Neal Spangler

Work Automation & AI

An Obsidian runbook with a nightly AI filing job

I keep fixes, incidents and decisions in a Git-tracked Obsidian vault. A nightly AI job connects new notes, while a wrapper reverts any edits it makes to existing notes.

Explore the knowledge graph

The nightly gardener run At 23:59 a timer starts the gardener. It snapshots the vault so my own edits are kept, runs a headless Claude session, then checks every change. New notes are filed into their index and committed. Any edit to an existing note is reverted. Timer nightly at 23:59 Snapshot my edits kept AI run headless Claude Diff check new or edited Commit filed in its index edit to an existing note: reverted Create-only, enforced in code
The wrapper keeps new notes and reverts edits to existing ones.
818runbook notes in one searchable vault
490commits, so every change can be reviewed
186gardener commits over 76 nights
26commits that retract or correct a claim

The problem

Fixes, task cards and daily logs lived in separate places and rarely linked up. An AI could do the linking, but I didn't want a model rewriting notes I depend on during an outage, with no record of what it changed.

What I did

  • A Git-tracked Obsidian vault linking fixes, incidents and decisions, so I can give models the relevant notes and earlier reasoning.
  • A nightly gardener at 23:59: a headless Claude Code run reads the day's notes and may create a few new notes that connect related work.
  • A create-only wrapper: it snapshots the vault first, reverts any AI edit to an existing note, keeps my uncommitted work and commits new files under the gardener's own Git author.
  • Scripts mirror the task board before the AI run and file each new note into its topic index afterward.
  • Sunday checks: a lint report for broken links and orphan notes, and a self-review where the gardener suggests changes that I accept or drop.

What I took from it

One night the gardener tried to edit two index notes and a graph file. The wrapper reverted all three. That run showed the create-only restriction held when the model went beyond its instructions.

Automation & AI

AI agents with command limits and approval checks

Claude Code and Codex handle remote tasks with approval checks and command-blocking rules enforced outside the model.
Privacy