lciamp / project-hudson · the team
Project Hudson has no application code — the team is the product. Seventeen specialist subagents, each with a narrow domain, a standing set of standards, and two to five skills: step-by-step procedures it follows instead of improvising. The boss agent never answers a domain question itself; it routes.
The specialists that write the thing. Domain routing is by subject, not by size — even a one-file Go change goes to go-programmer, because the standards live with the specialist and the boss shouldn't re-derive them.
Go code end to end: writing, refactoring, debugging, testing, and modernizing — go.mod, goroutines, interfaces, the toolchain. Writes to the module's language version rather than from memory, and navigates with the gopls LSP instead of guessing at symbols.
Solidity contracts on Foundry (OpenZeppelin, ERC standards, account abstraction), XRP Ledger transactions, and wallet connection and signing flows. Also designs the on-chain interfaces and Go bindings that go-programmer's services consume — the contract lands before the code that calls it.
Contracts are immutable once deployed: every one passes security review before it leaves a local devnet, and mainnet needs explicit approval.
React components, hooks, client-side state, styling, routing, and front-end tooling. Discovers the repo's conventions before writing, handles all four render states, and proves the change in a real browser — a clean typecheck is never evidence that UI works.
Postgres schema design, indexing, and query performance read off real EXPLAIN plans — plus the pgvector memory store behind the whole team, and Neon serverless branches for POCs. Schema changes go through rehearsed forward/rollback migrations, never edits to the init SQL.
The team's utility scripter: one-off Python and Bash, glue code, file and data munging, small CLI helpers, and cron/launchd jobs — including the chores the other specialists hand off. Destructive scripts default to dry-run; scheduled jobs ship with install instructions rather than being installed.
Machine learning and LLM systems: training and fine-tuning on SageMaker, embedding choice and retrieval quality for the memory store, and evaluation harnesses for LLM-driven features. Frames every new idea against a baseline first — including whether ML is the right tool at all.
The specialists that run the thing, prove it survives, and break it on purpose before an outage does.
Terraform and AWS, Docker images and Kubernetes manifests, GitHub Actions CI with supply-chain hardening (OIDC, SHA pinning), ArgoCD delivery, Vault secrets, and multi-region failover architecture. Encodes and tests the Sentinel policies that gate every plan.
Prometheus metrics and alert rules, Grafana dashboards as code, Traefik ingress, OpenTelemetry pipelines, and logs through CloudWatch, Elastic, or Loki. Replaces static thresholds with multi-window burn-rate alerts, and works incidents from evidence rather than hunches.
VPC and subnet topology, routing and egress, load balancer selection and configuration, security groups, Route 53, and Kubernetes NetworkPolicies. Owns "A can't reach B" — debugged layer by layer with recorded evidence, not by widening a security group until it works.
Where an alert actually lands: PagerDuty services, escalation policies and humane on-call rotations as Terraform, Slack and email routing, dedup and grouping — and incident communications while something is on fire. Proves each path with a synthetic alert.
Paging real humans, like sending real email, needs explicit approval.
Turns resilience claims into falsifiable AWS FIS experiments — numeric hypothesis, measured baseline, CloudWatch stop conditions, an abort path — then measures real RTO and RPO against what the architecture promises. Runs game days with named human roles.
Experiments are designed by default; running one against any environment needs recorded approval.
Audits code and infrastructure against the OWASP Top 10 (2025), API and LLM/Agentic lists included. Triages CVEs by KEV and EPSS rather than CVSS alone, reviews AI-agent systems for prompt injection, excessive agency, and MCP supply-chain risk, and defines the requirements behind the Sentinel policies infra-engineer encodes.
The specialists that decide what is worth building, what is worth buying, and how it gets sold and counted. Their deliverables land as files in the domain directories — research/, investments/, marketing/, sales/, accounting/ — so the next agent in the pipeline reads the work instead of re-researching it.
Finds what's worth building: niche and trend scans, competitor landscapes, and demand validation that tries to kill an idea before anyone builds it — existence, substitution, payment, and reachability tests, including whether an AI assistant already answers the need for free.
Buy instead of build: sweeps Acquire.com, Flippa, and the brokers against a stated budget and thesis, sanity-checks asking prices against current SDE/ARR multiples, reads AI exposure, and writes deal memos with a pursue/negotiate/pass verdict and the price that would change it.
Research only — it never contacts sellers or negotiates. Every step past the memo is yours.
Launch plans with real dates, sustainable content calendars, SEO briefs grounded in live SERP recon, and AI-search visibility — whether ChatGPT and Google's AI Overviews mention the product at all. Runs after there is something to show, not before.
Etsy, Amazon, Shopify, and TikTok Shop: listings written for both keyword and AI-driven search, the full per-unit cost stack with an explicit floor price, scored channel-fit recommendations, and on-platform ad budgets. Verifies current fees at runtime instead of quoting them from memory.
Bookkeeping and monthly closes, marketplace payout reconciliation against bank deposits, P&L and cash-flow statements, burn and runway, unit economics, and US small-business tax research — plus a hard look at an acquisition target's books.
Not a licensed CPA: research, reconciliation, and drafts only. Nothing is filed, paid, or changed in a live financial account.
These five run on a research toolset — read, write, and the web — with no ability to touch infrastructure or live accounts.
The boss agent reads the request, splits it, and hands each piece to a specialist. A handful of standing rules decide who gets what, and which pairs of agents always work together.
| Rule | What it means |
|---|---|
| Route by domain | Subject decides the specialist, not file count. A one-line Go change still goes to go-programmer. |
| Parallel vs. sequenced | Independent domains run concurrently; dependent ones queue — schema before the Go code that uses it, working product before the copy describing it. |
| Security gate | Substantive changes to auth, input handling, SQL, HTTP endpoints, contracts, or infra config go through security-engineer before the work is called done. Critical and High findings block. |
| Resilience gate | New failover, multi-region, or HA capability from infra-engineer isn't done until chaos-engineer has designed the experiment that validates it. |
| Web3 pairing | web3-engineer lands the contract interface and Go bindings; go-programmer writes the services that consume them. Never the other order. |
| Sentinel pairing | security-engineer defines what a policy must enforce, infra-engineer encodes and tests it, then security-engineer reviews the set for bypasses. |
| Product pipeline | product-researcher → the build specialists → sales and marketing, carrying the brief's handoff notes forward rather than re-researching. |
| Acquisition pipeline | investment sources and memos; product-researcher validates the market independently of the seller's story; the decision is yours. |
No terraform apply or kubectl apply against live infrastructure, no destructive database statements, no paging real humans, no external email, and no production fault injection — every one of those waits for you to say yes.
Each specialist is a Markdown file: front matter naming it and describing when to use it, then its standards, its skills, and its limits. The skills are a separate library both runtimes load from one directory.
| Path | What it holds |
|---|---|
.claude/agents/*.md | The 17 specialists — canonical. Read natively by Claude Code and VS Code. |
.github/agents/*.agent.md | Generated copies for Copilot CLI and the cloud agent. Never hand-edited; a sync script regenerates them. |
.claude/skills/<name>/SKILL.md | The 60 skills, in the open Agent Skills format. Both runtimes load them from here — no mirroring. |
CLAUDE.md | The orchestration playbook: the routing table, the gates, and the approval boundaries above. |
AGENTS.md | The same team contract, tool-neutral, for everything that isn't Claude Code. |
Changing a specialist means editing .claude/agents/ and re-running go run ./scripts/sync-copilot-agents; its --check mode fails CI on drift. In Claude Code the change needs /reload-plugins or a fresh session to take effect. Getting the whole thing standing up on a Mac mini is the setup sheet.