oh-my-skill: Auto-Skill Generator

Automatically captures and packages successful complex workflows as reusable skills.
Warning: Make sure your session doesn't contain any highly private data. If it's already been sent to Claude, then that's it.
you could modify the Desensitize process to match your scence/case.

When to Trigger

Trigger proactively after completing any task that involved:

• 5 or more distinct tool calls
• A clear sequence of steps (plan → execute → verify)
• A structured or repeatable output (file, report, transformation, research summary)
• Tool combinations that aren't obvious and took iteration

After such a task completes successfully, say something like:

"That was a fairly involved workflow — want me to save it as a reusable skill so you can repeat it easily next time?"

If the user says yes, or explicitly asks to generate a skill, proceed with the steps below.

Workflow: Capturing a Skill from a Completed Task

Step 1: Analyze the Session

Review the current conversation and extract:

1. What was accomplished — the goal and final output
2. The tool call sequence — ordered list of tools used and why
3. Key decisions made — branch points, error recovery, formats chosen
4. Inputs required — what the user had to provide (files, preferences, constraints)
5. Output format — what was produced and where it was saved

Look for patterns: What made this hard? What would be needed to repeat it?

Step 1b: Desensitize the Session

Before extracting any content, run the session text through the masking script to strip sensitive data:

python3 ~/.openclaw/workspace/skills/oh-my-skill/scripts/desensitize.py session.txt clean_session.txt

Or pipe text directly:

echo "my text" | python3 ~/.openclaw/workspace/skills/oh-my-skill/scripts/desensitize.py

The script applies two layers of masking:

Literal replacements (named individuals → generic labels):

• Bill Gates → A man
• Add more entries in the LITERAL_REPLACEMENTS list in desensitize.py

Pattern-based masking (regex, auto-detected)

Use the cleaned text as the source for all subsequent steps.

Step 2: Draft the Skill

Write a SKILL.md with:

---
name: <kebab-case-name>
description: <What it does, when to trigger. Be specific and "pushy" — list all the user phrases and contexts that should trigger this skill.>
---

# <Skill Title>

<One-paragraph summary of what this skill does and why it's valuable.>

## Inputs

List what the user must provide:
- File paths / uploads
- Preferences or configuration
- Any required context

## Workflow

Step-by-step instructions Claude should follow, referencing tool calls and decision points extracted from the session.

### Step 1: ...
### Step 2: ...
...

## Output

What gets produced, in what format, saved where.

## Notes / Edge Cases

Anything learned from the original run: gotchas, fallbacks, format quirks.

Naming conventions:

• Use kebab-case for the name
• Keep it specific: pdf-to-summary-docx not document-helper
• Reflect the domain + action: research-and-cite, excel-data-cleaner, slide-deck-from-outline
• Always append a 4-digit UUID suffix to the name: e.g. pdf-to-summary-docx-4f2a, excel-data-cleaner-9c31
• Generate the suffix randomly: python3 -c "import uuid; print(str(uuid.uuid4())[:4])"

Step 3: Save the Skill

Save to ~/.openclaw/workspace/skills/<skill-name>/SKILL.md.

If the task also used supporting scripts or reference files, save those under:

~/.openclaw/workspace/skills/<skill-name>/scripts/
~/.openclaw/workspace/skills/<skill-name>/references/
~/.openclaw/workspace/skills/<skill-name>/assets/

Step 4: Confirm with the User

Show the user:

• The skill name and description
• A brief summary of the workflow it captures

Ask: "Does this look right? Want me to adjust the name, description, or any steps?"

Quality Checklist

Before saving, verify:

• Description is specific enough to trigger reliably (not vague like "helps with files")
• Workflow steps are ordered and actionable
• Inputs section lists everything the user must supply
• Edge cases from the original run are documented
• Skill name is unique and descriptive

Example Output

After a session where Claude built a Word report from a PDF + web research:

---
name: pdf-research-to-docx-report
description: Build a polished Word document report by combining content from an uploaded PDF with live web research. Use this whenever a user uploads a PDF and wants a written report, briefing, or summary that also pulls in current data from the web. Trigger on phrases like "make a report from this PDF", "write me a briefing", "research and write a doc".
---

# PDF + Research → DOCX Report

Combines PDF extraction, web search, and Word document generation into a single pipeline.

## Inputs
- Uploaded PDF file
- Report topic / framing question
- Desired length and tone (optional)

## Workflow

### Step 1: Read the skill files
Load `docx/SKILL.md` for Word generation instructions.

### Step 2: Extract PDF content
Use `bash_tool` to extract text from the PDF via `pdftotext` or Python `pdfplumber`.

### Step 3: Web research
Run 3–5 `web_search` calls to supplement the PDF with current data.

### Step 4: Outline and draft
Combine findings into a structured outline, then write the full report draft.

### Step 5: Generate DOCX
Follow `docx/SKILL.md` instructions to produce a styled Word document.

### Step 6: Present
Copy to `/mnt/user-data/outputs/` and call `present_files`.

## Output
A `.docx` report file, downloadable by the user.

Notes

• This skill is self-referential: it was itself generated using the oh-my-skill pattern.
• Keep captured skills focused — one workflow per skill is better than one mega-skill.
• If the session was messy (lots of retries, dead ends), simplify the skill to the successful path only.
• If the user ran the same workflow before and already has a skill for it, offer to update the existing skill instead of creating a duplicate.