The Method

The day starts before I do. Three launchd agents run the loop overnight — no server, no cloud, just this machine keeping its own house.

06:17launchd: com.geoandr.nightly-gate-digest — every focus repo's gate.sh + evals.sh06:17jim-agent — ✅ pass (9s) · evals — ✅ pass (2s)06:17grocery-buddy — ✅ pass (3s) · procurement-agent — ✅ pass (2s)06:17dj-agent — ✅ pass (7s)06:45launchd: me.ops-report — the fleet report measures, sanitizes, commits, pushes; the site redeploys07:15sundays: me.curator — every entry drift-checked against its repo; publishes only through its own gates

Drift gets caught within 24 hours instead of at the worst possible moment — and I don't have to remember to check, because the checking isn't mine anymore. The gate lines above are quoted from today's digest file; the first session of the day opens with them already in hand.

No dashboard, no status meeting with myself. A SessionStart hook injects three things before I type a word:

SessionStart · session-context.sh →Git: on 'main', 0 uncommitted change(s). Last commit: 15c85a8 …Gate digest (2026-07-06): 4 repos ✅ · evals ✅## Next steps — the last handoff's ordered list, yesterday-me briefing today-me

The triage rule, in order: a red gate or a yellow eval regression is the first task — you don't build on a broken foundation. Green → start from the handoff's next steps. No handoff → pick a rung below.

Ceremony matches stakes. Every piece of work enters through one of these doors — the expensive mistake is over-planning small things and under-planning big ones.

The deep tier is an economic decision, not a default: the skill frames a self-contained brief and a co-located script runs a bounded fan-out — a few hundred thousand tokens of parallel scrutiny on the decisions that justify it, and nothing on the ones that don't.

The session shape that changes throughput isn't a better prompt — it's conduct, don't type. One session writes briefs and reviews diffs; implementation runs in parallel lanes — background agents and git worktrees — that verify themselves against their own gates.

$revamp /method + the project deep dives3 explorers fan out — the site · jim-agent · grocery + djplan approved — two implementation briefs writtenreviewing diffs, not token streams

brief: goal · constraints · files it owns · verify command · what done meanscompressing the topology · drawing the real sub-graphsverify: tsc ✅ · content check ✅

same brief shape, different territory — lanes can't stomp each otheringest transaction · generation loop · critic thresholdsverify: tsc ✅ · content check ✅

The ramp is deliberately boring: one conductor plus one executor, adding a lane only when the current count feels idle. The ceiling isn't the model — it's brief-writing and diff-review bandwidth. A lane gets touched when it pings or at its checkpoint, never watched; if lanes keep blocking on me, the fix is fewer lanes or sharper briefs.

The highest-leverage question in the whole system is asked before any task starts: can the agent verify this without me? If the answer is no — no test to run, no eval case, no seeded data — that gets built first. It feels slower; it's faster by mid-session, because the agent self-corrects instead of queueing on my eyeballs.

Evals are the same idea for the nondeterministic parts: anything with an LLM decision carries an offline, zero-credential evals.sh — JIM's replays 87 cases in about two seconds for $0.00, armed with a baseline so a regression is a verdict, not a feeling. The full harness is drawn inside JIM's dossier. And --no-verify stays available as the deliberate escape hatch — used rarely, said out loud in the commit message.

Everything above compounds through one rule and one ritual. The rule: anything done or explained three times becomes a skill that session. The ritual: a thirty-minute weekly retro over one question — where did I babysit this week?

Instruction files ride the same loop, treated as code: one canonical AGENTS.md per repo, containing only what an agent gets wrong without being told. Every re-explanation in a session is a bug in that file, fixed with one line, then — not someday.

The proof is this site. It watches the fleet's transcripts, reads the night's digests, files its own report every morning, and syncs its entries against their source repos weekly — gated by its own content check and build rather than by my attention. The archive maintains itself and shows its work in the ops room.

The design rule underneath all of it: the main context window is the most expensive real estate in the system, so heavy reading happens in someone else's window and only conclusions come back. Below is the machine itself — every layer clickable down to the file that enforces it.

The harness is why a one-person shop ships agents on payment rails with red-team suites. Same nervous system as the products, pointed at the producer — and increasingly, running without him.