Highest-leverage upgrades for a solo operator: add automatic prompt caching to the Railway harness (one line, cuts every per-client API call), wire the Windows safety and brand hooks, install Context7 for stale docs, write a DESIGN.md so the site stops looking generic, and cap parallel sub-agents at 5-6.
The 5 highest-leverage moves (start here)
01
Automatic prompt caching in the agency harness
One field on every messages.create call cuts recurring cost up to 90% per turn, extending the buyer's $40/mo cap.
E1 Agent SDK · Effort S
02
Cap parallel sub-agents at 5-6 + /compact first
A 10-agent batch burns 7-10x tokens and can close your 5-hour window in under 2 hours. Free behavior change.
E2 Operations · Effort S
03
Context7 MCP for current API docs
Solves the stale-SDK-docs pain built-in WebFetch cannot, at near-zero context cost (fires only on library mentions).
A2 MCP · Effort S
04
Windows safety + brand-rule hooks
Destructive-command warn closes the bypassPermissions gap; em-dash scrubber automates the brand rule. Under an hour.
C2 Skills & Hooks · Effort S
05
DESIGN.md + 7-exemplar reference library
Brand tokens are already locked. This is what makes the site stop looking like generic AI slop, and the site feeds the content engine.