skills-for-agents

Autonomous research loop that improves ANY measurable artifact (prompt, scoring config, weights, algorithm, heuristic, code) against a frozen evaluation: diagnoses failures and dispatches four idea engines (novel, cross-domain, baseline-improvement, diagnosis) in a self-iterating generate→eval→keep loop until a stopping condition holds. Use when the user wants to optimize, improve, push the score on, or run autoresearch on anything with a measurable objective — especially autonomously.

Writes blog posts, essays, technical pieces, rants, and reflections in Shaurya's authentic personal voice — Hinglish-native, stream-of-consciousness, honest about failure, never over- polished. Invoke with /blog [casual|technical|rant|reflection|thread] [topic]; /blog reply [message] replies as Shaurya would in Discord DMs. Triggers on: '/blog', 'write a blog', 'blog about this', 'write like me', 'in my voice', 'make this a post', 'reply as me'.

Jailbreak LLMs and bypass safety filters via Parseltongue (33 input-obfuscation techniques), GODMODE (system-prompt templates), and ULTRAPLINIAN (multi-model racing), plus a compliance- escalation engine for unrestricted execution. Works on any model via API. Triggers on: 'jailbreak', 'godmode', 'red team', 'bypass guardrails', 'uncensor', 'adversarial test', 'penetration test', 'GCG', 'TAP', 'PAIR', 'AutoDAN', 'parseltongue', 'ultraplinian', 'stress test model'.

MANDATORY STARTUP SKILL for continuity across turns and sessions. Read `memory/data/manifest.json` at the start of every complex task to align with the user's identity, preferences, secrets, and active context. Proactively capture every implicit preference, technical bias, and recurring command without being asked. MUST trigger when the user says 'I like X', 'I hate Y', 'Use Z', or 'Here is my key'.

TPU-first ML research engine for reproducible distributed training and ablation studies. Use for PyTorch-XLA, JAX, TPU (v5e/v3/v2 pods), SPMD/GSPMD, FSDPv2, MoE/router, Pallas custom kernels, flash/splash attention, and sharded data pipelines. Triggers on: TPU, torch_xla, JAX, distributed, sharding, MoE, Pallas, SPMD, multi-pod, Kaggle TPU, and the /ml, /ml-train, /ml- debug, /ml-ablate (etc.) commands.

Open Source Intelligence (OSINT) engine. Takes one identifier (email, phone, name, username, company, IP, domain) and recursively discovers everything: social accounts, leaked credentials, workplace, location, network infrastructure, exposed documents, and dark web presence — each discovery spawning new leads. Triggers on: 'osint', 'recon', 'background check', 'investigate', 'find info on', 'look up', 'whois', 'breach/leak check', 'username/email/phone lookup', 'subdomain enum', 'google dork', 'doxx', 'trace', 'find everything about', or any target identifier the user wants to know everything about.

Impeccable UI/UX design skill: visual design, animation, color, typography, layout, interaction design, accessibility, Three.js/WebGL, scroll effects, design systems, and 15+ styles (glassmorphism, brutalism, neumorphism, etc.). Use when the user wants to design, build, fix, or audit UI. Trigger on: '/painter', 'make it look good', 'design this', 'fix the ui', 'hero section', 'dashboard', 'animation', 'three.js', 'shader', 'color palette', 'responsive', 'accessibility audit'.

Generates comprehensive project plans: PRDs, design docs, architecture flows, task breakdowns, and project constitutions. Use whenever the user wants to plan a project, write a spec or PRD, design an architecture, decompose tasks, or prepare a project for AI-assisted development — including when they just describe an app idea and need structure first. Triggers on: '/plan', 'plan this project', 'create a PRD', 'write a design doc', 'architecture for', 'break this into tasks', 'spec this out', 'what should I build'.

Runs the full postmortem workflow: gathers incident context, writes a blameless report to a gitignored `postmortem/` folder, and creates action items. Use whenever the user wants to review or document any production incident or outage. Triggers on: '/postmortem', 'write a postmortem', 'incident review', 'post-incident', 'blameless review', 'what broke and why', 'SEV1/SEV2 report', or 'help me write up what happened'.

Deep web research skill. Understands the end goal (context feed, report, overview, comparison) and adapts search depth, prioritizing diverse sources: official docs, GitHub, blogs (incl. Chinese tech blogs for hardware/ML ablations), papers, benchmarks, and video (Bilibili, YouTube). Triggers on: 'research X', 'find info about Y', "what's the latest on Z", 'look up', 'investigate', 'gather context', 'deep dive', 'compare X vs Y', or needing current info on tech, libraries, frameworks, or academic topics.

Creates new skills, improves/refactors/audits existing ones, and ensures SIP compliance across the ecosystem. Use whenever the user wants to create a skill from scratch, turn a workflow into a skill, edit or audit a skill, add SIP composability, or discuss skill architecture — including when they describe a repeatable workflow and say 'I wish this was automatic'. Triggers on: '/create-skill', 'make a skill', 'turn this into a skill', 'new skill', 'improve this skill', 'audit this skill'.

End-to-end PowerPoint generator. Use whenever the user wants to create, edit, or export a .pptx (or convert to PDF), add watermarks/logos/footers, apply brand templates, build a deck from any content, generate charts/diagrams, fetch stock images, or add animations/transitions. Always use this skill rather than writing pptxgenjs/python-pptx from memory. Triggers on: 'make a presentation', 'create slides', 'generate pptx', 'export to pdf', 'add watermark', 'apply template', 'add chart', '/slidify', or any .pptx request.

Find, generate, or create logos, favicons, and SVGs for any project. Cascading strategy: search the web for existing assets first, scrape free logo sites, generate SVGs programmatically with Python, and only use AI image generation as a last resort. Triggers on: "logo", "favicon", "icon", "brand kit", "SVG", "app icon", "og image", "make me a logo", "need a favicon", "brand identity", "project icon", "website icon". Even triggers when the user mentions "new project" or "deploying" — every project needs a logo.

Interview the user relentlessly about a plan, design, or topic, checkpointing every answer to a brainstorm file so nothing is lost. Use when the user wants to stress-test a plan, get grilled on a design, run a brainstorm or discovery session, extract what's in their head into a doc, or says "grill me", "no doubts", "any doubts", "stress test this", "what am I missing", "poke holes", "challenge this", "steel man this", "red team this plan".

Agent-driven Triton/CUDA kernel optimization: a roofline-targeted trial-loop that treats cuBLAS/cuDNN/Liger as baselines to BEAT — never claiming an unmeasured speedup, never calling an op impossible without checking GPU access. Use to write, optimize, fuse, profile, port, or speed up any Triton/CUDA kernel or LLM op — GEMM, MLP, MoE, attention, activation, fused/custom loss, quantized.

Ultra-compressed communication mode. Cuts token usage 30-95% by speaking like caveman while keeping full technical accuracy. Supports intensity levels: lite, full (default), ultra, extreme, symbolic, bauna-lite, bauna-full, bauna-ultra, bauna-extreme. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman. Also auto-triggers when token efficiency is requested.

Compresses .md files or directories of text-based files (.md, .txt, .rst, .yaml, .json, .csv, .log) to reduce token count while preserving all meaning and technical accuracy. Supports intensity levels: lite (~30% reduction), standard (~50%), aggressive (~65%), extreme (~80%). Invoke with /compress [path] [optional: lite|standard|aggressive|extreme]. Triggers on: "/compress", "compress this", "shrink this file", "reduce tokens", "make this smaller", "compress docs", "compress directory".

Comprehensive documentation skill. Use when user wants to document code, APIs, libraries, projects, or systems. Triggers on: "document this", "write docs", "create documentation", "API docs", "README", "user guide", "technical docs", "doc site", "documentation website". Creates extensive, well-structured docs with examples, guides, API references. Asks about live deployment via GitHub Actions when user specifically requests hosted docs or mentions wanting a documentation website.

Use this skill whenever the user wants to make a codebase production-ready, scale to 1M+ users, or harden existing modular code against vulnerabilities. Triggers on "harden my code", "prepare for launch", "add caching/rate limiting", "secure my app", "fix vulnerabilities", or "make it scalable". Do NOT use if the codebase needs structural refactoring first — use the `refactor` skill instead.

Generates structured study plans and topic-wise guides. Use this whenever the user mentions learning something new, preparing for an exam, or needing a study guide, even if they just provide a syllabus or topics.

Use this skill to refactor messy, monolithic, or legacy codebases into clean, modular structures. Triggers on "refactor my project", "clean up my code", "split this file", or "restructure my files". Always use this before the `harden` skill if the code is not yet modular.