Production AI systems and workflows built across Claude Projects, GPT Projects, and Anthropic API deployments. Brand voice and campaign production. Market intelligence with kill-switch discipline. Autonomous multi-channel marketing. Satellite image analysis for mineral prospectivity. Every system keeps a human in control of the decisions that matter.
// Skills & Guardrails
I’ve got skills.
Skill files and Markdown guardrails steer frontier models to reliable output and automate work at scale, from copywriting to proposals. Touch a name to scatter it. Four I share the most.
Brand Voice · Copy System
Sunset Marquis Copy System
The operating voice for Sunset Marquis. Speaker, pillars, banned moves, fact rules, and channel guidance that keep AI-drafted copy on brand and accurate. The same system behind the hotel’s campaigns.
A playbook for building an AI-native company. MTP constitution, agent fleet, workflow interfaces, live dashboards, governance, and unit economics, from idea to operating model.
A disciplined system for fundable proposals. Funder fit, evidence labels on every claim, reconciled budgets, portal-ready answers, and a human review gate. Never fabricate, never submit.
Turns a working prototype into a launchable product. A 13-layer readiness scorecard, P0/P1/P2 triage, security gates, and a hard definition of done. A working demo is not a finished product.
4Operating stages: audit, field study, AIO build, tracking
GSCSearch Console: clicks, impressions, CTR, position
LiveRank feed: found, pending, improvement states
Problem
Small teams need SEO that works for classic search and AI answer systems, but the work fragments across plugins, keyword tools, spreadsheets, dashboards, and someone guessing what to change next.
System
Singularity SEO turns ChatGPT into the control room for a connected WordPress site: managed pages, rendered metadata, technical health, GSC visibility, and live rank checks. ChatGPT audits the site, studies competitors, and builds page-level SEO and AIO recommendations.
Workflow
Connect WordPress, verify the site, run a read-only baseline, then review competitors, choose keyword lanes, produce one governed full-site proposal, apply approved updates, and preserve rollback-ready change IDs.
Outcome
SEO becomes an operating loop instead of a plugin screen: audit, compete, optimize for search and AI answers, measure the result, then improve again.
I build my production agents by hand, then wire them into the tools a business already runs on, like Zapier and Telegram, so the work shows up where the team already works. Narrow agents, bounded authority, and a human on the decisions that matter.
03 · Production · Brand Voice · RAG
Agency689 Writing Systems
½×Production time per campaign cycle
3+Viable first drafts per run
3/wkSunset Marquis runs in active cycles
3Active client brands in production
Problem
Agency work creates constant demand for strong first drafts across email, banners, ad creative, landing pages, and social, with zero tolerance for voice drift. Generic AI output is the enemy.
System
Brand context files, RAG-retrieved voice samples, SKILL.md governance, and persona specs are assembled before any draft. The system produces multiple drafts, routes them through a standards AI for voice and quality, then to a human for approval via Telegram. Nothing ships without sign-off.
RAGSKILL.mdTelegramStandards AI
04 · Production · Stock Intelligence
VRT2 · CLAW
27Data streams ingested daily
23Active hypotheses across 6 tiers
5Verification states per signal change
20%Drawdown fires a system-wide kill-switch
System
Node.js + SQLite backbone. Claude Haiku handles routine daily reviews; Claude Sonnet handles advanced analysis. Ingests Finnhub, EDGAR, earnings transcripts, analyst revisions, and macro feeds daily, synthesized into structured intelligence briefs.
Discipline
Every signal moves UNVERIFIED → SYNTAX_OK → SQL_VERIFIED → RUNTIME_VERIFIED → PRODUCTION_VERIFIED. New signals run at 0.50× weight until proven; sub-55% hit rates auto-deactivate; a 20% drawdown fires the kill-switch. Nothing trades automatically.
A single Claude brain reads book metadata and generates platform-specific copy for all six channels in one run. OAuth2 and OAuth 1.0a handle auth server-side. KDP ranking data feeds back in via Rainforest API for attribution learning.
What it is
The open-source Lite release of the internal Agentic689 book and music marketing system. One book, one author voice, one Claude brain, six platforms. The Lite release strips to the essentials so the architecture is readable.
10World-class gold deposits as training fingerprints
5×5Max search grid across 100-mile radius
System
AuScan trains Claude Vision on the spectral and structural signatures of ten world-class gold deposits: Witwatersrand, Carlin Trend, Kalgoorlie, and others. Multi-sensor data is assembled into terrain fingerprints for each deposit.
Validation
In testing it scored a HIGH-confidence hit at Oregon’s Pueblo Mining District, a documented former gold producer, cross-referenced against the USGS Mineral Resources Data System.
What happens when you point these models at something genuinely hard and learn by making them make it. Case in point.
F = ma → Quantum Gravity
18Days from Newtonian mechanics to the frontier
38Reference documents written from scratch
5New results across 3 theorems in Paper I
1Integer-valued exponent gap nobody has explained
On April 23, 2026 I asked Claude where to start with quantum mechanics. No physics degree behind the question. The conversation ran in one unbroken thread from Newtonian mechanics to the frontier. By May 11 it had produced a paper, “Complexity-Sensitive Complementarity in Non-Isometric Holographic Codes,” headed for the arXiv.
