Every principle, skill, plugin, sub-agent, connector and hook this machine loads, under a two-tier model: one global operator layer every session inherits, and project layers that defer to it. Every number is probed from the live machine at build time, never typed.
Slug-named, one idea each, distilled from AI tools in use. A reference to pull from on demand, never pasted wholesale into a prompt. Source: ~/Projects/dotfiles/docs/AI-Principles.md.
Slugs replaced P0NN, so a citation can be sanity-checked where it sits and a merge needs no tombstone stub. The two concrete instructions were carried verbatim: derive from immutable sources and freeze the system for one full cycle. Seven files cited principles by number, not the four the proposal predicted, and all seven were re-pointed in the same change.
The persona and epistemics loaded into every session, across code repos and context vaults alike. Project files defer here for operator-level rules.
Auckland. The Creators Win (B2C mentorship, 35% profit share), WIN·WIN MEDIA (B2B UGC agency, 50%), and DHC advisory. Each client has its own vault.
Shortest answer that is still true. Bullets, no preamble, no recap. Fewest tokens that achieve the goal, then stop. Report a finding; don't build the fix unasked.
Current state only. No superseded sections, no strike-throughs, no evolution notes. The past lives in git and chat, never in the document.
Every non-trivial claim tagged confident / best guessing / wildly speculating. Don't round up to certainty; the wobble is information.
A delivery layer under the operator layer: it governs how Claude talks, not how it codes. Installed from alexgreensh/attention-span, switched with /style. The active style costs roughly 900 input tokens per session, cached after the first request (installed range 580–1820).
Terse, warmth stripped, maximum compression. Arrow markers, bold carries the whole answer, no cushioning.
The flagship. Answer first, plain English, expands only on what would cost you a mistake. Warm where Spartan is blunt.
Briefing mode. TL;DR, then a status checklist and tagged next moves. Optimises for scannability over brevity, so it works for standups and fights the shortest-answer rule elsewhere.
The global CLAUDE.md already mandates short, bullets, no preamble. A style adds what prose rules cannot enforce: a format contract. It reaches the main conversation only, so sub-agents answer in default prose.
What this stack can actually do, grouped by the job rather than by how it is delivered. The marker says what backs each one: S skill, P plugin, M MCP server, A sub-agent, C claude.ai connector. Filter the band to see one kind at a time.
Repo standard and docs gates, design and accessibility passes, then a typed Next.js build onto Vercel.
DNS and registrar moves, Workers and Durable Objects, KV/D1/R2, email routing, Turnstile. The one surface with both skills and live MCP behind it.
Stand up a vault, run the memory cycle, turn meetings, docs and Slack into wikilinked summaries. The plugin half is explicit-invocation only.
Connect configuration, billing and tax, key security. Scoped per repo, never global.
Verify a claim against real papers, audit whether existing citations say what they are cited for. Metered: one credit per fresh run.
Architect layered system prompts, audit against the checklist, adversarial verification, coach a single outgoing prompt.
Mail, calendar, files and the Notion studio, plus the WhatsApp surface behind the concierge product.
Keyword sets, landing-page copy, and finding a skill that already does the job.
Fan out read-only searches, plan an implementation, answer harness questions. The conclusion returns to the main thread, not the file dumps.
22 of the 70 plugin skills are explicit-invocation only, so the model never offers them; only you typing the command reaches them.
Connectors are a dated claim, not a probe — last checked 2026-08-20. Slack and VoC Sales are gone from the last session checked.
A second plugin surface exists that this build cannot see: claude.ai keeps its own list, and sessions have loaded ecomrads-studio and cowork-plugin-management skills that appear in no local file.
Authored but not enabled: mohios-orchestrator sits in the marketplace directory and no session loads it.
Built-ins are declared, not probed: ~25 harness skills and 6 sub-agents are not enumerable from disk.
The parts that run without you asking, and the parts that only look like they do.
SessionStart → sync.sh
Stop → stop-gate.mjs
Stop → memory-autopush.sh
Stop → stack-drift-hook.sh
UserPromptSubmit → style-reminder.sh
One fact per file, indexed by MEMORY.md. Recall surfaces relevant facts as background context each session. A Stop hook pushes it to git at the end of every session.
The directory holds 6 tasks. 2 are registered and enabled. 0 are orphans: a SKILL.md on disk that the scheduler has never returned.
The scheduler keeps its registry where this build cannot read it, so a task directory can look exactly like a live job while nothing runs it. This page now fails to build if a directory appears that the data file has not classified.
Each defers to the operator layer and embeds website-build-rules v2, re-synced from one source and never hand-edited. Worktrees fold into their parent. The public ones run as the portfolio at mohios.com: chatbot · dashboard · menu · stack (this page) · tcw roadmap · angelic reiki.
The portfolio’s double proof, live at angelicreiki.nz. One: Vercel + Resend replace Squarespace and WordPress outright (same pattern the Hostinger Exit case study documents). Two: the website IS the business back office — customer portal, bookings, Stripe payments and packages, CRM, finances and accounting, run through AI on the same stack this page describes.
Mohios code-audit engagement: 45 repos across interview, GTM and product assets.
Embeddable Mohios chatbot widget, the lead-capture engine behind the sites.
Claude-with-code workshops: agent-battle, decomposition, eval-driven dev, rightmodel, and more.
Personal and consulting site, off Hostinger entirely since 2026-08-14: site on Vercel, DNS, mail and registration at Cloudflare. Renders the ported WordPress copy verbatim until the repositioning pass.
Canonical backup of global Claude config, and the home of this page. The live ~/.claude files symlink in here.
Desktop and IPC tooling, with a scoped WhatsApp MCP.
Singapore holding-company page. A digital brass nameplate, intentionally nothing to do.
NZ small-biz AI automation site: AI receptionist, WhatsApp concierge, lead-capturing chat.
NICH Cycling NZ marketing and lead-capture site, built for Kevin Cantwell.
Open-source orchestration for teams of AI agents. Symlinks skills into the global set.
Most of the global skills, the hook scripts, the launchd plist and the website build kit.
The Creators Win (b2c mentorship) plus WIN·WIN MEDIA (b2b growth studio): one repo, host-based routing, two apex domains.
Public-but-noindexed mirror of the TCW roadmap at tcw.mohios.com, resynced from tcw origin/main every Friday by a scheduled task.
The Website Menu at menu.mohios.com: 65 website elements with wireframe visuals, business value, live examples and effort tags. The scoping menu for client conversations.
WhatsApp MCP surface feeding the concierge product.
Client-ready records of work this stack shipped, written up in the DHC vault first and published second. The build fails if a vault record named here has moved, so a card can never point at nothing.
A consulting site off WordPress and off its host in one working day — website, DNS, email and the domain registration, four vendors, not one. Zero downtime, no URL changed, annual cost USD 95.88 to 10.46. The sellable half is not the HTML: it is the twelve pre-existing faults the record-by-record pass surfaced (two SPF records, so no SPF; no DMARC; an AAAA record that would have left every IPv6 visitor on the old site) and a cutover order that never risks the mailbox and the website in the same operation.
read the artifact · ~/Documents/ai-context-dhc/Curated-Context/Knowledge/Case-Studies/20260814-Hostinger-Exit-WordPress-Migration.md
GitHub Actions billed ~1,800 minutes a month across the org; rebuilt per-run from the API, about a third of it was rounding on jobs that finished in 13 seconds, because billing rounds each job up to a whole minute. Six repos folded to one job each with filtered triggers, and the fault was fixed in the build-kit template so the next site does not re-buy it. Now ~550. Two repos also had a docs gate and no build gate, which is worse than no CI: the green tick reads as "this builds".
read the artifact · ~/Documents/ai-context-dhc/Curated-Context/Knowledge/Case-Studies/20260822-CI-Minutes-The-60-Second-Line.md
Pushback from the current pass. Replaces the previous set rather than sitting beside it.
Every count is probed at build time; the prose file contains no numbers at all. Claims that cannot be probed carry a visible date instead of pretending to be facts.
The paperclip skills were unlinked from ~/.claude/skills on 24 Aug; they still sit in the repo, so nothing looked broken. They resolve to a warning, not an error, which is why the page shipped dead chips until this pass.
The build fails on a scheduled-task directory the data file has not classified, which catches the next orphan. It cannot catch a task that gets disabled tomorrow, so a green build is not proof the schedule is what you think.