.cursor/rules/taste.mdc)
npx taste-code add antfu --target mdc

# Blend Antfu's TypeScript strictness with Karpathy's terse, outcome-driven tone
npx taste-code blend --style antfu --tone karpathy

Non-Destructive Marker Boundaries

The CLI manages rules using non-destructive block boundaries (<!-- TASTE:STYLE:... --> and <!-- TASTE:TONE:... -->). You can re-run and re-stack at any time without duplicating content or overwriting your own custom rules.

Supported Rule Formats & Targets

Target Flag	Generated File	Target Agent / IDE
(default)	CLAUDE.md	Claude Code (Root)
--target dotclaude	.claude/CLAUDE.md	Claude Code (Nested)
--target mdc	.cursor/rules/<name>.mdc	Modern Cursor (with YAML frontmatter)
--target cursor	.cursorrules	Legacy Cursor
--target agent	.agent/rules.md	Google Antigravity / Gemini Agent
--target agents	AGENTS.md	Multi-Agent Specification standard

---

3. 💧 Token Compactor: taste prune

Agent instructions (CLAUDE.md, .cursorrules, AGENTS.md) are injected into the LLM context on every single prompt or turn. Verbose corporate platitudes, excessive politeness, and conversational slop silently inflate your API bills and dilute the model's attention.

taste prune strips conversational filler and tautological fluff while preserving 100% of functional engineering commands, linters, and architectural invariants:

# Preview token savings and distilled prompt
npx taste-code prune CLAUDE.md

# Prune and overwrite file in-place with automatic .bak backup
npx taste-code prune CLAUDE.md --write --backup

What It De-slops:

• 📉 Politeness Tax: Removes "Please", "Kindly", "If you make a mistake apologize immediately".
• 🧽 Tautological Fluff: Cuts "Always write clean, readable code and follow all best practices".
• 🛡️ Zero Loss of Invariants: Strictly preserves testing commands (npm test, pytest), linter commands (eslint --fix), and negative constraints (never introduce 'any').
• ⚡ Context Economy: Reduces instruction token overhead by 30% to 70%.

---

4. ⚖️ Taste Philosophy Differ: taste diff

Engineering cultures collide. Should an agent be an obsessive TypeScript perfectionist, or a single-file raw PyTorch hacker?

taste diff runs an architectural and philosophical contrast between two flavors or existing files:

# Compare two master archetypes
npx taste-code diff antfu karpathy

# Compare existing configs
npx taste-code diff CLAUDE.md .cursorrules

Sample Output:

┌─ ⚡ TASTE PHILOSOPHY DIFF ─────────────────────────────────────────────────────────────────────────┐
│ 🥊 Comparing: antfu vs karpathy                                                               │
│                                                                                               │
│ 💥 The Philosophy Clash:                                                                      │
│   antfu demands rigorous type safety, modular packages, and modern ESM tooling.               │
│   karpathy prioritizes raw readability, single-file scripts, zero unnecessary dependencies,   │
│   and immediate algorithmic transparency over formal abstractions.                            │
│                                                                                               │
│ 🔍 Key Contrasts:                                                                             │
│   • antfu enforces strict TypeScript compilation and automated linting.                      │
│   • karpathy favors flat script simplicity and zero dependency overhead.                      │
│                                                                                               │
│ 🎯 When to choose antfu: Multi-package TS libraries, production enterprise codebases.          │
│ 🎯 When to choose karpathy: Machine learning experiments, proof-of-concepts, hackathons.      │
└───────────────────────────────────────────────────────────────────────────────────────────────┘

---

5. 🌶️ Taste Roast: LLM-Driven Linus Torvalds Critique

How good are your agent instructions? Are you paying Anthropic or OpenAI to generate verbose corporate apologies?

Run the Taste Roast:

npx taste-code roast [path/to/CLAUDE.md]

What It Audits (Powered by LLM):

• 🧠 LLM Semantic Evaluation: Evaluates constraints, actionable commands, and token economics.
• 📉 Fluff & Platitude Index: Detects useless corporate clichés ("write clean code", "strive for excellence", "be helpful").
• 💸 Politeness Tax: Calculates token budget wasted on polite greetings and apologies ("please", "kindly", "sorry").
• 🛡️ Negative Armor: Verifies whether you gave the AI explicit boundaries ("never", "do not", "prohibit").
• ⚙️ Verifiable Tooling: Checks whether the agent was given exact test/linter commands to verify its output.
• 💯 Taste Score (0-100): From CRIMINAL TOXIC WASTE to CHEF'S TASTE.

Sample Output:

┌─ 🌶️ TASTE ROAST REPORT ────────────────────────────────────────────────────────────────────────┐
│ Target File: CLAUDE.md                                                                        │
│ Lines: 7  |  Tokens: ~61  |  Words: 45  |  Engine: 🧠 LLM Reasoning                            │
│                                                                                               │
│ Taste Score: 12 / 100 [███░░░░░░░░░░░░░░░░░░░░░░░░░░░]                                        │
│ Verdict:     CRIMINAL TOXIC WASTE                                                             │
│                                                                                               │
│ 🔥 Sins & Pathology Detected:                                                                 │
│   ❌ Zero specificity — 'clean code' is undefined and unmeasurable                            │
│   ❌ 'Be helpful' is a vibe, not an operational instruction                                  │
│   ❌ No output contract or verifiable testing commands                                       │
│                                                                                               │
│ 🎙️ Linus Torvalds Roasts Your Taste:                                                         │
│   "This isn't a code review, it's a fortune cookie. 'Write clean code' is what every         │
│    professor says before assigning homework they won't grade."                                │
│                                                                                               │
│ 💊 Remedy & Prescription:                                                                     │
│   Run npx taste-code blend --style minimalist --tone terse to replace fluff with discipline. │
└───────────────────────────────────────────────────────────────────────────────────────────────┘

---

6. 🪪 Dynamic Profile Taste HUD Card & Micro-Badge: taste card & taste badge

Showcase your developer taste DNA directly on your GitHub Profile README, project documentation, or website with dynamic, dark-mode SVG HUD vector cards and sleek micro-badges.

1. Cyberpunk Profile HUD Card (495x195)

Features a frosted dark canvas, circular neon radial score gauge, AI Verified live pulse beacon, technical blueprint grid, and multi-theme colorways (cyber, matrix, midnight, sunset, noir):

# Generate local taste-card.svg with default Cyberpunk Rose theme
npx taste-code card --user yourname --theme cyber --output taste-card.svg

# Generate with Matrix Emerald hacker theme
npx taste-code card --user yourname --theme matrix --output taste-card.svg

2. Sleek Vector Micro-Badge (28px Pill)

For repository headers, status badges, and minimal profile headers:

# Generate Taste Score Badge ([ 🍷 TASTE | 94/100 · Chef's Taste ])
npx taste-code badge --type score --output taste-badge.svg

# Generate Taste DNA Badge ([ 🍷 TASTE DNA | antfu + karpathy ])
npx taste-code badge --type dna --output taste-badge.svg

# Generate Archetype Badge ([ 🛡️ ARCHETYPE | Defensive Architect ])
npx taste-code badge --type archetype --output taste-badge.svg

3. Live Dynamic Serverless HTTP API

Embed real-time SVG cards and badges directly using the serverless API:

<!-- Full HUD Profile Card -->
[![My Developer Taste](https://taste-your-taste.vercel.app/api/card?user=antfu&style=antfu&tone=linus&score=96&theme=cyber)](https://github.com/joe1chief/taste-your-taste)

<!-- Vector Micro-Badge -->
[![My Taste Badge](https://taste-your-taste.vercel.app/api/badge?type=score&score=96)](https://github.com/joe1chief/taste-your-taste)

Card & Badge API Query Parameters:

Parameter	Default	Description
user	developer	GitHub username displayed on card
style	antfu	Active developer style module
tone	karpathy	Active persona / agent tone
archetype	Anti-Slop Minimalist	Architectural archetype (Defensive Architect, Hacker Velocity, etc.)
score	94	Taste score (0 - 100)
theme	cyber	HUD colorway: cyber, matrix, midnight, sunset, noir
type	card	Display format: card (495x195 HUD) or badge (28px pill)
badgeType	score	Micro-badge type: score, dna, archetype, slop, roast

---

🏛️ Star Open-Source Projects (Top-Level Directories)

Each open-source project is directly accessible as a first-level directory in the repository:

Project Directory	Source Repository	Stars	Language	Taste Archetype	Key Philosophy & Invariants
📁 openclaw/	openclaw/openclaw	⭐️ 390.8k	Rust / C	Defensive Architect	One owner per responsibility, proof scoped, and zero blind retries.
📁 andrej-karpathy-skills/	andrej-karpathy-skills	⭐️ 216.0k	Python	Anti-Slop Minimalist	Karpathy workflow: single-file scripts, stdlib-first, zero dependency bloat, runnable code.
📁 ponytail/	DietrichGebert/ponytail	⭐️ 148.7k	JavaScript	Engineering Craft	"Lazy senior dev mode: the best code is the code you never wrote." Minimal abstractions.
📁 VoiceStudio/	debpalash/VoiceStudio	⭐️ 49.8k	Python / Electron	Anti-Slop Minimalist	Strict token economy: 1-line status updates max, electron desktop only, direct git diffs.
📁 agentsmd/	agentsmd/agents.md	⭐️ 24.7k	Specification	Pragmatic Taste	The open standard specification for guiding coding agents with explicit build & test contracts.
📁 open-saas/	wasp-lang/open-saas	⭐️ 16.0k	TypeScript / React	Engineering Craft	Modern full-stack SaaS boilerplate; strict package boundaries; production-ready TypeScript.
📁 nvidia-openshell/	NVIDIA/OpenShell	⭐️ 11.6k	Rust	Anti-Slop Minimalist	Injected on every prompt; strictly prohibits unsolicited restructuring or stylistic churn.
📁 securego-gosec/	securego/gosec	⭐️ 8.9k	Go	Defensive Architect	AST walking invariants, rule ID stability, and deterministic test invocation.
📁 claude-token-efficient/	claude-token-efficient	⭐️ 6.1k	Markdown	Anti-Slop Minimalist	Extreme token reduction; eliminates all conversational preamble; pure diff responses.
📁 devin-cursorrules/	grapeot/devin.cursorrules	⭐️ 5.9k	Rules	Hacker Velocity	Autonomous self-debugging loops, automated verification, diff-first execution.
📁 gravity/	marcobambini/gravity	⭐️ 4.6k	C	Defensive Architect	Memory safety invariants in C, zero memory leak tolerance, strict Valgrind testing.
📁 rtags/	Andersbakken/rtags	⭐️ 1.8k	C / C++	Defensive Architect	Notes for future LLM sessions: persistent symbol database, client/server protocol, zero regression.
📁 keras-cv-attention-models/	keras_cv_attention_models	⭐️ 1.7k	Python	Hacker Velocity	black -l 160, single-line multi-assignments, compact procedural clarity, thin wrapper pattern.
📁 graphframes/	graphframes/graphframes	⭐️ 1.2k	Scala / Java	Engineering Craft	Mission-critical Apache Spark codebase rules: strict backward compatibility, zero regressions.
📁 cbmc/	diffblue/cbmc	⭐️ 1.1k	C / C++	Defensive Architect	28k-char comprehensive AI coding assistant guide: SAT/SMT verification invariants, GOTO pipeline.
📁 tidyverse-readr/	tidyverse/readr	⭐️ 1.1k	R / C++	Defensive Architect	Boundary between Edition 2 (lazy parsing) and Edition 1 (eager C++ parser).
📁 claude-red/	SnailSploit/Claude-Red	⭐️ 7.1k	Security / Shell	Defensive Architect	Offensive AI red-teaming skills: API security abuse, secret leak audits, exploit simulation.
📁 awesome-design-skills/	awesome-design-skills	⭐️ 2.9k	Design / Tokens	Engineering Craft	Agentic UI design skills, design system tokens, typography scales, layout guardrails.
📁 sepia-skills/	Nanako0129/sepia	⭐️ 2.9k	Writing / Rules	Anti-Slop Minimalist	De-AI Hemingway writing skill: strips corporate synthetic filler and platitudes from agent output.
📁 claude-osint/	elementalsouls/Claude-OSINT	⭐️ 2.7k	Shell / Python	Hacker Velocity	Autonomous open-source intelligence gathering skill, cloud footprint discovery.
📁 markit/	shift-labs-ai/markit	⭐️ 1.3k	TypeScript / MD	Engineering Craft	Document to markdown converter skill: transforms complex PDFs and HTML into clean LLM context.
📁 repo-task-proof-loop/	repo-task-proof-loop	⭐️ 732	Specification	Defensive Architect	Spec-driven task proof loop with autonomous subagent spawning and exit verification.
📁 prismer/	Prismer-AI/Prismer	⭐️ 800+	Python / PyTorch	Pragmatic Taste	Multi-modal vision-language architecture, PyTorch distributed training guardrails.
📁 stripe-java/	stripe/stripe-java	⭐️ 600+	Java	Engineering Craft	Exact just test runners, Spotless formatting commands, and HTTP abstraction map.
📁 browser-operator/	browser-operator-core	⭐️ 510+	TypeScript / Python	Hacker Velocity	Headless browser control invariants, DOM extraction resilience, zero flakiness.
📁 rust-on-nails/	purton-tech/rust-on-nails	⭐️ 444	Rust	Hacker Velocity	Multi-agent separation: documentation agent, CLI agent, architecture boundary validation.

---

🧠 Curated Agent Skills Library (skills/)

In addition to whole-repository rules (CLAUDE.md, AGENTS.md), modern AI coding agents (Claude Code, Google Antigravity, Cursor 2.0, Codex) support specialized Agent Skills (SKILL.md) — self-contained behavioral playbooks that can be dynamically loaded into context when needed.

All modular skills are indexed in skills/:

Modular Skill	Direct Path	Source	Capability
karpathy-guidelines	skills/karpathy-guidelines/	⭐️ 216k	Andrej Karpathy's guidelines: simplicity first, surgical diffs, no speculative code.
offensive-ai-security	skills/offensive-ai-security/	⭐️ 7.1k	Red-teaming and defensive audits: secrets scanning, attack surface analysis.
agentic-design	skills/agentic-design/	⭐️ 2.9k	Generative UI & token architectures for web applications.
sepia-hemingway	skills/sepia-hemingway/	⭐️ 2.9k	De-AI humanized writing: strips apologies, corporate fluff, and synthetic sycophancy.
claude-osint	skills/claude-osint/	⭐️ 2.7k	Open-source intelligence: cloud recon, public DNS & asset mapping.
markit	skills/markit/	⭐️ 1.3k	Turns messy HTML, PDFs, and rich media into clean LLM markdown context.
repo-task-proof-loop	skills/repo-task-proof-loop/	⭐️ 732	Spec-driven task proof loop with autonomous subagent spawning and exit gates.

💡 How to Load Skills into Your AI Coding Agent:

• Claude Code: Direct Claude to read any skill: @skills/karpathy-guidelines/SKILL.md or copy it into .claude/skills/.
• Google Antigravity: Place the skill in .gemini/skills/ or reference via agent directives.
• Cursor & Codex: Reference skills/<skill-name>/SKILL.md in prompt context or .cursorrules.

---

📡 The Autonomous LLM Radar

Powered by GitHub Actions and scripts/radar.py, this repository monitors trending open-source projects daily using large language models instead of rigid regex heuristics:

flowchart LR
    A["🔥 GitHub Trending & High-Star Repos (⭐️ >= 1000)"] --> B["🤖 GitHub Actions (Daily Cron)"]
    B --> C["🔍 scripts/radar.py Engine"]
    C --> D{"Taste Files Found?"}
    D -- Yes --> E["🧠 LLM Architectural & Taste Assessment"]
    E --> F["📬 Create Review Issue & Archive Rules"]
    F --> G["📡 Rebuild README Activity & Gallery Catalog"]
    G --> I["🌐 Publish Gallery to GitHub Pages"]
    F --> J["🏆 Maintainer Curates to Hall of Fame"]
    D -- No --> G

Loading

• Strict Quality Gate: Only captures repositories that are either on GitHub Trending (Daily/Weekly) or have ⭐️ 1,000+ Stars (rejecting personal/low-impact repos).
• ⚡ Prompt Evolution Tracking: Stores Git Blob SHAs to detect when authors iterate or revise their prompts. When an author updates their rules, Radar detects the diff and triggers an evolution report issue!
• 🚀 Automated PR Staging (--create-pr): Optionally creates git branches and automated Pull Requests to stage newly discovered tastes straight into the repository.
• Isolated Archives: New radar imports live in discovered/<owner>/<repo>/, with META.json verifying repository ownership before every write. Upstream file paths are preserved. Existing curated top-level directories and skills/ remain unchanged; evolution can read historical archives only when their metadata matches. The recovered mattpocock/skills import is in discovered/mattpocock/skills/.
• Daily Visible Updates: Every completed scan regenerates the latest-activity section above and docs/catalog.json, commits them with the archives, and explicitly publishes the gallery. Pending discoveries appear immediately with their review status; curated snapshots remain unchanged. The scan timestamp advances even when no new projects are found.
• Rebuild Locally: Run python3 scripts/build_catalog.py. Use python3 scripts/build_catalog.py --check to detect stale generated files. Only run --record-scan --run-url <workflow-run-url> immediately after a completed Radar scan; local rebuilds retain the existing scan evidence.
• Pages Setup: Select Settings > Pages > Source > GitHub Actions once, then run the Publish Gallery workflow or push a gallery change to main. Radar calls the same publisher after its own commit, since a GITHUB_TOKEN push does not trigger another Pages build. Preview locally with python3 -m http.server 8000 --directory docs.
• Regression Tests: Run python3 -m unittest discover -s test -p "test_*.py" for Radar and catalog coverage, node --test test/gallery.test.js for gallery filters, and npm test for the CLI suite.
• Zero Spam: History is tracked in data/seen_repos.json to prevent duplicates.

---

⚙️ Configuration & Security

Both the taste CLI and Radar Engine natively leverage OpenAI-compatible LLM endpoints:

# Optional: Set your preferred LLM provider (defaults to standard OpenAI)
export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://api.openai.com/v1"
export LLM_MODEL="gpt-4o-mini"

Security & Privacy Guarantees:

• Zero Credential Leaks: API keys are never stored, printed, logged, or bundled.
• Offline Deterministic Fallback: If no API key is provided or the network is unavailable, all CLI commands (prune, diff, roast, card) fall back to deterministic local heuristic analyzers without crashing.
• Zero Runtime Dependencies: The entire CLI runs directly on native Node.js standard libraries.

---

🤝 Contributing

1. Submit via Issue: Click ✨ Submit a Developer Taste.
2. Submit a Taste Lego Brick: Add a new style or tone to registry/styles/ or registry/tones/ and submit a Pull Request!

---

📄 License

Distributed under the MIT License. All curated taste files and skill playbooks remain the copyright of their respective authors under their original open-source licenses.