Obsidian Edit

For each target file:

1. Read the full file
2. Identify all {...} annotations — skip those inside code blocks (between ``` markers) and YAML frontmatter (between --- markers at file start)
3. For each annotation, working top to bottom:
   • The text inside {...} is the edit instruction
   • The surrounding paragraph is the context to edit
   • If the annotation references other content (e.g., {link to X}, {cite source}, {connect to related concept}), use qmd to find the relevant note:

     qmd search "concept name" -c obsidian
     qmd vsearch "what the annotation is asking about" -c obsidian
     

     Use the found note's title to create a proper [[wiki link]] woven into the text.
   • Apply the instruction: improve the surrounding text accordingly
   • Remove the {...} annotation
4. Write the edited file using the Edit tool

3. Sync

After all files processed:

cd ~/obsidian-vault && git add -A && git commit -m "vault: process spatial edits" && git push
qmd update

4. Report

Output a summary per file:

Processed spatial edits:

Knowledge/Agentic Memory/example.md (3 annotations):
1. "{feels abstract}" -> added concrete mechanism
2. "{expand this}" -> added supporting examples
3. "{too technical}" -> simplified language

Total: 1 file, 3 annotations processed

Skip rules

• {text} inside fenced code blocks -> skip (likely an example)
• {text} inside YAML frontmatter -> skip
• Empty {} -> skip
• {unclosed or unopened} -> treat as literal text, skip

Examples

Example 1: Clarity

Input:

vaults give claude memory
{feels abstract -- what kind of memory?}

Output:

vaults give claude persistent memory across sessions by storing context in files it can read and write

Example 2: Strengthening

Input:

without persistent storage claude forgets everything between sessions
{this is the key point, make it hit harder}

Output:

without persistent storage claude starts fresh every conversation -- you re-explain the same context, rebuild the same understanding, lose the compound effect of accumulated knowledge

Example 3: Showing instead of telling

Input:

the solution is simple
{dont say simple, show}

Output:

the solution: store everything in markdown files that claude can traverse

Example 4: Link resolution (uses qmd)

Input:

the vault uses progressive disclosure to filter what agents read
{link to the note about this}

Process:

qmd search "progressive disclosure" -c obsidian
# -> finds: Knowledge/Agentic Memory/ContextEngineering.md

Output:

the vault uses [[ContextEngineering|progressive disclosure]] to filter what agents read

Quality checks

• All annotations in target files processed (run find-annotations.sh to confirm none remain)
• Code blocks and YAML frontmatter unchanged
• Changes committed and pushed
• qmd index updated
• Report shows file, count, and description for each edit