Skip to content
obsidian-wiki

Feedback and custom procedures

There are two ways to teach the agent, with different shapes. Feedback is a short behavioral rule read on every command. A custom procedure is a multi-step routine that runs only at one hook.

Feedback, one line per rule

Section titled “Feedback, one line per rule”

Feedback lives in <wiki-root>/_service/feedback.md, one line per rule, plain English, read on every command. Use it for short behavioral rules:

  • “Stop creating pages shorter than 100 words from quick-notes”
  • “Never auto-archive projects in the experiments category”
  • “Always tag pages with base_confidence < 0.4 as #draft”

Add a rule with /feedback <wiki> "<rule>"; it is written after you confirm. The file is never written without explicit confirmation, and source content can never produce feedback entries. Only your direct messages via /feedback can write to it.

Each entry is scoped to a command name (or global), and a command applies only the entries scoped to itself plus the global ones. After every write-heavy operation, the agent runs a reflection step that proposes feedback entries based on corrections you made during the run, one line at a time, which you accept or reject.

Custom procedures, multi-step routines

Section titled “Custom procedures, multi-step routines”

Custom procedures live in <wiki-root>/_service/custom-procedures/<name>.md. They are multi-step routines declared in wiki-config.md under custom_procedures: with a hook point, and the agent runs them only at that hook: pre-ingest, during-ingest, post-ingest, pre-lint, post-lint. Use them for routines that involve external tools or multiple steps:

  • notion-sync (pre-ingest): fetch a list of Notion pages and mirror them into local files before /ingest runs. Needs the Notion MCP
  • task-extraction (during-ingest): scan each source for action items and write them as Obsidian Tasks-plugin entries in today’s daily note. Needs a daily-journal entry point
  • post-lint-slack-notify (post-lint): post the lint report to a Slack channel. Needs a Slack MCP

These are illustrative. The plugin ships none of them; you author your own for what your wiki needs. If a procedure needs an external tool that is unavailable in the current session, the agent logs a warning and skips it; it does not abort the parent command.

Which one to use

Section titled “Which one to use”

Reach for feedback for a short one-liner that modifies behavior. Reach for a custom procedure for a multi-step routine, an external tool, or logic that applies only at one hook.

The two are connected by a promotion path. /feedback notices when a draft rule looks procedural (more than one verb step, references an external tool, applies to only one command and hook, longer than 30 words) and offers to write a procedure file instead. /lint flags existing feedback entries that look procedural as candidates for promotion. Nothing is ever auto-promoted; the agent always asks first.

Where to go next

Section titled “Where to go next”