Taiwan-built, self-hosted AI agent harness in a single Go binary — writes its own tools, sandbox-tests them, and shares them over MCP with Claude Code and Codex.
A curated monorepo of Agent Skills — 102 Tier-2 skills (the catalog is one of them) + the showcase renderer. Compatible with Claude Code, Codex, Cursor, Gemini CLI, Goose, Copilot, and any agent that consumes the SKILL.md standard.
Layout: no root SKILL.md (the README is the discovery surface). Skills are bucketed by single-noun category at skills/ / /SKILL.md (depth-2). skills.sh discovers depth-2 by default — requires CLI ≥ v1.5.8 — and --skill resolves path-independently, so install commands don't reference the category.
A curated monorepo of Agent Skills — 111 Tier-2 skills (the catalog is one of them) + the showcase renderer. Compatible with Claude Code, Codex, Cursor, Gemini CLI, Goose, Copilot, and any agent that consumes the SKILL.md standard.
Layout: no root SKILL.md (the README is the discovery surface). Skills are bucketed by single-noun category at skills/<category>/<name>/SKILL.md (depth-2). skills.sh discovers depth-2 by default — requires CLI ≥ v1.5.8 — and --skill <name> resolves path-independently, so install commands don't reference the category.
Quick install
# List everything in the monorepo
npx skills add broomva/skills --list
# Install a specific Tier-2 skill
npx skills add broomva/skills --skill handoff
npx skills add broomva/skills --skill make-spec
# The catalog skill itself (browse-the-inventory surface)
npx skills add broomva/skills --skill skills-catalog
# Install everything (use sparingly — large prompt-budget footprint)
npx skills add broomva/skills --skill '*'
Note:requires skills.sh CLI ≥ v1.5.8 (PR #1272), which discovers category buckets skills/<category>/<name>/SKILL.md at depth-2 by default — on both the local and remote/blob install paths. No flag is needed (that's only for depth-3+); buckets are exactly one level. On older CLIs the bucketed skills are not found — run or upgrade. resolves by frontmatter , so it is path-independent (the category never appears in the install command).
The monorepo. Skills bucketed by single-noun category at skills/<category>/<name>/ (depth-2), each with SKILL.md + optional references//scripts//assets/ per the agentskills.io spec. Includes the catalog skill at skills/tooling/skills-catalog/, and one skill that also vendors the runtime it drives at skills/simulation/parallax/runtime/.
bstack primer — the agent-readable contract for the Broomva Stack's twenty automation primitives (P1–P20) that turn an agent-driven workspace into a self-operating system (the primer skill; the bstack CLI is a separate clone + bootstrap product)
Audit and rightsize a context surface (CLAUDE.md, AGENTS.md, SKILL.md, prompts) against the Claude-5 context-engineering reversals — measures token budget, hard-rule ratio, duplicate sections and contradiction candidates, then adjudicates keep / relocate / delete by whether an independent mechanism already enforces the prose
Build or adversarially audit an evidence-first legal-readiness system for a software product, SaaS, AI app, API, marketplace, or website. Inventory every public and contractual claim; determine…
The content contract for a design doc — what must be IN it, as against make-spec which owns how it LOOKS. Synthesised from five primary sources read verbatim (Lynch/Refactoring English + his worked…
Unattended multi-lens review of a plan, launch plan, spec, strategy doc, ADR or proposal. The autonomous counterpart to grill-with-docs: instead of interviewing the human, the agent asks the grilling…
Run an unattended arc to a finished, clean end state as one invocation (/loop 30m /arc): artifact-checklist goal, fleet snapshot, research-instead-of-ask, fan-out by shape, cleanup end state
Use when the user has agreed on a plan or selected from suggested options and wants the agent to execute the work autonomously without further instruction
Turn any work-queue + enforcement pipeline into a self-driving, self-healing, human-minimal autonomy loop with a control-systems safety envelope — a metacognitive governor that drives isolated arcs and never performs the irreversible act itself
Terminal-message contract for an autonomous arc that has run out of agent-executable work and genuinely needs a human — a work order of imperative asks, each carrying a default so silence is never fatal
Restore an arc killed mid-flight by an API error, a dropped connection or an expired login — reconstruct what died from the transcript, recover dead subagents' work off disk, and continue rather than conclude
Claude subscription account manager for macOS (Claude Code CLI, Orca, Keychain): account discovery, usage telemetry, one machine-wide balancer that fails over without killing running sessions, safe…
The shared context core, phase 1: a read-only shared board. Hooks append each session's start, ARC-STATUS keyword and death to a per-scope log as structured fields only; SessionStart injects a factual brief of the other sessions on the same branch. Coordination only, not a security boundary
The Broomva fleet: every Claude Code and Paseo session, PR and loop the coordinators drive, and the workflows that manage it. reconcile is the hourly launchd tick (observe, classify, report, owner…
Audit Codex, Claude, Gemini, and Cursor token/cost traces plus Antigravity quotas; emit JSON/CSV/text or a self-contained HTML insights dashboard—no CodexBar runtime dependency
Rewrite a requirement so it can only be read one way — distilled from ASD-STE100 Simplified Technical English, with a stdlib checker for the mechanical ambiguity
Scaffold a substantive human-readable design doc (spec / plan / ADR / report / PR explainer) as native HTML using the workspace's canonical Broomva dark theme
Change or extend an app, website, game or device you own but whose source you don't control, by the cheapest rung that reaches the idea; the running original is the oracle, a stall breaker forces a re-rank, and every run ends in a linted field note. Ships a macOS app recon adapter and a worked Obsidian plugin
Verify a claim whose evidence is an ABSENCE — pairs every denial with a positive control that must succeed, because "everything is denied" and "nothing ran at all" are the same observation; returns INVALID rather than PASS when the controls did not fire
Find absence-assertions that carry no attempt-record — code returning the same empty value whether the work ran and found nothing or was skipped entirely; zero-config AST tripwire that names what it could not audit
Takes raw unstructured thoughts, voice transcript dumps, or stream-of-consciousness text and auto-files them into the right Obsidian vault folders with tags,…
Reversible payload compression — shrink any blob (tool output, log, RAG chunk, file) before it enters context, while caching the original locally for byte-exact retrieve(handle) on demand. The payload-axis counterpart to kg's retrieval axis. Content-routed deterministic compactors (JSON skeleton / code outline / text head-tail), stdlib-only, no ML
Knowledge engine over the Colombian Truth Commission report Hay Futuro Si Hay Verdad (2022) — findings, statistics, actor responsibilities, differential harms, lexicon, and the 67-recommendation non-repetición roadmap, with a kg/LLM-wiki retrieval engine and an align policy-vs-roadmap scorer
Explain the concepts a session used, at the operator's register — ranked by what blocks understanding, anchored to where each term appeared, grounded in the knowledge graph; degrades to a re-pitch when the slice is short
Ingest-and-integrate an artifact someone points at with a terse, deliberately under-articulated directive — "check this out", "lets research this", "look into this",…
Work the way "The New Baseline" playbook describes: every delegated piece of work runs under a five-term contract (intent, bounds, evidence, escalation, review), attention is rationed to what can be…
Takes one set of project facts and generates three versions: technical for engineering, business-impact for leadership, and customer-facing for success teams
Produce high-retention, dopamine-aware video content using brainrot editing techniques — fast cuts, word-by-word captions, sound design, visual velocity, pattern…
Full-stack AI content studio — orchestrates visual DNA compilation, cinematic generation (via Higgsfield CLI or MCP), browser-automated tool execution, and…
Produce polished product launch videos using the Liquid Glass aesthetic — dark void backgrounds, 3D perspective floating UI panels, particle effects, spring…
Directed, reproducible code-first motion: a render-contract linter, a Playwright freeze-frame renderer with a labelled contact sheet and purity check, and drift-free motion math (beats, springs, GIF timing)
Algorave-grade livecoded music workflow — TidalCycles patterns (Haskell DSL driving SuperDirt over OSC) + Hydra-synth visuals (browser or VS Code Simple Browser via a…
Local TTS, voice cloning, voice design, and video dubbing via the OmniVoice Studio MCP server (open-source ElevenLabs alternative; nothing leaves the machine, runs on…
Speak an explanation out loud while working in any project — tiered text-to-speech with a pluggable backend (ElevenLabs by default and quota-guarded, macOS say via --fast for free instant local…
Distill a visual style from reference sites/products into a validated dual-mode design system and a ready-to-run Claude Design handoff — a composition skill that fires existing tools in sequence rather than reimplementing them
Premium design engineering skill for agentic workflows — produces high-end, distinctive UI designs using DESIGN.md as the portable contract across Pencil MCP (in-IDE…
Remove the vibecoded / AI-slop look from any frontend codebase at the root, autonomously — full-repo survey with root-cause attribution, substance tells (legal, loading states, placeholders, real product evidence), and a deterministic crafted-floor gate; composes the impeccable detector
Build or restyle anything in Liquid Midnight. The look: Apple Midnight blue-black grounds (hue 245-252, never purple), liquid glass rimmed in travelling liquid metal (a conic chrome edge with cyan…
Build or restyle anything in Liquid Dawn, the calm daylight sibling of Liquid Midnight. The look: warm paper grounds (Dawn, primary) and a sea-slate Dusk mode; nature drawn in type (lagoon caustics…
Provision and harden a fresh Linux VPS into an autonomous-agent dev host using the capability-preserving model — the box IS the sandbox: full agent autonomy inside it…
Give an agent session a disposable sandboxed bash shell whose entire filesystem is one JSON file — snapshot, fork, branch and resume a workspace at ~0.24ms per fork instead of seconds per container
Set up Claude Code messaging channels for Telegram and Discord — bot creation, plugin installation, token configuration, access control (pairing, allowlists, guild…
Per-channel remote sessions for Claude Code via Discord and Telegram — each channel, thread, or chat gets its own isolated Claude Code session via tmux, with…
Software-defined radio (SDR) and satellite reception toolkit — what to install, what you can hear from space, and how to compose the open-source stack (SatDump,…
Use Meta's TRIBE v2 brain encoder to validate cortical alignment of AI model representations (LLaMA, V-JEPA2, Wav2Vec, or any encoder) and inform model selection in…
Personal health knowledge graph — local-first ingest of Garmin (Apple Health, Whoop, Oura, CGM in v2+) traces into SQLite, projected to Obsidian daily-note…
Comprehensive bioinformatics and ocean genomics skill for eDNA metabarcoding, metagenomics, protein structure prediction, and marine biodiversity analysis
Turn a question about the world into a table Parallax can accept, with every field typed observed or simulated at the moment it is written. You search; it records findings against the artifact they were read from, hashes it, judges each column by rule, and emits the exact parallax propose invocation
Drive Parallax — the ontology simulation layer: propose an ontology from a context, a human accepts it before anything runs, then roll it forward with every answer typed observed or simulated. Routes state to the next command and a typed refusal to its remedy, for all 52 error codes. Ships the runtime it drives at runtime/, so installing the skill installs the layer
Crawl a set of sites into a verified entity graph — companies, the people who run them, the relationships between them — where every claim points at the bytes it came from and nothing expands until…
Shop Tiendas D1 (Colombia) from the command line — search, resolve your nearest store, price a basket against its real stock, quote delivery; builds carts but never pays
Find the cheapest fuel/GNCV near a Colombian location — engine over live per-station open data (SICOM GNCV via datos.gov.co) with an explicit freshness verdict and honest coordinate resolution (municipal-centroid distances shown approximate)
Estimate optimal waiting time for Alkosto's "every 25/50 customers" promotion using either checkout-flow observations or winner announcement timestamps
Catalog inventory
The 111 skills bucket into 23 single-noun categories — one row per skills/<category>/ directory:
See CONTRIBUTING.md for both flows: the root-level catalog flow (adding inventory entries, updating the showcase) and the Tier-2 monorepo graduation flow (adding a skill at skills/<category>/<name>/).
