Build with AI tools

Give your assistant the docs

AI tools write better code when they read the real reference instead of guessing. Every page here has a Copy page for AI button, and the whole API is available in machine-readable form.

Add a project rules file

Put this in CLAUDE.md for Claude Code, .cursor/rules for Cursor, or .github/copilot-instructions.md for Copilot. It teaches the assistant how talks works before it writes a line.

## talks API
- Base URL: https://api.talks.africa/v1
- Auth: header "Authorization: Bearer $TALKS_KEY". Read the key from env, never hard-code it.
- Use tk_test_ keys in development and tests. Never commit tk_live_ keys.
- Phone numbers must be international format: +254712345678.
- Send: POST /v1/sms/send, POST /v1/email/send, POST /v1/whatsapp/send.
- Bulk: POST /v1/sms/send-bulk (10 req/min), not loops of single sends.
- Errors return { error: { type, code, message, param, request_id } }.
  Retry only 429 (respect Retry-After) and 5xx with backoff. Never retry other 4xx.
- Delivery results arrive by webhook; do not poll in a tight loop.
- Full reference: https://talks.africa/docs/llms-full.txt

Prompts that work

1. 1

   Point it at the reference

   Ask the assistant to read the docs first, so it uses real endpoints and fields.

   Read https://talks.africa/docs/llms-full.txt, then add a
   sendOrderSms(order) function that texts the customer when
   an order ships. Use $TALKS_KEY and handle errors as the docs say.

2. 2

   Ask for error handling up front

   Name the cases you care about: out of credits, invalid numbers, rate limits.

   Handle insufficient_credits by alerting #finance on Slack,
   invalid_phone_number by flagging the customer record, and
   rate limits with Retry-After.

3. 3

   Test with a test key

   Have it write tests that run against a tk_test_ key, so nothing real is sent.

   Write integration tests for sendOrderSms using TALKS_TEST_KEY.
   Assert on status and credits_used in the response.

MCP server

Stay safe