3 min read
Indirect Prompt Injection: Tool Output Is Not Instructions
A retrieved document, an email, or a tool result can tell the model to take an action. Delimiters do not stop it. The tool allowlist after untrusted text does.
5 min read
Copilot Auto Model Selection Now Has Tiers: Efficiency, Balance, Intelligence
GitHub Copilot's Auto picker gained efficiency, balance, and intelligence tiers in September 2026, alongside GPT-6 Astra and GPT-6.1 Sol. How the tiers change the cost-quality trade, and what admins control.
5 min read
MCP Went Stateless: What the 2026-07-28 Spec Changes for Your Server
The Model Context Protocol 2026-07-28 release removes sessions and the initialize handshake, adds routing headers and cacheable lists, moves Tasks to an extension, and deprecates Dynamic Client Registration. What to change in an MCP server.
7 min read
GitHub Copilot AI Model Comparison: Which Model to Use for Each Task
AI model comparison for GitHub Copilot Chat: GPT-5.6, Claude, Gemini, Grok, and Kimi, plus when Auto is the cheaper default.
3 min read
MCP Tool Descriptions That Route the Model to the Right Call
How to write MCP tool names, descriptions, and schemas so the model picks one tool on purpose, and what to do when two tools sound the same.
3 min read
Put the Workflow in Code When the Agent Keeps Improvising
Which steps belong in a fixed program and which belong in a model, so an agent stops skipping the check that production depends on.
4 min read
MCP Resources vs Tools vs Prompts: Stop Putting Everything in a Tool
The three primitives in the Model Context Protocol, what the model actually sees, and a rule of thumb for Dataverse, files, and internal APIs.
7 min read
Evaluating LLM Applications: Getting Past 'It Looks Good to Me'
Shipping LLM features on vibes works until a prompt tweak silently breaks ten other cases. Here is how to build evals that catch regressions before your users do.
7 min read
Getting Structured Output From an LLM Without the Heartbreak
Asking a model to 'return JSON' and parsing the result is how you get 3 a.m. pages. Tool schemas, constrained decoding, and validation turn a probabilistic text generator into a reliable API.
7 min read
Chunking Strategies for RAG: Where Retrieval Quality Is Won or Lost
Most RAG systems that retrieve bad context aren't failing at embeddings or reranking — they're failing at chunking. How you split documents quietly decides what your model can ever find.
10 min read
An MCP Server for Dynamics 365 Finance and Operations: Natural-Language Access to ERP
How to expose Dynamics 365 Finance and Operations to Claude through MCP — the architecture, the data entity choices, the auth model, and the operations that should never be one prompt away.
6 min read
Claude Code Hooks: Automating Repo Guardrails Without Pre-Commit Fatigue
Use Claude Code hooks to enforce policy at the moment actions happen — before edits, before tool calls, on session start — without relying on pre-commit hooks everyone learns to bypass.
7 min read
Building Your First MCP Server in Python: A Hands-On Walkthrough
From zero to a working Model Context Protocol server in about 100 lines — tools, resources, a local client test, and the traps that will bite you on day one.
7 min read
Prompt Caching with Claude: What It Saves, Where It Fails, and How to Set It Up in Production
A precise look at Claude's prompt caching — how the cache hit math actually works, the 5-minute and 1-hour TTL variants, and the subtle mistakes that silently disable it.
8 min read
Tool Use with Claude: Parallel Calls, Retries, and the Patterns That Survive Production
The tool-use features of the Claude API look straightforward in the docs. These are the production patterns that separate a demo from a system that handles real traffic — parallel tool calls, retry loops, error recovery, and tool_choice tactics.
7 min read
Claude Agent SDK vs Direct API: When to Reach for Each
The Claude Agent SDK is the fastest path to a capable coding-style agent. The raw messages API is the fastest path to understanding what your agent is doing. Here's when each is the right tool.
8 min read
Designing a Production-Ready RAG System: What Actually Matters
Retrieval-augmented generation demos look great. Production RAG is a different engineering problem — chunking, hybrid retrieval, reranking, evaluation, and the failure modes nobody mentions at conferences.
8 min read
Vector Databases Compared: pgvector, Qdrant, Pinecone, and When You Don't Need Any
A practical comparison of the vector databases people actually deploy in 2026 — and an honest look at when a vector database is the wrong tool for the job.
8 min read
Event-Driven Architecture for LLM Pipelines: Queues, Retries, and the Cost of Synchronous Chains
LLM-powered pipelines break in the same ways old monoliths did — when you stack synchronous calls. Queues, events, and the patterns that make a six-step AI pipeline survive a partial outage.