Everyone has the same models. My harness is the difference.
2026-11-17
The models I use are the same ones anyone can use. What I have built around them is mine: rules, hooks, skills, scripts, memory files and tests that turn a general model into one that works the way my business works. I call that the harness. Right now I think it is where the advantage is. This is how it stood on 17 November 2026, with a month of its own logs. It changes almost daily, so read it as a snapshot.
What do I mean by a harness?
I work in Claude Code. Around it sit layers that each do a different job.
A rules file tells the model how I work. It is prose, and the model reads it.
Skills are procedures that load only when a task needs them, such as publishing a post or replying to a prospect. I have 27.
Hooks are code that runs at set points in a session, most of them around a tool call. Before a call, a hook can block it or ask me. After the call, it can leave a note. I call them gates. I have 33 numbered gates in use. Together with the hooks that have no gate number, they use 11 hook events and about 10 200 lines of code.
Scripts and memory files hold the rest. Before a finished text or a code change goes out, an independent check reads it.
The rules file says what to do. A gate enforces it for the calls it recognizes. Some gates began as a trap I had written down after a mistake. The heredoc gate is one.
I give the gate names below as examples. I am not publishing the code, and the names are in Finnish because that is how my workspace names them.
What did a month of gate logs show?
A gate that blocks, asks or flags something writes a row in a log. From 18 October to 16 November the log has 2 963 rows under 25 gate names, with rows on every day.
Here is what those rows were:
| Outcome | Rows |
|---|---|
| Blocked the call | 1 013 |
| Asked me | 309 |
| Recorded an observation | 317 |
| Let the call through, then flagged something about it | 222 |
| Stopped a session or a subagent from ending | 104 |
| No outcome recorded, almost all from before the log began recording one in late October | 998 |
The number I find most useful is who caused them. Subagents caused 1 860 rows, the main session 1 102, and one row has no actor. Some subagent types, such as Explore and Plan, do not see my rules file, and a general-purpose subagent gets it as one more piece of text. General-purpose subagents caused 1 617 of the 1 860 subagent rows. A hook runs on the tool calls it watches, whoever makes them.
A block measures friction. It does not measure harm prevented. Many blocks enforce working rules, such as no heredoc with a doubled backslash and no git in a session. The heredoc gate had rows on every one of the 30 days. Day after day the same habit came back. The log shows only the instances the gate recognized, and it blocked all of them.
Five gates, by name
| Gate | Rows | What it does |
|---|---|---|
kenoviivapari-heredocissa | 605 | Blocks a heredoc that contains a doubled backslash, because the shell tool collapses the pair and the file on disk would differ from what the agent wrote. |
gotchat-ennen-repokoodia | 343 | The first write to each repository file in a session is blocked once, and the block hands the agent the traps recorded for that file. |
suojattu-porttitiedosto | 167 | Before a write to the gate files it asks me. |
dok-kirjaamatta | 114 | A session that changed work files and wrote nothing down is stopped once at the end and told to record what it did. |
git-sessiossa | 48 | No git inside an agent session, because several sessions share one working tree. |
How were the blocks answered?
The answers came at four levels. First, the block message is the instruction. Most of them name the right way to do the thing. On 16 November a subagent hit the heredoc gate and, about 13 seconds later, wrote its edit to a file with the file tool. The log records the actions and not the reasoning, so the link between the block and the change is my reading of the order of events.
Second, when a gate keeps firing, my first answer is usually a tool rather than a stricter gate. In the last week of October, the gate that protects the gate log asked me well over 150 times, most of them for subagents analyzing the log with their own scripts. The answer was a read-only tool that knows where the log is, and a hook that gives the house rules to each subagent I start with the Agent tool. On 1 November I also changed that gate and three others to refuse a subagent's call where they would have asked me, because a subagent's question would come to me as a prompt in the middle of other work. The change took effect when I installed it on 4 November. From 4 to 13 November the heredoc gate blocked 187 times, so on 14 November I added a patch tool that takes its edit from a file. Heredoc blocks did not fall on the day it arrived or the day after: 45 on 14 November and 34 on 15 November, both heavy subagent days. On 16 November there were 17. I am not claiming the tool helped yet.
Third, the set of gates moves. There were 25 numbered gates in use on 18 October. On 31 October I read in the hook documentation that, by default, a hook that crashes or times out does not block the call. A per-hook setting can make such a failure block the call. The same day I added a gate that records calls whose gate did not finish. On 17 November I read three publications of the OWASP GenAI Security Project. Its listing gives LLM Exploit Generation Guide 2.0 the date 27 October, GenAI Data Security Best Practices v2 Part 1 the date 3 November and Securing Agentic Applications V2 the date 10 November. The same day I built a gate that asks me before a third-party MCP tool that can write is called, unless I have approved its current definition.
Fourth, false alarms. My rule is that a gate does not go red on a guess, because when a gate goes red for the wrong reason, the agent learns to work around it. When a gate cannot decide, it lets the call through, or, at a few gates, asks me. At most of those gates a subagent gets a refusal with a note on what to do instead, so that its question does not interrupt me. Since 16 November I can mark a block as unfounded. So far 0 of the 2 963 rows are marked. That is too early to mean anything, and I am saying so rather than reading it as a result.
How do I know a gate still blocks?
With mutation testing. I run two suites, one for the gate code and one for my other tools. A mutation run breaks one place in the code at a time and checks that some test notices. If no test goes red, that code could break unseen.
The gate suite had 1 086 mutations on 24 October. On 17 November it had 2 114. In 48 parallel parts 2 112 were caught, and one more was caught on a rerun by itself. The last one changed the code in a way no test can observe, so I removed it, and the suite now has 2 113.
The tool suite had 1 002 mutations on 22 October and took 64 minutes. On 17 November it had 1 941. All were caught. The 48 parallel parts left 14 open, and each of those was caught on a rerun by itself. That run took 80 minutes on an Intel Core i9-14900KS, and by my own reading the processor sat at about 250 W throughout. I run 48 parts, because at 96 parts Windows reported a system error. I have not looked for the exact limit between the two.
The models run in the cloud, but every hook and every mutation run runs on my own machine, so for the way I work I would not want a slower one: the 80-minute run would only get longer. Intel rates its laptop Core Ultra 9 285HX at 160 W maximum turbo power. In an 18-inch gaming laptop that Notebookcheck reviewed, that processor settled at 174 W under Prime95 alone. In the Extreme Performance mode used in the review, the laptop's fans reach nearly 60 dBA under heavy load, measured at 15 cm. My desktop stays quiet even at full load. The cooling is an Arctic Liquid Freezer II with three 120 mm fans pushing and three pulling on its radiator, plus six 120 mm and three 140 mm case fans.
Why does this matter when the models are the same?
In my logs the same mistakes come back day after day, and I expect the same model brings them to anyone else. My harness is the record of my own mistakes, turned into code: a trap written down, then a rule, then a gate or a tool. It is hard to copy whole, because each part answers a mistake made in my own work. That is also why I show names and numbers here and not the code. What you can take is the method: when a mistake repeats, write it down, and when it repeats again, make it a gate.
This is a snapshot
Between 18 October and 17 November the numbered gates in use went from 25 to 33. The gate mutations grew from 1 086 on 24 October to 2 113, and the tool mutations from 1 002 on 22 October to 1 941. Many recent plugin releases changed only skill text and left the hooks as they were, so a release is not always a gate change. This post describes my setup on 17 November 2026. By the time you read it, some of it has changed.
Updated 2026-11-17. The paragraph on heredoc blocks ended at 15 November, although the log window runs to 16 November. It now includes 16 November.
Corrected 2026-11-17. Several sentences said more than their sources do. Subagent permission questions do reach me, so a gate refuses them to spare me the interruption, not because no one would answer. By default a crashed or timed-out hook lets the call through. A setting can change that. A gate enforces only what it recognizes, and the heredoc gate blocked the instances it recognized. I run 48 parts as a limit I chose, not a measured maximum. The gate change of 1 November took effect on 4 November. The three OWASP publications are now named.