9 min read
Screenshots Fill Every Later Turn: Downscale, Then Drop the Pixels
A screenshot is not a log line. It is input on every later request. Downscale on ingest, extract the text, and drop the pixels after the model has read them.
Theme
AI & Agents
Practical guides on Claude, the Model Context Protocol (MCP), agent architectures, prompt caching, and tool use in production.
24 articles
7 min read
Tool Results Fill Every Later Turn: Truncate Before You Append
A tool_result is not a log line. It is input on every later request. Cap rows at the tool, keep pairing, and compact only completed turns.
6 min read
Claude Thinking Blocks: Echo the Signature or the Tool Loop Breaks
Pass thinking and redacted_thinking back unchanged. Effort steers depth; max_tokens caps spend.
7 min read
Streaming LLM Output: SSE Events, Partial Tool JSON, and Abort
Tokens arrive before the JSON is valid. Handle SSE events, concatenate partial_json, and abort the request.
7 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.
7 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.
5 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.
8 min read
Agentic ERP in September 2026: What Is Available, in Preview, and Still a Date
A precise reading of Microsoft's 23 September 2026 Dynamics 365 ERP announcement, the 10.0.49 previews behind it, and the financial-journal dates that are not this fall.
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.