Your agent needs a context file for every API (or it'll reinvent OAuth every time)

I watched my coding agent spend 47 minutes trying to figure out the Stripe API yesterday. Not because Stripe is complicated, but because it kept making the same authentication mistakes over and over.

The problem isn't that Claude doesn't know APIs. It's that every conversation starts from zero. Your agent rediscovers rate limits, reinvents error handling, and re-reads documentation that hasn't changed in months.

The fix is dead simple: context files for every API your agent touches.

Here's what I mean. Instead of letting your agent fumble through Stripe docs every time, create stripe_context.md:

# Stripe API Context

## Authentication
- Use test key: sk_test_... (stored in .env as STRIPE_SECRET_KEY)
- Always set Stripe-Version header: 2023-10-16

## Common Patterns
- Customer creation: always include email and name
- Payment intents: confirm=true for immediate processing
- Webhooks: verify signature with stripe.webhook.construct_event()

## Rate Limits
- 100 requests/second in test mode
- Use exponential backoff on 429 errors

## Error Handling
- card_declined: show user-friendly message
- insufficient_funds: suggest lower amount
- All others: log and show generic error

Now your agent doesn't rediscover that Stripe needs the version header. It doesn't waste 15 minutes figuring out webhook verification. It just works.

Do this for every API:

• GitHub — token scopes, rate limits, webhook payloads
• OpenAI — model capabilities, context windows, pricing tiers
• Twilio — phone number formats, SMS limits, error codes
• AWS — IAM patterns, region preferences, service limits

The pattern is always the same: authentication, common operations, gotchas, and error handling. Anything your agent has to figure out twice goes in the context file.

Pro tip: Update these files every time your agent hits a new error. Turn every API mistake into institutional knowledge.

I started doing this three months ago. My coding agent went from spending 30% of its time on API discovery to maybe 5%. Tasks that used to take an hour now take 20 minutes.

The difference between a coding agent that works and one that flails around? Documentation that travels with the conversation.