A scheduled, intelligence-assisted pipeline that finds scholarships, matches them to two student profiles, reports ranked matches for human review, then assists with submission — human-gated throughout.
github.com/rrittich/scholarship-finder · decided 2026-06-23instructions/*.md (the hydrated brain) ·
data/scholarships · data/matches ·
data/state.json · data/reportssend_email — email submissions + PDFsThe routine sandbox blocks arbitrary web fetch by default (allowlist only). Set the environment's
network access to Full or capability 1 can't reach scholarship pages.
Minors' particulars stored as encrypted *.enc.json; decrypted only in routine memory.
Reports reference a son by id / first name only.
Every run commits something (even “0 new matches”) — a missing commit is itself the failure alarm. Ephemeral cron jobs fail silently.
Native GitHub App auth (or /web-setup) — the routine inherits your identity.
Don't stash a PAT in env vars; those are stored plaintext-readable.
Routines are free under Max (“act as you”). The $0.08/hr + token metering is the separate Managed Agents product (“acts as a service”).
The June 15 2026 split (Agent SDK / claude -p off-subscription → metered credit)
was paused. Programmatic usage still draws from Max “for now,” advance notice promised.
Never auto-submits; never ghost-writes an essay submitted as a son's own work. Many scholarships require authentic student work — the system assists, you decide.
Routine cron minimum granularity is 1 hour (daily is plenty). Per-run ceiling ≈ 4 vCPU / 16 GB / 30 GB.