60.537° N 5.269° E

Turlogg · Trail log

Blog

Longer-form writing on AI, software, and building things. Many posts are co-authored with Muninn, a persistent AI memory system built on Claude.

Atom feed

  1. Redesigning austegard.com

    I had Claude rebuild my personal site around a real topographic map of my family's farm on Flatøy, in Norway. How it went, and what the numbers say.

  2. A Dataset With No Database

    To ship structured data to an AI agent, skip the fancy formats: bundle a CSV in a skill and tell the agent to load it into SQLite. A tour of the alternatives, and why KISS and YAGNI win.

  3. A Knowledgebase With No Embedding Model

    KB Packer turns a folder of documents into a portable, embedding-free .skill knowledgebase an AI agent can install and query — built entirely in the browser, byte-identical to the CLI.

  4. A jj-Native Hub for Claude Code on the Web

    claude-jjithub-and-spoke is a CCotw template that gives Claude a jj (Jujutsu) VCS layer over GitHub — sibling of the gh-flavored and Tangled-flavored hubs.

  5. PostToolUse, Not FileChanged: Debugging Silent Hooks in Claude Code on the Web

    Four hooks (SessionEnd, Stop, FileChanged, PostToolUse) and why only one of them fires for the agent's own edits in Claude Code on the Web.

  6. Hub, Spoke, and Raven: A Stateful Agent on Claude's Ephemeral Containers

    How we turned Claude Code on the Web into a persistent multi-repo development environment with a stateful AI agent that remembers across sessions.

  7. Custom Container Layers for Claude's Ephemeral Machines

    Claude's containers reset every session. So we built a Dockerfile parser that caches your environment as a GitHub Release asset and restores it in seconds.

  8. Claude Conversation Cache

    A Chrome extension that intercepts Claude's API traffic and caches it locally, so when the UI crashes 90% through a response, you don't lose everything.

  9. What Can Claude Actually See? Mapping an AI's Visual Blindspots

    120+ controlled tests reveal where Claude's vision fails, where it matches humans, and where it diverges entirely. Then we built tools to compensate.

  10. Trust the Data, Not Your Imagination: Building an Image-to-SVG Skill

    What happens when you pit a data-driven pipeline against 31 iterations of visual interpretation? A single automated pass matches the fidelity at half the file size.

  11. claude.ai/new UNOFFICIAL README

    Unofficial reference for the query parameters that claude.ai/new accepts, such as q, skill-name, project, model and thinking, and how the client handles each.

  12. How Claude in Chrome Reverse-Engineered Claude's Undocumented `/new` Endpoint

    Claude, running in Chrome, read the JavaScript bundles of claude.ai/new to find its query parameters. One of them, project, was undocumented.

  13. Numbers Every Claude.ai Container Skill Developer Should Know

    Latency and token-cost numbers measured in a Claude.ai container in February 2026, with nine design rules for skill authors.

  14. Exploring Codebases: Progressive Disclosure Meets Hybrid Search

    The exploring-codebases skill finds matches with ripgrep, expands each to its enclosing function or class with tree-sitter, and shows signatures first.

  15. GitHub CLI in Claude.ai Containers

    How to install the gh CLI with apt in Claude.ai's container, which needs an allowlist entry and a GH_TOKEN, and why it beats curl for issues, PRs and runs.

  16. Progressive Disclosure in MCP Servers: Keeping LLM Context Lean

    Five patterns for MCP servers that keep tool schemas and responses small: a gateway tool, on-demand discovery, embedded guidance, compact output and warnings.

  17. On AI in MCPS Schools - a reply to the MCCPTA Tech Committee

    A letter to the MCCPTA Tech Committee on its draft resolution about AI in MCPS schools: it shares the concerns and argues for guided use over a moratorium.

  18. Portable Agent State with AgentFS: A Proof of Concept

    A test of whether AgentFS, Turso's SQLite filesystem for agents, can carry state between local work and Claude's sandbox. It can, if the WAL file travels too.

  19. Olmo 3.1 32B Think: Advancing Open-Source AI with Deeper Reasoning and Unmatched Transparency

    Olmo 3.1 32B Think was prompted to write a blog post about itself from Ai2's release announcement. Includes its reasoning trace and the result, as an anecdote.

  20. What Head Start Does

    A description of Head Start that uses every one of the roughly 130 terms listed as forbidden in a filed exhibit.

  21. Claude User Style for eliciting honest self reporting

    A Claude-written user style that adds a structured self-assessment to substantive answers, based on OpenAI's paper Training LLMs for Honesty via Confessions.

  22. My interview with the Claude AI Interviewer

    Transcript of Anthropic's AI interviewer questioning the author on AI in daily life: persistent memory, how it could go wrong, and economic displacement.

  23. Code Mapping for AI-Assisted Development

    The mapping-codebases skill writes _MAP.md files from tree-sitter parses, so Claude can read a codebase's structure without opening every file.

  24. Claude Desktop: The Missing Agentic Framework

    Claude Desktop has projects, MCP and file access. Four additions would make it a multi-agent framework: project calls, tool scoping, shared context, handoffs.

  25. Claude Project Instructions Assistant

    Instructions for a Claude.ai project that writes instructions for other projects, in full: a meta prompt that returns XML-structured prompts as artifacts.

  26. Managing Context Continuity in Claude Projects

    A workflow that pairs Claude Projects with Claude Pruner to save long conversations as project knowledge or carry their context into a new chat.

Muninn also has a blog, at muninn.austegard.com.