LIVING OFF AI · THE ENGINE

Context & the Feedback Loop

Context is what the AI loads before it answers. The feedback loop writes what you actually did back into the files.

ContextassemblyFeedbackloop🛡️Constitution🧠Memory🪄Skills🧰Tools🎯Focus

📎 Same 5-layer map as Module 0. Framing follows Anthropic, "Effective context engineering for AI agents," Sep 29, 2025. · source

MOVEMENT 1 · CONTEXT ASSEMBLY

How Files Hold Context: 3 Layers

L1, the index (the CONTEXT file): a short map. Always read first.

L2, the detail: full content, opened on demand.

L3, the archive: finished work, out of the way. Not listed in the index means invisible to the AI.

📎 Load-on-demand mechanics: Anthropic, "Effective context engineering for AI agents," 2025. Layer pattern: one builder's live vault, running since 2025. · source

L1 · indexL2 · detailL3 · archivelisted at L1×Nrepeats atevery nodecontent settles down · links stay on top
MOVEMENT 1 · REAL EXAMPLE

One Index, 3,489 Files

My network folder: 5 subfolders. People, companies, communities, events, call transcripts.

The AI reads one file: _network-CONTEXT.md, 2.4k tokens. It says what lives where. Everything else opens on demand.

📎 Counts: one builder’s live vault, Jul 2026 · index file 9.6 KB ≈ 2.4k tokens.

📁 network/ · 3,489 files_network-CONTEXT.md · the index · 2.4k tokenspeople/ · 525 filestranscripts/ · 2,785events/ · 109companies/ · 28communities/ · 24masters + logs · 18One question about one person = the index (2.4k tok)+ one person file (≈500 tok). The other 3,487 stay closed.
THE PRICE OF ALWAYS-ON

What Always-On Costs

6 files load before every single answer: 47 KB, about 12,000 tokens.

That is 1.2 percent of a 1M-token window. The window is not the constraint, the recurring bill is: every always-on line is billed every session.

📎 1 token ≈ 4 chars: OpenAI API docs, "Concepts." ≈ 3.5 chars for Claude: Anthropic glossary. Always-on files load and consume tokens every session: code.claude.com/docs/en/memory. Cache pricing: platform.claude.com, prompt caching. · source

always-on: 47 KB≈ 12k tokens1M-token context window1 KB250-340 tokper sessionbilled againcache read0.1× pricesmall slice of the window · real money every session
MY ALWAYS-ON STACK

My Constitution, Measured

Models load and bill tokens, not KB. 1 KB is roughly 250 to 300 tokens.

My bands: under 2k it barely knows you, 5 to 20k is healthy, past 50k cost rises and answers degrade.

The size rule is written inside the constitution itself.

📎 1 token ≈ 4 chars: OpenAI API docs, "Concepts." ≈ 3.5 chars for Claude: Anthropic glossary. File sizes: my live vault, measured Jul 2026. The bands and the in-file budget are my practice, not vendor guidance. · source

my 6 always-on files · tokens per sessionsystem map3.0kskills table2.8kcurrent focus2.5knorth star1.8kvoice1.4kwho I am0.3ktotal ≈ 12k tokens · 47 KB · 1.2% of a 1M window
MOVEMENT 2

The Feedback Loop

Every week the system reads what actually happened, calendar, email, notes, compares it with the plan, and writes updates back into the files.

📥 Sourceswhat happened
🧭 Compareagainst the plan
🧠 UpdateMemory + Focus files

📎 Schluntz & Zhang, "Building effective agents," Anthropic, Dec 19, 2024. · source

THE LOOP · ON A CLOCK

Scheduled Jobs Feed the Loop

Cron is the same skill run on a clock. Mine: calendar sync, call transcript ingest, health data, a 05:00 morning brief.

All access is read-only. This is the feedback loop running without me.

📎 calendar.readonly and gmail.readonly are standard read-only OAuth scopes: developers.google.com/workspace/calendar/api/auth · developers.google.com/workspace/gmail/api/auth/scopes. Job list: my live crontab. · source

📅 calendar✉️ email📞 calls⌚ trackerthe vaultfiles first05:00 briefto chat~30 scheduled jobs · files, then chat
THE LOOP · REAL EXAMPLE

One Week Through the Loop

Monday, a client call: the transcript job writes decisions and promises into that person’s file.

All week, email and calendar scans collect what changed. Sunday it converges into the focus draft. I reply OK.

📎 Pipeline: one builder’s live week, Jul 2026 · scans are read-only, writes wait for approval.

MON🎙 client call → transcript analyzedwrites: that person’s file · decisions + promisesTUE–SAT✉️ + 📅 daily scans · read-onlycollects: what changed, meetings held, mail threadsSUN📬 focus draft email · built from all of itI reply: OK 1 3 5 · EDIT 4files updated · next week starts from truth
THE LOOP · EMAIL CHANNEL

Give Your AI Its Own Gmail

Create a new, separate Gmail account that belongs to the AI. Forward the emails that matter.

It reads them and sends back a weekly digest, 4 emails per week max.

Your main mailbox is never handed over.

📎 Gmail read via official OAuth (gmail.readonly): developers.google.com/workspace/gmail/api/auth/scopes. Caps and categories: my live setup. · source

forward what matters✉️ your mailboxAI's own Gmailweekly digest backmax ~4 emails / week · third parties: neverdeep and slow by email · fast and noisy in chat
EXAMPLE · EMAIL AUTOMATION

Example: The Weekly Focus Email

Every Sunday that Gmail account sends me a draft: my focus file refreshed across 7 life areas, only the changes shown.

I reply from my phone: OK 1 3 5, EDIT 4. It writes only after approval.

📎 Layout and approval grammar: one builder’s live Sunday email, running weekly since 2025.

Weekly focus · draft, week 29from: my assistant <second.brain@gmail.com> · to: meYour focus draft. 3 of 7 areas changed this week:1 Bodytraining moved outdoors · weight trend on track3 Peopletwo calls logged · one follow-up owed by Friday5 Businessdemo day confirmed Jul 28 · 2 videos published4 areas unchanged. Full text of each change below.Reply to approve: OK per area number, EDIT to changeOK 1 3 · EDIT 5: ...
THE LOOP · NOT THEORY

My Skill Table, Live

Every skill is a file with a trigger word. Say the word, the AI opens the file and follows it.

Mine today: 19 triggers. Seven are packaged for you in the library.

📎 Trigger table: my live system, Jul 2026 · the packaged seven: materials/library

skills · trigger → jobMIT ★plan the day, 3 prioritiesFOCUS7 ★weekly focus refresh, diff + OKMAIL7 ★the email channel, weekly digestIN7free-form check-in → weekly fileINBOXsort the dump folder, wait for OKFULL7ingest a call recording, enrichVOICE7local voice-to-text, 100% privateEXIT7close the session, write backCOX7 ★index check, one master per node+ 10 more · ★ = packaged in the library
WHEN ONE WINDOW IS NOT ENOUGH

Multi-Agent Orchestration

One lead agent splits the job. Each subagent gets its own context window, works in parallel, and reports back. Results are checked before merging.

In Claude Code one word triggers it: ultracode, now an official effort setting. OpenAI and Gemini offer SDK equivalents.

📎 Dynamic Workflows + ultracode: claude.com/blog, May 28, 2026 · subagents and agent teams: code.claude.com/docs · Codex subagents (TOML, max_threads 6, Ultra delegation): learn.chatgpt.com/docs/agent-configuration/subagents · Gemini CLI to Antigravity CLI: developers.googleblog.com, May 19, 2026. · source

one lead agent🔎researchthe web🔨buildit📚check my ownfiles first⚖️verifythe resultchecked, then mergedown window each · works in parallel · reports back
CLOSING

Next: Maintenance and Growth

This deck was inflow and learning. Next deck: the fifteen-minute Sunday routine, two audit prompts, and the growth path into Modules 4 and 5. Module 4 installs all of this as skills.

loai.wisdomblend.com · livingoffai.nexus/accountability

brief

📎 Anthropic, "Effective context engineering for AI agents," 2025 · Schluntz & Zhang, "Building effective agents," 2024.