Create Skill Workflow

Purpose

Guide users through creating new AgentOps skills with consistent structure and quality. Reduces friction for skill ecosystem growth while enforcing standards.

Mode Selection

Present mode options at the start:

Clone Mode Procedure

1. List available skills from

   .github/skills/

2. User selects one to clone
3. Read that skill's SKILL.md as base
4. Interview focuses on what to change/customize

Complexity Tiers

After mode selection, assess complexity:

Complexity Assessment Questions

Ask the user:

1. "Does this skill have a single main procedure, or multiple branching paths?"
2. "Does it need to invoke other skills or handle errors specially?"

If answers suggest simple → Simple tier (5 questions) If answers suggest complex → Complex tier (10+ questions)

Interview Questions

Simple Tier (5 Questions)

Use

agent-ops-interview

my-custom-skill

name

.github/skills/

agent-ops-

description

state_files

create_prompt

the {name} prompt section

IMPORTANT: If user provides a name starting with

agent-ops-

❌ Cannot create skill with

agent-ops-

prefix. This prefix is reserved for bundled assets managed by

aoc bundle install

. Please choose a different name (e.g.,

my-tags

instead of

agent-ops-tags

).

Category Assignment (After Q2)

After collecting the skill name and description, assess and assign category:

Step 1: Auto-detect category based on description keywords:

git

analysis

docs

core

core

utility

extended

Step 2: Assess certainty

Step 3: Display category assignment

Always show the category before proceeding:

📂 **Category Assignment**

Based on your description, this skill fits the **{category}** category.

Available categories:
- **core**: Essential workflow skills
- **git**: Git operations and branch management  
- **analysis**: Code analysis and review
- **docs**: Documentation management
- **utility**: Supporting utilities and helpers
- **extended**: Specialized/domain-specific skills

Is **{category}** correct? (yes / or type a different category)

If certainty is LOW, present as a question instead:

📂 **Category Selection**

I couldn't confidently determine the category from your description.

Available categories:
- **core**: Essential workflow skills (constitution, baseline, planning...)
- **git**: Git operations (git, branch-workflow, selective-copy...)
- **analysis**: Code analysis (context-map, code-review, project-sections...)
- **docs**: Documentation (docs, mkdocs, versioning...)
- **utility**: Supporting utilities (interview, guide, tools, debugging...)
- **extended**: Specialized skills (api-review, docker-review, research...)

Which category fits best?

Step 4: Record final category for use in frontmatter generation.

Complex Tier Additional Questions (6-10+)

invokes

invoked_by

Follow-up questions (as needed based on answers):

• "What are the preconditions before this skill can run?"
• "What completion criteria determine success?"
• "What are common anti-patterns to avoid?"
• "Can you provide an example invocation?"

Skill Template

After interview completion, generate SKILL.md using this template.

CRITICAL: Do NOT wrap the output in code fences. The SKILL.md file must be plain markdown starting directly with the

---

```skill

```markdown

Correct output (raw markdown, no fence):

---
name: my-skill
description: "..."
...

Wrong output (wrapped in fence — NEVER do this):

```skill
---
name: my-skill
...
```  ← This fence breaks skill loading

Template Structure

---
name: {name}
description: "{description}"
category: extended
invokes: [{invokes}]
invoked_by: [{invoked_by}]
state_files:
  read: [{read_files}]
  write: [{write_files}]
---

# {Title} Workflow

## Purpose

{purpose_description}

## Procedure

{main_procedure}

## Completion Criteria

- [ ] {criterion_1}
- [ ] {criterion_2}

## Anti-patterns (avoid)

- ❌ {antipattern_1}

Template Field Mapping

{name}

{description}

{purpose_description}

{category}

{main_procedure}

{read_files}

{write_files}

{invokes}

{invoked_by}

Generation Procedure

1. Validate answers:

   • Name does NOT start with

     agent-ops-

     (reserved prefix)
   • Name is unique (check

     .github/skills/

     doesn't have directory)
   • Category is valid enum
   • Description is non-empty

2. Create directory:

   • Path:

     .github/skills/{name}/

3. Generate SKILL.md:

   • Populate template with interview answers
   • Add sections based on complexity tier

4. Generate prompt file (if requested):

   • Path:

     the {name} prompt section

   • Use prompt template (see below)

5. Confirm with user:

   • Show generated content
   • Ask "Does this look correct? (yes/edit/cancel)"

6. Save file(s):

   • Write to

     the {name} skill section

   • Write to

     the {name} prompt section

     (if requested)

Prompt Template

When user requests a prompt file, generate using this template:

Use the `{name}` skill for {short_description}.

## Quick Usage

{brief_usage_example}

## When to Use

- {use_case_1}
- {use_case_2}

## Options

{any_modes_or_options_from_skill}

Mapping:

• {name}

  → skill name from interview

• {short_description}

  → description from interview (lowercase, no period)

• {brief_usage_example}

  → derived from main procedure

• {use_case_1/2}

  → inferred from purpose

Registration Procedure

After skill file is created, auto-register in SKILL-TIERS.md:

1. Read

   .github/SKILL-TIERS.md

2. Determine tier from category: extended Category	Tier
   core	Tier 1
   utility	Tier 3
   analysis	Tier 5
   git	Tier 4
   recovery	Tier 4

3. Find table for target tier (pattern:

   ### Tier {N}:

   )

4. Insert row at end of table:

   | `{name}` | {description} | {invoked_by or "User request"} |
   

5. Save file

6. Confirm: "✅ Skill registered in SKILL-TIERS.md under Tier {N}"

Post-Creation Validation

After generation, verify:

• File exists at

  the {name} skill section

• Frontmatter is valid YAML
• Name in frontmatter matches directory name
• Name does NOT start with

  agent-ops-

  (reserved prefix)
• Description is non-empty
• At least one procedure section exists
• State files declared (read/write)
• Registered in SKILL-TIERS.md
• Prompt file exists at

  the {name} prompt section

  (if requested)

If validation fails: Report specific failure, offer to fix or regenerate.

Examples

Example 1: Simple Skill Creation

User: "Create a skill for managing project tags"

Interview:

Q1: "What should this skill be called?"
A1: "my-tags"

Q2: "Describe in one sentence what this skill does:"
A2: "Manage project-level tags for categorizing issues and documentation"

Q3: "What category does this fit?"
A3: "utility"

Q4: "What is the main procedure?"
A4: "1. List current tags from .agent/tags.md
     2. User selects action (add/remove/rename)
     3. Execute action and update file"

Q5: "What state files does it read and write?"
A5: "read: tags.md, focus.md; write: tags.md, focus.md"

Q6: "Should this skill have an accompanying prompt file for slash command usage?"
A6: "yes"

Generated output:

• the my-tags skill section

• the my-tags prompt section

Example 2: Clone Mode

User: "Create a skill similar to agent-ops-docs but for API documentation"

Interview:

Mode: Clone agent-ops-docs

Q1: "What should this skill be called?"
A1: "my-api-docs"

Q2: "What's different from agent-ops-docs?"
A2: "Focuses on API endpoint documentation, generates OpenAPI specs"

... (remaining questions focus on differences)

Generated output: Based on agent-ops-docs structure with API-specific modifications

Cancellation Handling

If user says "stop", "cancel", or "nevermind" at any point:

1. Acknowledge cancellation
2. Report progress: "Skill creation cancelled at question {N} of {total}"
3. No files created unless user explicitly confirmed

Anti-patterns (avoid)

• ❌ Creating skill with reserved

  agent-ops-

  prefix
• ❌ Creating skill without interview (skipping validation)
• ❌ Registering skill in wrong tier based on category
• ❌ Creating duplicate skill name
• ❌ Generating SKILL.md without user confirmation
• ❌ Skipping validation checklist after creation