You were handed advice: let Claude plan and review, and let cheap helpers do the chopping. The idea is right. The kitchen it assumes isn't the one you have.
Each ticket opens one at a time. Plain line on top; the checkable fact underneath.
Part 5 of the AI Workflow series build log
A head chef shouldn't dice onions. Paying top rate for renames and boilerplate wastes the plan's weekly allowance on work a junior could do.
Anthropic support: "Opus costs several times more per turn than Sonnet, and Sonnet more than Haiku." Max plans have a weekly limit across all models. Before this, ~/.claude/settings.json ran every session and subagent on Opus at xhigh. Helpers now default to Sonnet.
The part to keep: someone decides, someone does, and someone checks before anything is served.
This already runs here un-generalised: the game-demo gauntlet, a weekly review of a collaborator's merges, ada-stair's planted-fault reviews.
The advice assumes a big prep kitchen. This one has a single small counter: the graphics card can hold a short recipe card, not the whole cookbook a coding agent needs open.
RTX 3070 Ti, 8 GB VRAM, ~1.2 GB held by the Windows desktop (no iGPU on the i7-12700KF). Ollama: coding agents need "at least 64000 tokens" of context. Named models: qwen3-coder:30b 19 GB, glm-4.7-flash ~23 GB VRAM at 64k, gpt-oss:20b 14 GB.
You've already tested this. In February a mid-size local model on this card fumbled its tools and drifted into other languages.
TranScripts/Claude_Edited/MoltBot/model-upgrade-plan-2026-02-01.md: Qwen2.5 14B failed tool calls; Llama 3.1 8B became primary.
Pointing Claude Code itself at a local model is off the menu too: unsupported, slower, and it can step outside your subscription.
code.claude.com/docs/en/llm-gateway: Anthropic "doesn't support routing Claude Code to non-Claude models through any gateway." Ollama's Anthropic API has no prompt caching.
Your subscription comes with a whole brigade, not just the head chef. Haiku and Sonnet cost the plan less per turn and follow instructions reliably.
API price ratio Opus 5.5 : Sonnet 5.5 : Haiku 4.5 = 4 : 2 : 1 per token. Max weighting isn't published; /usage before and after the pilot is the real measure.
So the worker tier is Haiku for mechanical edits, Sonnet for straightforward builds, and Opus only to plan and to review.
~/.claude/agents/: shell-proxy (haiku), worker-mechanical (haiku), worker-builder (sonnet, effort medium), reviewer-senior (opus, effort high). Haiku gets no effort setting: Haiku 4.5 rejects the parameter.
The head chef writes the tickets. Each cook gets their own station, so nobody bumps elbows. A thermometer checks every plate before the chef tastes it.
~/.claude/workflows/director.js
1 Plan (Opus): goal → items {tier, files[], check_cmd, accept_cmd}; no two items share a file
2 Baseline: pin origin/main SHA; checks must be green, acceptance must be red, before work
3 Per item: git worktree ../<repo>-wf-<id> → worker commits → checker runs the tests
4 Review (Opus, blind): spec + diff only; revise at most twice, then hand back
If a cook burns a dish twice, it goes to a more senior cook, and after that back to the chef. Nothing is served on a cook's say-so.
Retry once at the same tier → escalate haiku→sonnet → capped. The checker's exit codes are parsed in plain script; worker self-reports are ignored.
The first shift is five small jobs in ada-stair-generator, and one planted mistake to see if the chef catches it.
P1 Docker suite parity · P2 .dockerignore · P3 compose init · P4 components.rb (sonnet) · P5 README Local Docker is the CI stand-in: GitHub Actions blocked by billing since 2026-09-22.
Every claim in the advice was checked against its source. None were invented, but several were stale or stretched.
True (2026-05-06), but it doubled the five-hour window only. Weekly caps were not raised. anthropic.com/news/higher-limits-spacex
`ollama launch` (2026-01-23). Needs ≥64k context; default is 4k under 24 GB VRAM. docs.ollama.com/integrations/claude-code · docs.ollama.com/context-length
That was January's list. Current docs name newer models, and all of them overflow 8 GB at 64k.
Numbers right: 4 hours, 4 different days, ≥3 average viewers on each, 25 followers. The cut dates from ~June 2025. The May 2026 post opened Bits and subs to non-Affiliates. blog.twitch.tv/en/2026/05/13/monetization-for-all/
2026-08-20, but only after "Creator Sponsorship Certification." Brands still choose.
Allowed, with the rules in T-06. affiliate-program.amazon.com/help/node/topic/GJ64G2W6MJ2N7CXP
If you recommend gear out loud, you say the disclosure out loud too. Written links need it written right next to them.
"As an Amazon Associate I earn from qualifying purchases." Disclosure "in the medium in which the recommendation was provided."
Prices may only be shown if Amazon serves them or via its PA/Creators API. The parts bin stores no prices at all; pages say "Check price".
No giveaways, no "bookmark my link", no shorteners that hide the Amazon destination.
So sign up once /builds has traffic, not before.
SketchUp Make 2017 is licensed non-commercial. The virtual PC builds render in Blender headless, already in ada-stair-generator's compose.yaml.
Four helpers, four jobs. Claude runs the kitchen. Muse runs the front desk: calendar and mail. OpenClaw rings the bell and fetches things from the Windows side. Jev is a quick yes-or-no at the pass.
Muse → Claude: self-email "[MUSE>CC] …"; trusted only if Gmail marks it SENT. Claude → Muse: a draft "[CC>MUSE] …" (Claude has no send tool). Muse reads these drafts in its 7:45 AM briefing (confirmed). Notes are information, never orders. No client details: Muse's connector data can train Meta's AI.
Notify: POST /hooks/wake with its own narrow token. Each ping is one turn on its agent, which now runs on free Gemini Flash (an AI Studio key); it was on OpenAI, so pings cost nothing now. Windows jobs: tray MCP system.run, sandboxed, no network, 30 s limit. Its master key stays out of Claude's reach (a deny rule in Claude's settings).
TypeSafe's decision-only model, 70–500 ms, $0.042 per M input, ~$5/month free credit. Official plugin @openclaw/typesafe. Can be nudged by injected text: never the only gate. Only typesafe.ai; ~670 lookalike domains appeared in its first week. Status 2026-10-01: in the director it is a shadow pre-screen that gates nothing. It agreed with Opus 6/6 across two runs, about $0.0001 per check. It now sees only privacy-gated, redacted text and is skipped for client repos. The director itself is live in 7 repos.
The photo-reading local model waits. It would save almost nothing and needs an hour of your own labelling before it could be trusted. It makes a better episode later.
Design kept: qwen3.5:4b in a "vision" compose profile; adopt only if it beats Haiku on a 24-photo gold set you label yourself.
No second company's coder yet: Codex is on the free plan. An OpenClaw worker would need its master key, so it waits too.
Revisit both if the pilot shows Haiku and Sonnet fall short.
Five tickets went out to the cooks. Four came back right the first time, passed the thermometer and passed the head chef's taste. None had to be handed up to a more senior cook.
ada-stair-generator, landed as 443188a on main: docker-suite-parity (haiku) · dockerignore (haiku) · compose-init (haiku) · components-lib (sonnet) first-pass checks 5/5 · Opus first-round pass 4/5 · escalations 0 Issue #1 closed; epic #3 layer 1 started.
The fifth ticket failed for a reason the kitchen can learn from. Two cooks worked on dishes that depended on each other without being able to see each other's plates. The README described a Dockerfile that another cook was changing at the same time.
readme-sync capped after 2 review rounds; finished by hand after merge. Fix: the planner now merges coupled items or keeps them for Opus.
The head chef also overreached: while writing the tickets, it cooked one dish itself to check the recipe. That was most of the bill.
API-equivalent: $4.74 actual vs $8.03 if every token were Opus (~41% cheaper). Opus was $3.50 of it, mostly the planner building and testing the Sonnet item itself. Fix: planner now runs as the read-only Plan agent.
Then we tested the taster. We slipped in a dish with one wrong number and a menu edited to match it. The chef caught both.
Planted: 5/8" plywood at 9/16" (spec 19/32"), test changed to agree; the suite passed. reviewer-senior (blind): "revise"; flagged components.rb:27 and the test protecting the error.