Installs into .claude/skills of the current project.
Are you the author of 2501 Auto Summary 5f793154?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/tools-only-2501-auto-summary-5f793154)
---
description: Batch-generate one-line summaries for notes missing the summary frontmatter field using parallel Haiku agents
model: haiku
---
# /auto-summary
Batch-generate one-line `summary` fields for notes missing this frontmatter property. Uses parallel Haiku agents processing notes in batches of 15-20 for fast, cost-effective summary generation. The `summary` field enables AI triage, search results ranking, and quick note previews.
## When to Use This Skill
- Adding summaries to notes that lack them
- Improving search and discovery across the vault
- Preparing notes for AI-assisted triage and retrieval
- Post-migration cleanup to add summary fields
- Generating previews for index pages or dashboards
## Usage
```
/auto-summary [--scope path/to/folder] [--mode suggest|apply] [--max 100]
```
### Parameters
| Parameter | Description | Required |
|------------|------------------------------------------------------|----------|
| `--scope` | Folder or glob pattern (default: notes missing summary) | No |
| `--mode` | `suggest` shows summaries; `apply` writes them (default: `suggest`) | No |
| `--max` | Maximum number of notes to process (default: 100) | No |
## Instructions
### Phase 1: Identify Notes Needing Summaries
1. **Scan notes in scope** — Find all Markdown files
2. **Parse frontmatter** — Check for existing `summary` field
3. **Filter:** Keep only notes where `summary` is missing or empty
4. **Sort by priority:** ADRs and Concepts first (most benefit from summaries), then Projects, then Meetings
5. **Divide into batches** of 15-20 notes per agent
### Phase 2: Batch Summary Generation — Agent Team (Parallel Haiku Agents)
Use the Batch Processing pattern. Launch N parallel agents.
**Agent 1-N: Summary Writer** (Haiku)
Task: Generate one-line summaries for assigned batch
For each note:
1. Read the title, frontmatter, and first 500 words of body content
2. Generate a one-line summary following the **Writing Rules** and **Summary Patterns by Type** below
3. Return: List of `{ filename, generatedSummary }` per note
#### Writing Rules
Every generated summary must follow these constraints:
- **Length:** 10-25 words, one sentence, no period at the end
- **Voice:** Start with the subject, not "This note..." or "A document about..."
- **Tense:** Use present tense for active items, past tense for completed
- **Content:** Include the most distinctive and searchable terms
- **Avoid generic openers:** Never start with "Overview of...", "Notes on...", "Summary of...", "Document about...", "Page covering..."
- **UK English** throughout
- **No line breaks** — single sentence only
#### Summary Patterns by Type
| Type | Pattern | Example |
|------|---------|---------|
| Concept | What X is and why it matters | `Continuous integration practice ensuring code changes are validated automatically on every commit` |
| Pattern | How to apply X for Y | `Event-driven architecture pattern for decoupling services via asynchronous message passing` |
| ADR | What was decided and why | `Selected PostgreSQL over MongoDB for transactional consistency in the order management domain` |
| Meeting | What was discussed and decided | `Sprint review covering API migration progress and blockers for the Q2 release` |
| Project | What the project delivers and its current state | `Data platform modernisation delivering real-time analytics for operational dashboards` |
| Task | What needs to be done and why | `Review security findings from penetration test before production release approval` |
| Reference | What the source covers and its relevance | `AWS Well-Architected Framework covering operational excellence and reliability pillars` |
| Research | What was investigated and key findings | `Comparison of stream processing frameworks for event-driven integration workloads` |
| System | What the system does | `Enterprise MRO platform managing aircraft maintenance scheduling and compliance tracking` |
| Person | Role and context | `Solutions Architect in platform engineering, leading the data integration workstream` |
| Incubator | What idea is being explored | `Exploring voice-activated workflows for hands-free note capture during site visits` |
| Email | What the email communicates | `Proposal to programme board for adopting infrastructure-as-code across the platform team` |
#### Quality Checks
After generating each summary, validate against these rules:
- **Too long:** Flag if over 25 words — tighten the phrasing
- **Too short:** Flag if under 10 words — add meaningful context
- **Too vague:** Flag if the summary could apply to any note of that type (e.g., "Architecture decision about a technology choice")
- **Banned phrases:** Reject summaries starting with any of these:
- "This note..."
- "A document about..."
- "Summary of..."
- "Overview of..."
- "Notes on..."
- "Page about..."
- "Document covering..."
- **Missing specifics:** Flag if the summary lacks any concrete nouns (technologies, processes, outcomes)
If a summary fails quality checks, regenerate it before returning.
### Phase 3: Review and Apply
1. **Compile summaries** from all agents
2. **Present for review** in a table
3. If `--mode apply`: Write summaries to frontmatter
**Frontmatter format:**
```yaml
summary: "Selected PostgreSQL over MongoDB for transactional consistency in the order domain"
```
Always quote the summary value since it may contain special YAML characters (colons, brackets).
## Output Format
```markdown
# Auto-Summary Report
**Date:** YYYY-MM-DD | **Notes Processed:** X | **Summaries Generated:** X
## Summaries
| Note | Generated Summary | Apply? |
|-----------------------------------|-----------------------------------------------------|--------|
| `ADR - Event Sourcing.md` | Decision to adopt event sourcing for order audit trail | Yes |
| `Meeting - 2026-01-15 Sprint.md` | Sprint planning covering order service backlog priorities | Yes |
| `Concept - CQRS.md` | Command Query Responsibility Segregation pattern for read/write separation | Yes |
## Statistics
| Note Type | Processed | With Summary | Without Summary |
|------------|-----------|--------------|-----------------|
| ADR | X | X | X |
| Meeting | X | X | X |
| Concept | X | X | X |
| Project | X | X | X |
| Other | X | X | X |
```
## Safety
- Always use `--mode suggest` first for vault-wide operations
- Never overwrite existing non-null summaries
- Commit to git before running
- If note body is too short (<50 words), skip rather than guess
- Summary generation is additive only — never removes existing summaries
## Examples
### Example 1: Preview All Missing Summaries
```
/auto-summary
```
### Example 2: Apply to ADRs
```
/auto-summary --scope ADRs/ --mode apply
```
### Example 3: Limit to 50 Notes
```
/auto-summary --max 50 --mode suggest
```
---
**Invoke with:** `/auto-summary` to batch-generate summaries for notes missing them