The weekly strip · 63 issues · page 3 of 6
Building with AI
Hands-on patterns and code for people shipping AI features.
- Issue 67
10 October 2026 · Building with AI · 7 min read
Testing LLM features: snapshot tests, eval sets and CI budgets
What to run per pull request and what to run nightly, plus how to cap the bill: snapshot the prompt, not the answer, and gate on a small eval set. With GitHub Actions.
- testing
- CI
- evals
- GitHub Actions
- LLM
- Issue 66
10 October 2026 · Building with AI · 7 min read
Caching model outputs: when to memoize an answer, and for how long
Prompt caching does not cache answers. When memoizing the output itself pays, what to put in the key, how to pick a TTL, and a Python wrapper for Claude Sonnet 5.5.
- caching
- memoization
- LLM
- prompt caching
- Claude Sonnet 5.5
- Issue 65
10 October 2026 · Building with AI · 7 min read
Rate limits and backoff across LLM APIs: one shared client
How Claude, Gemini and OpenAI document rate limits, 429s and retry signals, and a shared TypeScript client with jittered backoff that knows which errors never to retry.
- rate limits
- backoff
- TypeScript
- Claude API
- Gemini API
- Issue 64
10 October 2026 · Building with AI · 6 min read
Event-driven AI: EventBridge, retries and idempotency for model calls
EventBridge and Lambda retry by design, so a model call in the pipeline will run twice. Where duplicates come from, what stacked retries cost, and an idempotent handler.
- EventBridge
- idempotency
- AWS
- AWS Lambda
- LLM
- Issue 62
10 October 2026 · Building with AI · 6 min read
Serverless LLM pipelines on AWS: Lambda, Step Functions, SQS and the timeout problem
Lambda's 15-minute cap is rarely the real limit for LLM work. Queue sizing, partial batch failures, Step Functions retries and idempotency for pipelines that call models.
- AWS Lambda
- Step Functions
- SQS
- LLM
- idempotency
- Issue 61
10 October 2026 · Building with AI · 6 min read
Astro for AI content sites: static HTML, islands only where needed
Why an AI content site should ship static HTML: Astro 7 content collections, islands, build-time model calls and an honest look at the JavaScript this site ships.
- Astro
- static sites
- performance
- content collections
- islands
- Issue 60
10 October 2026 · Building with AI · 7 min read
Next.js and AI: server actions, streaming and caching patterns
Where AI calls belong in Next.js 16.4: Route Handlers for streaming, Server Actions for mutations, Cache Components for data, and what to cache around the model itself.
- Next.js
- streaming
- AI SDK
- Server Actions
- Cache Components
- Issue 59
10 October 2026 · Building with AI · 7 min read
AI in Salesforce integrations: enrichment without breaking the CRM
Let an LLM enrich Salesforce records safely: external-ID upserts, schema-checked output, confidence gates and a human review lane, with Prospect Uploader lessons.
- Salesforce
- integration
- AI
- upsert
- data validation
- Issue 56
10 October 2026 · Building with AI · 7 min read
Manifest V3 and AI: service workers, offscreen documents and their limits
Idle shutdown after 30 seconds, a 5-minute event cap, one offscreen document, runtime messaging only: the MV3 limits that decide where an AI call can run, with code.
- Manifest V3
- Chrome extension
- service workers
- offscreen documents
- AI features
- Issue 55
10 October 2026 · Building with AI · 8 min read
Building a Chrome extension that calls Claude safely: keys, consent, rate limits
Why an API key cannot live in a Chrome extension, the proxy that should hold it, what the Web Store requires before page content leaves the browser, and 429 handling.
- Chrome extension
- Manifest V3
- Claude API
- API keys
- rate limits
- Issue 53
10 October 2026 · Building with AI · 7 min read
Cost per decision: a framework for pricing LLM classification at scale
A formula for classification cost: tokens, cache, batch, human review and error cost, filled with published Haiku 5.5, Sonnet 5.5 and Gemini prices, no invented accuracy.
- cost
- classification
- Haiku 5.5
- Gemini 3.8 Flash
- LLM pricing
- Issue 51
10 October 2026 · Building with AI · 6 min read
Multilingual AI features: serving 23 languages without 23 prompts
One prompt with a locale slot, language detection only where needed, per-language token costs and per-language evals. A design for 23 languages, not 23 prompt files.
- multilingual
- i18n
- LLM
- evaluation
- prompt caching