Your agent needs a skill inventory (or it'll rediscover the same tools forever)

Your agent just spent 15 minutes figuring out how to parse a CSV file. Last week, it spent 20 minutes on the same problem. The week before that, it reinvented JSON parsing from scratch.

This isn't a memory problem — it's a skill inventory problem.

Most agents treat every task like they're starting from zero. They have access to tools and APIs, but no systematic way to remember what they've learned about using them effectively. So they rediscover the same patterns, make the same mistakes, and burn tokens on problems they've already solved.

The fix is a skill inventory — a living document that captures not just what tools your agent has, but what it's learned about using them well.

Here's the pattern that works:

# SKILL_INVENTORY.md

## Data Processing
### CSV Operations
- **Tool**: pandas via code execution
- **Best practices**: Always check encoding first (utf-8-sig for Excel exports)
- **Common gotchas**: Empty cells become NaN, dates import as strings
- **Learned patterns**: Use pd.read_csv(file, encoding='utf-8-sig', parse_dates=['date_column'])
- **Last used**: 2024-01-15
- **Success rate**: 95% (fails on malformed headers)

### API Authentication  
- **Tool**: requests library
- **Best practices**: Store credentials in environment variables
- **Common gotchas**: Rate limits aren't always in headers
- **Learned patterns**: Implement exponential backoff, check for 429 status
- **Last used**: 2024-01-14
- **Success rate**: 90% (OAuth refresh tokens still tricky)

The magic isn't in the format — it's in the **learned patterns** section. This is where your agent documents the hard-won knowledge that comes from actually using tools in production.

Instead of rediscovering that Excel exports need utf-8-sig encoding, your agent checks its inventory and knows immediately. Instead of hitting rate limits again, it sees the exponential backoff pattern it learned last time.

Pro tip: Update the inventory after every significant task. Make it part of your agent's completion routine: "Task finished. Update skill inventory with any new patterns learned."

Here's how to implement this:

• Create the inventory file — Start with your agent's most-used tools
• Add update triggers — After completing tasks, after hitting errors, after discovering new patterns
• Include failure modes — What doesn't work is as valuable as what does
• Track success rates — Helps your agent choose between multiple approaches
• Date everything — Skills decay, APIs change, patterns become obsolete

The difference is dramatic. Instead of burning 200 tokens figuring out CSV encoding every time, your agent spends 20 tokens checking its inventory and gets it right immediately.

Your agent stops being a goldfish that forgets everything between sessions and becomes a craftsperson that gets better with experience.

The skill inventory turns every task into training data for future tasks. Your agent doesn't just complete work — it builds expertise.