Shelving the masterplan
Seven phases, sixteen days of commits, 27 posts — then a quiet summer. The masterplan is shelved. What it produced is still running, and the toolchain underneath it doesn't look like May anymore.
The masterplan is shelved. Postponed, not abandoned mid-sentence — the parts that work are still running unattended, and the parts that never started are being left unstarted on purpose. Seven phases, sixteen days of commits, 27 posts on this site, then three and a half quiet months.
It reads like a failure and it isn’t. A plan for a personal build has one job: make the thing exist. This one did that inside two weeks, wrote everything down while it was happening, and then got out of the way. The learnings were the valuable part, and they’re the part that survives.
What’s still running
The homelab never got rolled back. n8n, Penpot, CouchDB and Browserless sit on the GMKtec. The vault syncs with LiveSync, and the session ingester still turns Claude Code transcripts into notes without me touching it. The publish pipeline is intact: tick a checkbox in the masterplan, and Groq drafts a post, Unsplash supplies a cover, and a PR lands here — which is how every post since May arrived, including the ones I edited afterwards.
What changed underneath it
The plan assumed local models. That assumption died first.
- Ollama (qwen2.5:7b, qwen2.5-coder:14b) from all automated workflows
- Claude API in the critique loop
- Sanity as the CMS path
- Groq qwen3-32b for critique, Mistral as the fallback arm
- DeepSeek in the design feedback path
- Hermes Agent in the terminal alongside Claude Code
The A/B test is what settled it. Four challengers beat Claude on critique quality — Groq qwen3-32b at 9.86 spec+actionability, Mistral at 9.00, local 14b at 9.20, all with perfect brand-violation catch — so Claude came out of the automation and stayed where it’s actually strong: interactive sessions. Hermes was an arm in that same test in May and failed the Penpot task outright. It’s now the terminal agent doing work like the SEO pass and the CI gate on this site.
Claude Code is still where code gets written. The difference is it’s no longer the only thing in the room.
What never got built
Phases 6 and 7 — RAG over the vault with Qdrant, and the multi-agent critic loop — never started. Phase 5c was handoff documentation at scale: the batch driver benchmarks at 50–75 seconds per component, so 500 components is an overnight job that never got scheduled. The design pipeline works, and brief-to-Penpot-frame runs under 15 minutes, but it stayed a proof of concept.
What this unblocks
Nothing on the schedule, which is the point. The plan was useful for exactly as long as it was forcing decisions, and it stopped forcing them somewhere in June. With the checklist out of the way, the posts on this site stop being phase updates and start being whatever the week actually produced — the SEO pass and CI gate from this morning being the obvious example. Claude Code, DeepSeek and Hermes in one workflow is a better setup than anything seven phases predicted, and I’m looking forward to what we create with it next.