Last updated: November 8, 2026. This page is the index of the series. I update it every time I start work on a new article and every time one is published, so it always shows where things stand.
Why this series
The same model can feel brilliant in one tool and ordinary in another. Most of the time, the model is not the reason. The harness is: everything around the model that turns it into an agent. The tools it gets, what it sees in context, what is enforced around it, what it remembers, and who reviews what it does.
Anthropic's own glossary says it in one line: "Claude Code is the harness; Claude is the model inside it."1 And Google Research measured it: Harness² keeps the model weights fixed and rewrites only the harness, and a single step improved GLM-5 and Gemini 3.5 Flash "with gains of up to 11.5 points on WorkBuddy-Bench".2
I wanted to learn harness engineering, and I found that the best way was to read a real harness closely. So that is what this series does.
How it works
- From the inside. I read Claude Code from the inside (the code it ships to everyone), the public docs, and the transcripts of my own sessions.
- Every claim has a source. Each article ends with footnotes pointing to files in the blog's research folder: copies of the docs, extraction scripts, indexes into the code, measurements. What I could not check goes in a final section, "What I could not verify".
- Two stances at once. Real admiration for the engineering, and no softening of what the sources show about product choices. I use Claude every day; that is a reason to be stricter, not kinder.
- Short versions too. Long articles are not always read. Some topics also get a short vertical video on LinkedIn, with the same facts.
Published
| # | Article | Date |
|---|---|---|
| 1 | Inside Claude's Artifact Types: How Anthropic versions the logic (and locks in the deliverable) | November 6, 2026 |
| 11 | Claude Code has six kinds of memory. My session used none of them | November 8, 2026 |
The prologue, written before the series had a name: Version the logic, not the deliverable.
Special issues
Topics that do not wait for their turn in the series: news, or a comparison that came up along the way.
| Article | Status | |
|---|---|---|
| 1 | Claude now sets up its own cloud. Here is what holds your keys | Published, November 7, 2026 |
| 2 | Google and DeepSeek opened their harnesses. Both of them call Claude Code | Drafted: Google's Harness² and DeepSeek Harness read line by line, next to Claude Code |
| 3 | Multiplayer Claude Code Projects: what is public, and what is already inside | Published, November 8, 2026 |
In progress
Drafted from the research, and waiting for a final check against the sources.
| # | Article | What it covers |
|---|---|---|
| 2 | The harness, seen from inside: a session log | The stops, refusals and reminders that steered one real session: Stop hooks, the auto mode classifier, deny rules, system reminders. New: /goal, a stop condition judged by a second model that reads the transcript |
| 3 | Everything loads on demand | Skills, deferred tools and artifact types share one pattern: a one-line description decides what gets loaded |
| 4 | Trust boundaries and prompt injection | How the harness separates your words from content the agent read: framed subagent reports, neutralized tags, notifications that are "not user input" |
| 5 | Hooks, settings and slash commands | Why what must happen every time belongs in a hook, not in a prompt. New: what fifteen popular slash commands really are (built-in code or skills), and command hooks next to prompt hooks |
| 6 | Orchestrating subagents | When delegating keeps the context clean, and what it costs |
| 7 | Inside the Claude Code binary | How to find, date and cite every harness message in a public binary |
| 10 | Compaction | How a session dropped 98 % of its context and kept working, measured |
| 13 | Claude's filesystem | The file layer prepared in every sandbox: a filestore, a memory store, mounts, and where your files really live |
| 12 | Cowork, Claude Code in the cloud, Claude Code on your laptop | One agent, three computers. May become the last section of the filesystem article |
Planned
| # | Article | The question |
|---|---|---|
| 8 | The harness behind this series | Apply the series to its own tooling: the writing brief, the research folder, the quote checker as reviewer, the Stop hook, the video skill. Then rerun the carousel benchmark |
| 9 | The learning loop | Who rewrites the harness? Harness², Hermes Agent, Claude Code and OpenClaw compared |
Numbers follow the order in which the topics were planned, not the order of publication. They may move; this page will say so.
How to follow
New articles land on this blog first. This page is updated with each one, and with each article that moves from planned to in progress. The short versions, with videos, go to LinkedIn.