~/projects/executive-assistant/README.md
Two halves of one personal operating system. The Executive Assistant is a Telegram bot backed by n8n, Claude and a pgvector memory: a weekday morning brief, scheduled nudges, voice notes turned into tasks, and questions answered from my own history. MangJoseOS is the long-term memory, an Obsidian vault run by Claude Code with an evidence layer, a synthesized wiki, and skills that turn repeated work into one-command procedures.
cat architecture.flow
architecture.flow
Morning brief
Capture
Second brain
cat problem.md
Every morning started with the same manual triage: check the commute and flood conditions, scan two mailboxes, review open tasks, check what my staff were waiting on, and skim the news. Deadlines buried in email were easy to miss.
Ideas and to-dos captured by voice or chat had nowhere reliable to land, and nothing remembered what had happened last week, so the same questions got re-answered from scratch.
AI coding sessions have the same problem at a larger scale. Each session starts with no memory of past decisions, incidents or procedures, so the same lessons get relearned.
cat solution.md
The Executive Assistant is a set of n8n workflows that fan out to traffic, weather, news, mail, calendar and my task system on a schedule. Claude writes the brief as structured JSON that becomes a Telegram message with one-tap "add as task" buttons. Afternoon nudges and a Monday finance report follow the same pattern.
Voice notes are transcribed by a local Whisper service with a mixed English and Tagalog vocabulary. Claude classifies each one as tasks or a note. Tasks go to TALA, and notes go to a diary and a pgvector memory that powers retrieval-based Q&A over my own history.
MangJoseOS is the long-term layer. It is an Obsidian vault with three tiers: immutable raw sources, a synthesized and cross-linked wiki, and PARA folders for live work. Claude Code is instructed to check the vault before answering from general knowledge, and to write back what it learns.
Repeated work becomes Claude Code skills: capturing knowledge into the vault, publishing to the blog, adding dated reminders to TALA, building n8n workflows, and ingesting YouTube tutorials. A harvester skill watches for repeated tasks and proposes new skills, which are only created after I approve them.
ls ./components
Schedules, the Telegram trigger, mail, calendar and HTTP calls all live in n8n. Each data source retries on its own and degrades to "unavailable" instead of failing the whole brief.
Claude runs as the CLI on the homelab under my subscription. n8n reaches it over SSH with a forced-command key, so no API key sits on the host and the key can't run arbitrary commands.
Brief, nudge, ageing and date logic live in plain JS modules with unit tests. A build script inlines them into n8n Code nodes, and a test checks that the build matches the source.
Diary entries and notes are embedded locally with Ollama (nomic-embed-text) and stored in PostgreSQL with pgvector, so retrieval Q&A never sends my history to a third-party embedding API.
The weekly finance report reads the Liquidation Tracker through a dedicated read-only role with explicit column grants. The assistant can report on the business but can't change it.
The vault's CLAUDE.md defines the operations (ingest, query, lint, new client project), page types and logging rules, so every session works the vault the same way.
./run --trace
At 07:30 the workflow fetches live and predicted commute time, weather, news, both mailboxes, staff status from TALA, and the diary, with per-source retries.
Sections that failed are listed as unavailable, and the whole prompt is trimmed to a byte budget.
Claude returns JSON: the brief as escaped Telegram HTML plus up to five suggested tasks.
The brief is sent to Telegram with one-tap buttons that create the suggested tasks in TALA.
./run --trace 2
Videos are treated as a source of claims, never a source of truth.
A skill calls an n8n workflow that runs a locked-down container (yt-dlp, with faster-whisper as a fallback when captions are missing) through a forced-command SSH key.
Concepts, procedures and factual claims are pulled from the timestamped transcript.
Each claim is checked against official documentation and recorded in a claim ledger as verified, contradicted or unverified.
Source, concept and entity pages are written with cross-links, and the index and log are updated.
ls ~/.claude/skills
Each skill encodes a procedure we did more than once, with its guardrails, so it runs the same way every time.
Records what a session learned or built into the right page type, with cross-links and an index and log entry.
Notices repeated multi-step work and proposes turning it into a skill. Nothing is created without approval.
Transcript → claims → independent verification → provenance-tracked wiki pages.
Writes a session up in my voice, generates a feature image and publishes to Ghost. It drafts by default.
Turns any date that comes up in a session into a dated TALA task through its machine API.
Builds, validates, deploys and tests n8n workflows on the homelab instance, with error alerting wired in.
tail -f incidents.log
cat package.json | jq .stack
echo $EXIT_STATUS
outcome.log
My weekday starts with one Telegram message instead of six apps, and anything I say into my phone ends up as a task or a searchable memory within seconds.
The second brain changed how I work with AI. Decisions, incidents and procedures persist across sessions, and repeated work becomes a reviewed, reusable skill instead of a re-explained prompt.