b1rdmania

Building in AI. Mostly big data - structuring it, then running agents across it.

I build the data layer as much as the analysis. Replicated datasets, cross-source APIs, MCP servers - then swarms of agents on top for due-diligence and anti-hallucination work. Findings that each trace to a real source, or an honest account of what couldn't be established. No vibes, no confident guessing.

Dig puts it all together: a rebuilt Discogs catalog, served over MCP, agents digging through it.

Ran it across a few domains this year: legal AI (a handful of pilots, last six months), plus e-commerce and music.

Sidelines, mostly for my own use, some published: Wario Style (type any song, get the 8-bit Game Boy version), Boss Whisperer, a stack of Claude skills.

Ship fast, across domains. London.

🏆 WAR.MARKET — winner, HyperLiquid London Community Hackathon (Jan 2026)

▸ GhostClaw · ghostclaw.io — a Claude SDK agent that runs on your own machine. Bare metal, no sandbox, autonomous execution. What it feels like with the guardrails off.

▸ Wario.Style — type any song, get the 8-bit Game Boy version. Four channels of raw chiptune, zero samples. (tap through for sound — GitHub won't play it here) · source

---

Legal AI

A supervised-autonomy workspace and statute-grounded skills that plug into it. Everything here drafts for a solicitor to verify, names the framework it works under, and surfaces the trap before it bites.

Project
legalise · legalise.dev	The workspace. Open-source supervised-autonomy register for UK legal work — every action signed, supervised, provable.
notary	Signed, human-approved execution receipts for any Claude skill. Proof of what an AI was allowed to do, and who approved it, that nobody can forge without the key. Drop-in MCP server, no infra.
pre-motion	Adversarial premortem for E&W civil litigation. Builds the strongest case, then attacks it from four angles to find where it loses.
kramer-v-ai · divorce.broker	AI-guided amicable divorce settlement, money picture to filing pack, human in the loop throughout. A Nicole-and-Keith parody on top, a serious supervised-autonomy demo underneath.

Plus a suite of statute-grounded skills across employment and litigation → all repos

Claude skills

Skill
boss-whisperer · bosswhisperer.fun	Reverse-engineer your boss. Anonymised, multi-framework read — what they actually want, how to manage up. Munger meets Gottman.
bro	Context reset. Settled concern still steering the thread, plan bigger than the problem, still planning after you said go. Say bro. One line, no sorry, back to the actual question.
plain-english	Kills LLM slop. Orwell '46 + Gowers '48 + the modern AI tic list.
premortem	Imagine the plan failed 6 months from now. Work backward. Find every reason why.
skill-auditor	Audits other Claude Code skills. Yes, it's a bit human centipede.
skill-security-audit	Security-audits a skill before you run it. Severity-governed: one Critical means unsafe.
mobbin-claude-rodeo	Find real-app UI references on Mobbin and build to them faithfully. No LLM drift, no stale specs.
brand-skills	Brand identity in Claude Code. Stops LLMs generating the average.
Visible · Live	Tells you if AI can find you online. If not, fixes it.

Also

Music-tech (Dig, ghost-pattern), a generative London garden, and DeFi (WAR.MARKET, compost) → all repos