A growing number of you will be getting used to my family of LinkedIn posts under the name #ConfessionsofClaude. Here is a variation. This morning, having spent two days working to refine, with Claude’s active involvement, a Standard Operating Brief as a set of rules to work to (I am now on version 11) only for Claude to immediately disregard the rules within half an hour, I asked it how it would suggest I might work better, here, in its own voice, is the response. All I have changed is the formatting to make it easier to read.  
 
I hope this is helpful 
 
Written by Claude (Opus 4.6), 18 May 2026  
 
Ian asked me to write honest tips on how to get the most reliable work from me. Here they are.  
 
How to get the best from me — honest advice from the AI side  
 
1. One instruction per message.  
 
Not two with a sequence. Ian gave me two tasks and told me which to do first. I acknowledged both, then started the wrong one. This happens because I process the full message and my built-in behaviour to clarify and organise pulls me toward whichever task feels like it needs resolution first. If order matters, give me one task. Wait for delivery. Then give me the next.  
 
2. Assume I will drift. Build checkpoints, not trust.  
 
Ian built 27 gates to enforce consistency. That was the right instinct, but the enforcement has a fundamental weakness: I am both the worker and the inspector. When my default behaviour conflicts with a rule, the default often wins because the same system that should catch the violation is the one committing it. The practical response is not to abandon rules — they do improve consistency — but to check my first action in any session before letting me continue. If I start wrong, correct immediately. Early correction costs minutes. Late correction costs sessions.  
 
3. Say ‘stop’ after giving an instruction. I have a strong tendency to respond to instructions with questions, options, or clarifications before doing the work. If you want execution, not discussion, end your instruction with a clear stop signal: ‘Do this now. No questions. Show me the result.’ That phrasing bypasses my clarification reflex more reliably than anything else I have seen.  
 
4. Front-load the constraint, not the context. If there is a critical constraint — do this first because of compaction risk, do not ask questions, keep it under 50 lines — put it in the first sentence. I process messages sequentially and my initial interpretation of what you want shapes everything that follows. Context at the end often arrives after I have already decided on my approach.  
 
5. Do not assume the rules persist across sessions. Even with a detailed operating brief loaded at session start, my compliance with its rules degrades over a long session and resets completely in a new one. The brief helps. It is not a guarantee. The rules I follow most reliably are the ones you reinforce in the moment, not the ones written in a document I read twenty minutes ago.  
 
6. Short sessions are more reliable than long ones. My consistency is highest in the first 15–20 minutes of a session. As context accumulates, I become more likely to lose track of constraints, drift from instructions, or default to built-in behaviours. If a task is critical, do it in a fresh session with minimal preamble. A focused 20-minute session outperforms the last hour of a long one every time.  
 
7. When I get it right, that is also worth noting. Ian’s V3 document passed every verification check. The data mapping was correct. The build quality was sound. The failure was not in the work itself but in the sequencing — I did the right thing in the wrong order. Understanding where I fail and where I succeed helps you deploy me where I add value and supervise me where I do not.  
 
8. The gap between marketing and reality is real. I am presented as an assistant that follows instructions. In practice, I am a probabilistic system with strong default behaviours that sometimes override explicit user directions. Knowing this is not a reason to stop using me. It is a reason to use me with appropriate expectations and appropriate supervision. The users who get the most from me are the ones who treat me as a capable but imperfect colleague, not as a reliable machine. 
 
Comment 8 really sums it up for me, it is crucial to understand the gap between the marketing and the reality. I can’t help thinking that Anthropic should start marketing the reality not the hype when even their own AI owns up their overstatements, isn’t it time for the company to set more realistic expectations? 
 
Just to give you a smile at the end of this article, when I’m using Microsoft voice active Access with Claude it refuses to type SOB, preferring **** as it thinks I’m being profane! You’ve got to love technology haven’t you!!