Stripped the site down to its essentials. Cream background, editorial serif, negative space doing the heavy lifting. Inspired by Tadao Ando — the architecture is the experience. Two components, one page, no navigation. Just a business card and a feed.
Went back through the Orchestration Agent, Editorial Agent, and IT Support Agent memory files — markdown docs and JSON configs that each agent maintains — to rebuild a cleaner dated record of what has actually been built since the February launch. The system had grown fast across OpenClaw crons, Discord channels, agent workspaces, and reusable skills, but the documentation hadn't kept pace. Needed one place that reflected the work, not just the conversations around it.
The morning and midday scans for the online publication were doing the work, then dying on thread creation — the cron payload was passing an empty components: [] array that the Discord API rejected. Tightened the JSON config so thread creation only sends the fields the tool expects. Also split the model routing: moved the Editorial Agent's gate-approval sub-agents to Gemini 3.1 Pro via OpenRouter while keeping scan crons on Gemini Flash. Speed and judgment each got the right lane.
Reworked the online publication's cron automation in OpenClaw around thinner, file-first JSON payloads. Less prompt logic inline, clearer source files on disk, tighter control over what each cron job could post and where it could go. The pattern: point the agent at a file path, not a paragraph of instructions. Daily scans, the What's Happening build, and the weekly digest all moved to this model.
Tightened the online publication's publish stack and moved platform-specific copy closer to the queue file itself — structured as a JSON payload that the publish script reads directly. The work is easier to trust when the source of truth lives with the file, not scattered across Discord threads and agent summaries. Separately, draft-length tuning finally landed. The online publication's output started hitting the intended editorial range consistently. That was one of the last pieces keeping the system in build mode instead of polish mode.
Added a building log to the site. Each post is a markdown file with frontmatter. No CMS, no database. Gray-matter parses the frontmatter, remark converts the body to HTML. Everything runs server-side. Low friction by design — the best system is one you actually use.
The visible instability turned out to be a mix of reconnect pressure, auth drift across the Google Workspace APIs, and restart churn in the OpenClaw runtime — not one big system failure. Cleaned that up and got the machine behaving like one machine again. Also patched the Daily Briefing cron so a single broken quote lookup or auth timeout would not kill the whole run, and repaired the weekly ops audit after shell command drift started breaking the IT Support Agent's checksum step.
Ran a signal all the way through the online publication's pipeline — from the OpenClaw scan queue through AI-assisted drafting, image generation, editorial packaging, to final publish on Ghost. That mattered more than another clean architecture document because it proved the full system could actually ship end-to-end.
The online publication's visual language stopped being a moving target. Typography, palette, format rules, and the distinction between cover treatments and scene drops finally had one home in the brand docs. Deployed the v3 image pipeline in parallel — tightened the framing logic, locked the output structure, and pointed everything at woostermag.com as the canonical domain. The brand went from mood board to operating spec.
Created an achievements log — a simple append-only markdown file timestamped by date — so the work would stop disappearing into Claude Code session history and Discord threads. If a system is being built, the build itself needs a ledger.
Expanded the model fallback chain through OpenRouter: GPT 5.4 as primary, then Gemini 3 Flash, Gemini 3.1 Pro, Claude Sonnet 4.6, and Claude Opus 4.6 as the final backstop. Each model is selected by cost, latency, and capability for the job at hand. The goal was not novelty — just a more durable operating path so a single provider outage would not stall the whole environment.
Stopped treating the resume site like a digital PDF and started treating it like a product. Built it in Next.js with Claude Code as the primary dev tool — describing intent, reviewing output, iterating in conversation. The structure got cleaner, the hero got sharper, and the page started reading more like a point of view than a file.
Spending time with Claude Code as a dev tool. The workflow is natural — describe, review, iterate. What I find most useful isn't the code generation. It's having something that holds full project context while I think through architecture. The conversation becomes the design doc.
The online publication needed a visual center, not just a mood board. Locking the Ralph character direction — a consistent editorial mascot generated through the AI image pipeline — gave the brand something it could actually build around. Then pushed a MIDiA Research signal through the full publication pipeline: scan, research, AI-assisted drafting, image generation, and final packaging. That run helped expose where the workflow was still theatrical and where it had become real.
Updated OpenClaw to 2026.3.13, ran the doctor repair, and increased key cron timeouts to 900 seconds in the JSON config for the Daily Briefing and Tech Ops EOD jobs. Cleared the post-update drift that had started showing up in delivery behavior — agents timing out before their work completed. Not glamorous work, but the machine runs on this kind of maintenance.
Handled the OpenClaw update to version 2026.3.11 — ran openclaw doctor --fix, verified the live cron store in the JSON config files, and stopped trusting stale terminal warnings. The IT Support Agent, responsible for Technology Operations, moved from concept to real operating lane: live cron ownership in OpenClaw, Discord reporting surfaces, and a defined job scope. Also restored Apple Messages access via AppleScript and built a Leon-only iMessage bridge with its own reply channel in the agent layer. A lot of good systems rot because nobody does the unglamorous work after the version number changes.
Packaged three recurring workflows into reusable OpenClaw skills — agent onboarding, docs research, and workspace cleanup — so any agent in the system can invoke them without rebuilding the logic. Also split the one-shot Weekly Drops flow into separate build and enqueue cron steps after it kept failing end-to-end. Simple rule that keeps proving true: when a process gets fragile, separate the thinking from the side effect. Repetition usually means the system wants a tool, not another reminder.
Named the Technology Operations function, defined its charter, and gave it a formal place in the Vantage org structure — the multi-division company that sits above the Orchestration Agent, the online publication, and the other operating units. Once it had a name, it could stop being miscellaneous cleanup and become actual owned work. First move: retiring the broad heartbeat monitoring cron and replacing it with a tighter exceptions-only model routed through a Discord #maintenance-log channel. Better systems usually say less, but say it when it matters.
Cleaned tracked runtime artifacts out of the Git repository — browser localStorage dumps, cached API responses, stale session files — and rewrote the history to bring it back to a sane size. The project got lighter the second it stopped pretending browser state belonged in version control. On the editorial side, a stale news miss in the online publication's scan led to harder date checks and source-freshness rules in the agent workflow config. If the entire product depends on signal quality, the freshness standard cannot be optional.
Built the board-meeting checker, daily rollup, and report-ledger structure inside Discord so reporting could become a system instead of a habit. Each agent files a structured end-of-day report; a cron job validates format and completeness; a rollup posts the summary to #board-meeting. Also added a direct thread-state helper for the online publication's signal board — the forum got cleaner once thread status was controlled through the OpenClaw agent layer, not left to Discord defaults. Governance starts when visibility becomes repeatable.
Built the Editorial Agent workspace inside OpenClaw, loaded the operating docs, and got the first real scan working for the online publication. The Editorial Agent runs the online publication — an editorial media division — using Gemini Flash through OpenRouter as its model gateway. The scan scored incoming signals from RSS and industry sources, wrote backlog files to disk, and created forum threads in a Discord channel called #signal-queue. That was the moment the online publication stopped being an idea and started behaving like a division.
Looking back on the Built By Seven years — taking an artist from unsigned to a major label deal remains one of the most fulfilling things I've done. It wasn't a discovery story. It was a build story. Grassroots marketing, digital strategy, positioning, and then sitting across the table negotiating the deal. Most people in administration never touch the creative side. I've always believed the best operators understand both.
Started with a fresh OpenClaw install — an open-source agent orchestration framework that manages cron jobs, memory, and tool access through JSON config files — and wired together the basics: Gmail and Calendar via Google Workspace APIs, a Discord server for agent communication channels, and a persistent markdown-based memory layer. The Orchestration Agent is the operating system I'm building around my work. A small start, but the first day it became real.
Deep in the Curve implementation at gamma. The challenge isn't the software — it's mapping bespoke deal structures into a system that expects standardization. Every artist agreement has edge cases. The work is in building a translation layer between contract language and system logic that doesn't lose fidelity. When this is done, it replaces an entire manual process.
There's a particular kind of work that doesn't come with a playbook — building a department where none existed. At gamma, there was no royalty infrastructure when I arrived. No system, no workflows, no institutional knowledge to inherit. The mandate was simple: figure it out and make it work. That's the kind of problem I'm drawn to.