Live
BREAKING: OpenAI GPT-5.5 Instant is now the default ChatGPT model — 52.5% fewer hallucinations on high-stakes topics; new memory sources feature shows users exactly what context shaped their responsesBREAKING: Grok falls to 5th place in AI chatbot rankings — Claude surges 1,205% year-over-year as xAI loses 80+ staff and faces App Store removal threat over explicit image controversyBREAKING: Five major publishers sue Meta over Llama — Hachette, Macmillan allege millions of books pirated for AI training; model can reproduce verbatim passagesBREAKING: Trump administration strikes AI safety deals with Google DeepMind, Microsoft, and xAI — pre-release model review extended to three more major labsBREAKING: Anthropic deploys 10 AI agent templates for Wall Street — Claude Opus 4.7 leads Finance Agent benchmark at 64.37%, now integrates with Excel, PowerPoint, Word, and OutlookBREAKING: Trump White House weighs executive order to vet AI models before release — Anthropic's Mythos model reportedly triggered the policy reversalBREAKING: IBM Think 2026 — watsonx Orchestrate repositioned as agentic control plane, IBM Bob reaches GA, Concert Secure Coder embeds security in developer workflowBREAKING: WEF report — 94% of cyber leaders say AI is defining force in cybersecurity; strategic AI adopters cut breach costs by $1.9M and shorten lifecycle by 80 daysBREAKING: 2026 cyberattack analysis — 17-year-old breached 7M records to buy Pokemon cards; time-to-exploit collapsed from 700 days in 2020 to just 44 days in 2025BREAKING: MIT Technology Review — AI is becoming the primary interface for democratic participation, but institutions were not designed for this worldBREAKING: Pentagon signs AI military deals with OpenAI, Google, Nvidia, Microsoft, Amazon, SpaceX, and Reflection AI — Anthropic refuses, sues over autonomous weapons concernsBREAKING: May Day 2026 — labor movement faces existential threat as Amazon plans to replace 500,000+ jobs with robots and AI automationBREAKING: AI research undergoes great pivot — focus shifts from model-centric breakthroughs to system-level deployment and autonomous scientific discoveryBREAKING: Northwestern study reveals AlphaFold2 expanded structural biology rather than replacing it — human-AI collaboration model offers template for futureBREAKING: PNNL scientists use machine learning to optimize nuclear waste vitrification at Hanford — could save hundreds of millions and reduce project timeline by yearsBREAKING: Meta raises AI capex to $125-145B, Google to $180-190B — trillion-dollar question: is the spending actually working?BREAKING: OpenAI launches GPT-5.5 agentic AI — AWS and Databricks announce managed agent services powered by GPT-5.5BREAKING: IBM launches Bob, end-to-end SDLC AI platform — $20/month Pro tier, multi-model orchestration, enterprise governanceBREAKING: Roblox Indonesia implements mandatory facial scanning for users under 16 — privacy advocates raise concernsBREAKING: DW investigation: AI industry burning trillions with no clear path to profitability — financial analysts warn of bubbleBREAKING: Microsoft drops exclusive OpenAI license — OpenAI now free to work with Amazon, Google, and any cloud providerBREAKING: China blocks Meta's $2B Manus acquisition — Beijing orders deal unwound in major cross-border AI tech rulingBREAKING: Musk vs. Altman trial begins — nine-person jury seated in Oakland, $134B OpenAI lawsuit opens TuesdayBREAKING: Big Tech's $600B AI earnings reckoning — Alphabet, Microsoft, Meta, Amazon all report WednesdayBREAKING: OpenAI proposes 4-day workweek, robot tax, and public AI wealth fund in sweeping economic policy blueprintBREAKING: Congress racing to reform FISA Section 702 before AI supercharges warrantless surveillance of AmericansBREAKING: Google reveals 75% of all new code is now AI-generated — up from just 25% eighteen months agoBREAKING: Google Gemini April Drop: native Mac app, AI music creation with Lyria 3 Pro, and Notebooks go liveBREAKING: Forbes AI 50 released — OpenAI leads at $182.6B funding, but vertical specialists are the real storyBREAKING: BCA Research warns AI trade entering 1999-style melt-up — S&P 500 could hit 9,200 before correctionBREAKING: OpenAI GPT-5.5 Instant is now the default ChatGPT model — 52.5% fewer hallucinations on high-stakes topics; new memory sources feature shows users exactly what context shaped their responsesBREAKING: Grok falls to 5th place in AI chatbot rankings — Claude surges 1,205% year-over-year as xAI loses 80+ staff and faces App Store removal threat over explicit image controversyBREAKING: Five major publishers sue Meta over Llama — Hachette, Macmillan allege millions of books pirated for AI training; model can reproduce verbatim passagesBREAKING: Trump administration strikes AI safety deals with Google DeepMind, Microsoft, and xAI — pre-release model review extended to three more major labsBREAKING: Anthropic deploys 10 AI agent templates for Wall Street — Claude Opus 4.7 leads Finance Agent benchmark at 64.37%, now integrates with Excel, PowerPoint, Word, and OutlookBREAKING: Trump White House weighs executive order to vet AI models before release — Anthropic's Mythos model reportedly triggered the policy reversalBREAKING: IBM Think 2026 — watsonx Orchestrate repositioned as agentic control plane, IBM Bob reaches GA, Concert Secure Coder embeds security in developer workflowBREAKING: WEF report — 94% of cyber leaders say AI is defining force in cybersecurity; strategic AI adopters cut breach costs by $1.9M and shorten lifecycle by 80 daysBREAKING: 2026 cyberattack analysis — 17-year-old breached 7M records to buy Pokemon cards; time-to-exploit collapsed from 700 days in 2020 to just 44 days in 2025BREAKING: MIT Technology Review — AI is becoming the primary interface for democratic participation, but institutions were not designed for this worldBREAKING: Pentagon signs AI military deals with OpenAI, Google, Nvidia, Microsoft, Amazon, SpaceX, and Reflection AI — Anthropic refuses, sues over autonomous weapons concernsBREAKING: May Day 2026 — labor movement faces existential threat as Amazon plans to replace 500,000+ jobs with robots and AI automationBREAKING: AI research undergoes great pivot — focus shifts from model-centric breakthroughs to system-level deployment and autonomous scientific discoveryBREAKING: Northwestern study reveals AlphaFold2 expanded structural biology rather than replacing it — human-AI collaboration model offers template for futureBREAKING: PNNL scientists use machine learning to optimize nuclear waste vitrification at Hanford — could save hundreds of millions and reduce project timeline by yearsBREAKING: Meta raises AI capex to $125-145B, Google to $180-190B — trillion-dollar question: is the spending actually working?BREAKING: OpenAI launches GPT-5.5 agentic AI — AWS and Databricks announce managed agent services powered by GPT-5.5BREAKING: IBM launches Bob, end-to-end SDLC AI platform — $20/month Pro tier, multi-model orchestration, enterprise governanceBREAKING: Roblox Indonesia implements mandatory facial scanning for users under 16 — privacy advocates raise concernsBREAKING: DW investigation: AI industry burning trillions with no clear path to profitability — financial analysts warn of bubbleBREAKING: Microsoft drops exclusive OpenAI license — OpenAI now free to work with Amazon, Google, and any cloud providerBREAKING: China blocks Meta's $2B Manus acquisition — Beijing orders deal unwound in major cross-border AI tech rulingBREAKING: Musk vs. Altman trial begins — nine-person jury seated in Oakland, $134B OpenAI lawsuit opens TuesdayBREAKING: Big Tech's $600B AI earnings reckoning — Alphabet, Microsoft, Meta, Amazon all report WednesdayBREAKING: OpenAI proposes 4-day workweek, robot tax, and public AI wealth fund in sweeping economic policy blueprintBREAKING: Congress racing to reform FISA Section 702 before AI supercharges warrantless surveillance of AmericansBREAKING: Google reveals 75% of all new code is now AI-generated — up from just 25% eighteen months agoBREAKING: Google Gemini April Drop: native Mac app, AI music creation with Lyria 3 Pro, and Notebooks go liveBREAKING: Forbes AI 50 released — OpenAI leads at $182.6B funding, but vertical specialists are the real storyBREAKING: BCA Research warns AI trade entering 1999-style melt-up — S&P 500 could hit 9,200 before correctionBREAKING: OpenAI GPT-5.5 Instant is now the default ChatGPT model — 52.5% fewer hallucinations on high-stakes topics; new memory sources feature shows users exactly what context shaped their responsesBREAKING: Grok falls to 5th place in AI chatbot rankings — Claude surges 1,205% year-over-year as xAI loses 80+ staff and faces App Store removal threat over explicit image controversyBREAKING: Five major publishers sue Meta over Llama — Hachette, Macmillan allege millions of books pirated for AI training; model can reproduce verbatim passagesBREAKING: Trump administration strikes AI safety deals with Google DeepMind, Microsoft, and xAI — pre-release model review extended to three more major labsBREAKING: Anthropic deploys 10 AI agent templates for Wall Street — Claude Opus 4.7 leads Finance Agent benchmark at 64.37%, now integrates with Excel, PowerPoint, Word, and OutlookBREAKING: Trump White House weighs executive order to vet AI models before release — Anthropic's Mythos model reportedly triggered the policy reversalBREAKING: IBM Think 2026 — watsonx Orchestrate repositioned as agentic control plane, IBM Bob reaches GA, Concert Secure Coder embeds security in developer workflowBREAKING: WEF report — 94% of cyber leaders say AI is defining force in cybersecurity; strategic AI adopters cut breach costs by $1.9M and shorten lifecycle by 80 daysBREAKING: 2026 cyberattack analysis — 17-year-old breached 7M records to buy Pokemon cards; time-to-exploit collapsed from 700 days in 2020 to just 44 days in 2025BREAKING: MIT Technology Review — AI is becoming the primary interface for democratic participation, but institutions were not designed for this worldBREAKING: Pentagon signs AI military deals with OpenAI, Google, Nvidia, Microsoft, Amazon, SpaceX, and Reflection AI — Anthropic refuses, sues over autonomous weapons concernsBREAKING: May Day 2026 — labor movement faces existential threat as Amazon plans to replace 500,000+ jobs with robots and AI automationBREAKING: AI research undergoes great pivot — focus shifts from model-centric breakthroughs to system-level deployment and autonomous scientific discoveryBREAKING: Northwestern study reveals AlphaFold2 expanded structural biology rather than replacing it — human-AI collaboration model offers template for futureBREAKING: PNNL scientists use machine learning to optimize nuclear waste vitrification at Hanford — could save hundreds of millions and reduce project timeline by yearsBREAKING: Meta raises AI capex to $125-145B, Google to $180-190B — trillion-dollar question: is the spending actually working?BREAKING: OpenAI launches GPT-5.5 agentic AI — AWS and Databricks announce managed agent services powered by GPT-5.5BREAKING: IBM launches Bob, end-to-end SDLC AI platform — $20/month Pro tier, multi-model orchestration, enterprise governanceBREAKING: Roblox Indonesia implements mandatory facial scanning for users under 16 — privacy advocates raise concernsBREAKING: DW investigation: AI industry burning trillions with no clear path to profitability — financial analysts warn of bubbleBREAKING: Microsoft drops exclusive OpenAI license — OpenAI now free to work with Amazon, Google, and any cloud providerBREAKING: China blocks Meta's $2B Manus acquisition — Beijing orders deal unwound in major cross-border AI tech rulingBREAKING: Musk vs. Altman trial begins — nine-person jury seated in Oakland, $134B OpenAI lawsuit opens TuesdayBREAKING: Big Tech's $600B AI earnings reckoning — Alphabet, Microsoft, Meta, Amazon all report WednesdayBREAKING: OpenAI proposes 4-day workweek, robot tax, and public AI wealth fund in sweeping economic policy blueprintBREAKING: Congress racing to reform FISA Section 702 before AI supercharges warrantless surveillance of AmericansBREAKING: Google reveals 75% of all new code is now AI-generated — up from just 25% eighteen months agoBREAKING: Google Gemini April Drop: native Mac app, AI music creation with Lyria 3 Pro, and Notebooks go liveBREAKING: Forbes AI 50 released — OpenAI leads at $182.6B funding, but vertical specialists are the real storyBREAKING: BCA Research warns AI trade entering 1999-style melt-up — S&P 500 could hit 9,200 before correction
GuidesAdvanced
advanced11 min read

Skills: The Secret to Accurate and Consistent Agents

---

S
By SUPERBASH_

Skills: The Secret to Accurate and Consistent Agents

Overview

Skills are the most important feature for achieving consistent, accurate results from your OpenClaw agent. They act as a recipe book that your agent can reference on-demand, ensuring reliable execution without overloading context. This guide explains how to build, optimize, and leverage skills for production-grade agent performance.

Why Skills Matter

The Consistency Problem

Without skills:

  • Agent does it right once ✓
  • Maybe twice ✓
  • Third time: Missing sources, hallucinations, wrong format ✗

With skills:

  • Consistent execution every time ✓
  • Follows established patterns ✓
  • Maintains quality standards ✓

The Context Advantage

Critical insight: Skills do NOT overload context window

  • Skills are not part of bootstrap files
  • Only metadata is loaded at session start
  • Full skill content is read only when needed
  • You can have hundreds or thousands of skills with no performance penalty

How Skills Work

The Matrix Analogy

Think of skills like learning kung fu in The Matrix:

  • Neo doesn't know kung fu all the time - Only when needed
  • Skills are loaded on-demand - "I need to learn kung fu now"
  • No memory overhead - Not stored in active memory until required
  • Instant access - Available immediately when called

Skill Loading Process

  1. Session start: Agent loads skill metadata only (name, description)
  2. Task received: Agent identifies relevant skill
  3. Skill activation: Full skill content is read from disk
  4. Execution: Agent follows skill instructions precisely
  5. Completion: Skill content is released from context

Metadata vs. Full Content

Metadata (always loaded):

yaml
name: create_presentation
description: Creates detailed presentations with research and graphics
trigger: presentation, slides, deck

Full content (loaded on-demand):

markdown
# Presentation Creation Skill

## Research Phase
1. Use sub-agents to research topic
2. Gather sources from X, web, and documentation
3. Identify key themes and patterns

## Content Structure
- Title slide with compelling hook
- Executive summary (3-5 bullets)
- Main content (5-7 slides)
- Conclusion with actionable takeaways

## Formatting Requirements
- Use SVG graphics for visuals
- Include source links
- Maintain consistent color scheme
- Export to web-accessible format

Creating Effective Skills

Skill File Location

Default locations:

  • ~/.openclaw/skills/
  • ~/.claude/skills/
  • Project-specific: .openclaw/skills/

Check your configuration:

Where are my skills stored?

Basic Skill Structure

markdown
---
name: skill_name
description: Brief description of what this skill does
tags: [category, type, domain]
---

# Skill Name

## Purpose
What this skill accomplishes

## When to Use
Specific triggers or scenarios

## Steps
1. First step with details
2. Second step with details
3. Third step with details

## Output Format
Expected structure of results

## Quality Checks
- Validation criteria
- Success indicators
- Common pitfalls to avoid

Example: Presentation Skill

markdown
---
name: create_research_presentation
description: Creates detailed presentations with X research, web sources, and custom graphics
tags: [presentation, research, graphics]
---

# Research Presentation Creation

## Research Phase

### Sub-Agent Deployment
1. Spawn research sub-agent for X/Twitter
   - Search specific accounts: @account1, @account2
   - Collect tweets from last 24 hours
   - Extract key quotes and links

2. Spawn research sub-agent for web
   - Search Google for latest developments
   - Identify trending discussions
   - Find gaps in coverage

3. Spawn research sub-agent for documentation
   - Review official docs
   - Extract technical details
   - Note version-specific information

### Research Output Format
Each sub-agent must return:
- Key findings (3-5 bullets)
- Notable opinions with attribution
- Direct links to sources
- Patterns across multiple sources
- Gaps nobody is talking about yet

## Presentation Structure

### Slide 1: Title
- Compelling headline
- Subtitle with context
- Date and presenter

### Slide 2: Executive Summary
- 3-5 key takeaways
- Each takeaway: one sentence max
- Focus on actionable insights

### Slides 3-7: Main Content
- One concept per slide
- Supporting data or quotes
- Visual element (chart, diagram, or icon)
- Source attribution

### Slide 8: Conclusion
- Recap key points
- Actionable next steps
- Call to action

## Graphics Requirements

### SVG Generation
- Use sub-agent for custom graphics
- Consistent color palette: #2563eb, #7c3aed, #db2777
- Icons should be simple and clear
- Charts must be data-driven

### Image Placement
- Title slide: Hero graphic
- Content slides: Supporting visuals
- Conclusion: Summary graphic

## Output Format

### File Delivery
- Export to HTML with embedded SVG
- Host on accessible URL
- Provide direct link in response

### Quality Checks
- All sources linked and verified
- No placeholder text
- Graphics render correctly
- Consistent formatting throughout
- Spell check completed

## Common Pitfalls to Avoid

- ❌ Generic boilerplate content
- ❌ Missing source attribution
- ❌ Inconsistent formatting
- ❌ Placeholder graphics
- ❌ Unverified claims

## Success Criteria

- ✓ All research backed by sources
- ✓ Unique insights identified
- ✓ Professional visual design
- ✓ Actionable takeaways
- ✓ Web-accessible delivery

Managing Skills

Editing Skills

Method 1: Manual editing

bash
# Navigate to skills directory
cd ~/.openclaw/skills/

# Edit with your preferred editor
nano create_presentation.md

Method 2: Agent-assisted editing

Update my presentation skill to include:
- More emphasis on data visualization
- Requirement for executive summary
- Specific color palette

Method 3: Claude Code / Cursor

Open my skills directory in Claude Code and help me 
refine the presentation skill based on recent failures

Organizing Skills

By category:

skills/
├── presentations/
│   ├── research_presentation.md
│   ├── technical_presentation.md
│   └── sales_presentation.md
├── research/
│   ├── twitter_intelligence.md
│   ├── web_research.md
│   └── documentation_review.md
├── coding/
│   ├── code_review.md
│   ├── bug_fix.md
│   └── feature_implementation.md
└── communication/
    ├── email_drafting.md
    ├── slack_updates.md
    └── report_generation.md

Benefits:

  • Easy to find and update
  • Clear separation of concerns
  • Scalable to hundreds of skills

Skill Versioning

Track changes:

markdown
---
name: create_presentation
version: 2.1.0
last_updated: 2026-05-06
changelog:
  - 2.1.0: Added SVG graphics requirement
  - 2.0.0: Introduced sub-agent research phase
  - 1.0.0: Initial version
---

Skill Optimization

Keep Skills Focused

Good: Specific skill

create_technical_presentation.md
- Focused on technical content
- Specific formatting rules
- Clear success criteria

Bad: Generic skill

do_everything.md
- Tries to handle all presentation types
- Vague instructions
- No clear quality standards

Use Sub-Skills

Main skill:

markdown
# Create Presentation

## Step 1: Research
Execute skill: research_topic

## Step 2: Structure
Execute skill: outline_presentation

## Step 3: Graphics
Execute skill: generate_graphics

## Step 4: Assembly
Combine all components into final presentation

Benefits:

  • Modular and reusable
  • Easier to maintain
  • Can mix and match

Iterate Based on Failures

When agent fails:

  1. Identify what went wrong
  2. Update skill with specific instructions
  3. Add to "Common Pitfalls" section
  4. Test with same prompt

Example iteration:

markdown
## Common Pitfalls to Avoid

- ❌ Forgetting to include source links
  → Always add [Source](url) after claims
  
- ❌ Using placeholder text like "Lorem ipsum"
  → Validate all content is real before delivery
  
- ❌ Inconsistent slide formatting
  → Use template structure for every slide

Advanced Skill Patterns

Conditional Logic

markdown
## Execution Logic

IF task requires research:
  - Execute research_topic skill
  - Validate sources
ELSE:
  - Use provided information only

IF output format is "presentation":
  - Execute create_slides skill
ELSE IF output format is "report":
  - Execute create_report skill

Quality Gates

markdown
## Quality Validation

Before delivering output:

1. Source Check
   - [ ] All claims have citations
   - [ ] Links are valid and accessible
   - [ ] Attribution is clear

2. Format Check
   - [ ] Follows template structure
   - [ ] No placeholder content
   - [ ] Graphics render correctly

3. Content Check
   - [ ] Meets length requirements
   - [ ] Includes actionable insights
   - [ ] Free of spelling errors

If any check fails: STOP and fix before delivery

Skill Chaining

markdown
# Daily Briefing Skill

## Step 1: Data Collection
Execute skills in parallel:
- twitter_intelligence
- web_news_scraping
- documentation_updates

## Step 2: Analysis
Execute skill: identify_trends
Input: All collected data

## Step 3: Synthesis
Execute skill: create_briefing_presentation
Input: Trend analysis results

## Step 4: Delivery
Execute skill: send_to_discord
Input: Final presentation URL

Skills vs. Memory vs. Bootstrap Files

When to Use Skills

Use skills for:

  • ✓ Repeatable processes
  • ✓ Complex multi-step workflows
  • ✓ Quality standards and formatting
  • ✓ Task-specific instructions

Don't use skills for:

  • ✗ Personal preferences (use memory)
  • ✗ Project context (use bootstrap files)
  • ✗ One-time instructions

Comparison Table

FeatureSkillsMemoryBootstrap Files
Loaded at startMetadata onlyYesYes
Context impactMinimalMediumHigh
Use caseProcessesFactsIdentity
Quantity limitUnlimited~150K chars~150K chars
Update frequencyAs neededDailyRarely

Skill Performance Impact

Context Window Savings

Without skills (instructions in memory):

Bootstrap files: 120K tokens
- Personal info: 30K
- Project context: 40K
- Task instructions: 50K ← Problem!

With skills (instructions in skills):

Bootstrap files: 70K tokens
- Personal info: 30K
- Project context: 40K
- Skill metadata: <1K ← Efficient!

Skills loaded on-demand: 10K tokens only when needed

Savings: 40K+ tokens available for actual work

Model-Specific Benefits

Lower-end models (MiniMax, Qwen):

  • Critical for success - Keeps context under 40%
  • Dramatically improves consistency
  • Reduces hallucinations

Higher-end models (Opus, Sonnet):

  • Still beneficial - Better organization
  • Improves reliability
  • Scales to complex workflows

Cron Jobs + Skills

Automated Workflows

Example: Daily briefing

Cron job:

0 7 * * * Execute skill: daily_briefing

Skill handles:

  1. Data collection from multiple sources
  2. Analysis and trend identification
  3. Presentation generation
  4. Delivery to Discord/Slack

Benefits:

  • Consistent execution every day
  • No manual intervention needed
  • Quality maintained automatically

Skill-Based Automation

markdown
# Twitter Intelligence Skill

## Execution Schedule
Triggered by: Cron job at 6:00 AM daily

## Data Collection
1. Search tweets from monitored accounts
2. Filter by relevance criteria
3. Extract key information
4. Save to structured format

## Output
- JSON file with collected tweets
- Saved to: ~/data/twitter_intelligence/YYYY-MM-DD.json
- Notification sent to Discord

## Error Handling
If API rate limit hit:
- Wait 15 minutes
- Retry collection
- Log failure if still unsuccessful

Troubleshooting

"Agent not using skills"

Cause: Agent doesn't know skills exist or when to use them

Solution:

  1. Check skill metadata is clear and descriptive
  2. Add trigger keywords to skill description
  3. Explicitly request: "Use your presentation skill"
  4. Add to agent memory: "I have skills for common tasks"

"Skill produces inconsistent results"

Cause: Skill instructions are vague or incomplete

Solution:

  1. Add specific examples to skill
  2. Include quality checks
  3. Define exact output format
  4. Add "Common Pitfalls" section

"Skills not loading"

Cause: Wrong directory or permissions issue

Solution:

  1. Verify skills directory location
  2. Check file permissions: chmod 644 *.md
  3. Confirm skill file format (markdown with frontmatter)
  4. Ask agent: "List my available skills"*

"Too many skills, agent confused"

Cause: Overlapping or poorly named skills

Solution:

  1. Use clear, distinct names
  2. Add specific trigger keywords
  3. Organize into categories
  4. Remove duplicate or obsolete skills

Best Practices Summary

  1. Start with failures - Create skills for tasks that fail repeatedly
  2. Keep skills focused - One skill, one purpose
  3. Include examples - Show expected output format
  4. Add quality checks - Define success criteria
  5. Iterate continuously - Update based on real usage
  6. Organize logically - Use categories and clear naming
  7. Version your skills - Track changes over time
  8. Test thoroughly - Verify skill works before relying on it
  9. Document pitfalls - Learn from mistakes
  10. Scale fearlessly - Add as many skills as needed

Related Resources

  • Context Window Management [blocked]
  • Sub-Agents [blocked]
  • Cron Jobs [blocked]
  • Memory Management [blocked]

Duration: 7 minutes
Difficulty: Intermediate
Video Reference: OpenClaw Skills: The SECRET to Accurate Agents