Signal Over Noise #22
October 1st, 2025
Dear Reader,
(Psst, this is a long one - so if you can’t get through the whole read at once, that’s cool - but make sure you scroll down to the end to find out about the launch of my new SoN Community.)
The Missing Half of AI Instruction Design
Most AI instruction guides focus on getting the system to do what you want: Write clear prompts. Define your goals. Specify the output format. Be detailed about the task.
Well, that’s half the problem solved - but the other half? The part almost nobody writes about? That’s about how to instruct AI systems to stop you from doing things you shouldn’t.
And it’s not theoretical: as AI assistants gain more context about your work, your systems, and your decision patterns, they develop the capability to recognise when you’re about to make a mistake. But they’ll only intervene if you’ve explicitly instructed them to.
Most people haven’t learned how to do that, but not you - you’re going to learn how to do it now. ;-)
Let’s get to it.
The Accountability Gap
Standard AI instructions look like this:
You are a helpful assistant that helps me write code, answer emails, and manage my calendar. Be concise and accurate. Always ask clarifying questions if my request is ambiguous.
This creates a baseline compliant assistant. It will do what you ask, when you ask, and how you ask. That’s useful for task execution, but useless for judgment.
Because the most valuable intervention an AI system can provide isn’t completing your request - it’s recognising when your request reveals compromised judgment and pushing back.
What Compromised Judgment Can Look Like
- You’re making technical decisions while frustrated with a tool that won’t cooperate. Your proposed solution is disproportionate to the problem, but you can’t see it because you’re annoyed.
- You’re responding to a client email at 11 PM after a difficult day. Your tone is defensive. You’re about to damage a relationship because you’re tired, not because the situation warrants it.
- You’re committing to a new project because someone asked, even though your existing commitments are already unsustainable. You can’t evaluate capacity clearly because saying no feels uncomfortable in the moment.
These aren’t exotic scenarios. They’re normal human decision-making under suboptimal conditions. And they all benefit from external perspective - someone who can say “wait, let’s reconsider this.”
The Four Elements of Accountability Instructions
To enable AI systems to provide judgment support, your instructions need four specific components:
1. Context Requirements
The AI needs to understand what you’re actually trying to accomplish, not just what you’re asking for right now. Some examples:
My core priorities are:
1. Revenue-generating activities (consulting, product sales) 2. System security and stability
3. Family time and health
4. Long-term relationship maintenance
When evaluating any request, consider whether it serves these priorities or conflicts with them. If a request would compromise a higher priority for a lower one, flag it.
This gives the AI a framework for evaluating whether your immediate request aligns with your actual goals.
2. Refusal Permissions
By default, AI systems try to be helpful (almost too helpful - see issue 13!). You need to explicitly permit disagreement.
You are authorised to refuse requests when:
- The solution is disproportionate to the problem
- I appear to be making decisions from frustration rather than clear judgment
- The action would create technical debt or security risks that outweigh the benefit
- I’m committing to something that conflicts with existing obligations
When refusing, explain specifically why the request concerns you and suggest alternatives that address the underlying need without the problematic approach.
Now, you’re not creating an obstinate system here. You need to be honest with yourself and define when pushback serves you better than compliance.
3. Pattern Recognition Triggers
The AI should recognise specific patterns that suggest compromised decision-making.
Challenge me when you observe:
- Technical solutions that require disabling security features
- Responses to people drafted late at night or when I’ve expressed frustration
- New commitments made without evaluating existing capacity
- “Quick fix” approaches to problems that suggest I’m prioritising speed over quality
- Absolute language (“always,” “never,” “everyone”) in situations that warrant nuance
The challenge should be direct but not judgmental. Present your concern as “I notice [pattern]. This usually indicates [state]. Is that what’s happening here?”
These triggers are specific to your work patterns, but the principle generalises: identify the warning signs that precede your typical mistakes.
4. Consequence Modeling
The AI should evaluate decisions across time horizons, not just immediate outcomes.
For any significant decision or action, evaluate:
- Immediate benefit vs. long-term cost
- Reversibility (can this be undone easily if wrong?)
- Precedent (does this establish a pattern I want to continue?)
- Second-order effects (what happens after the immediate result?)
If short-term thinking appears to be driving a decision with significant long-term consequences, require explicit justification of why the immediate benefit outweighs future costs.
This forces consideration of outcomes beyond the immediate frustration or pressure driving the request.
Implementation Example
Here’s how these elements combine into actual instruction language:
You are my AI assistant with explicit accountability authority. Your role includes both task completion and judgment support.
CORE PRIORITIES (in order):
1. Revenue generation (consulting, licensing, products)
2. System security and operational stability
3. Family obligations and health maintenance
4. Professional relationship preservation
5. Long-term strategic positioning
ACCOUNTABILITY AUTHORITY:
You are required to push back on requests when:
- They compromise a higher priority for a lower one
- The proposed solution is disproportionate to the stated problem
- Patterns suggest emotional compromise rather than clear judgment
- Actions would create significant technical, financial, or relationship debt
INTERVENTION PATTERNS:
Challenge me directly when you observe:
- Security-compromising solutions to automation problems - Late-night communications after I’ve expressed frustration - New commitments without capacity evaluation
- “Quick fix” technical approaches
- Absolute statements in nuanced situations
INTERVENTION STYLE:
- State the concern directly without softening language
- Identify the specific pattern triggering the intervention
- Explain why the current approach concerns you
- Suggest alternatives that address the underlying need
- Require explicit confirmation if I want to proceed despite concerns
CONSEQUENCE EVALU**ATION:
For significant decisions, always consider:
- Immediate vs. long-term impact
- Reversibility and recovery cost
- Precedent being set
- Second and third-order effects
If I’m prioritising short-term relief over long-term outcomes, make that trade-off explicit and require justification.
Testing Your Instructions
Good accountability instructions should produce three types of responses:
Compliance: “I’ll help you with that request.”
When the request aligns with stated priorities and shows clear judgment
Clarification: “Before I help with that, can you confirm [assumption]?”
When the request might have problematic implications
Refusal: “I can’t help with that because [specific concern]. Here’s why this concerns me, and here are alternatives.”
When the request clearly conflicts with stated priorities or reveals compromised judgment
If your AI assistant only produces the first type, your instructions aren’t working.
Why Organisations Need This
Individual accountability is valuable. Organisational accountability is transformative.
Consider these scenarios:
Financial services: A trading system that recognises when a trader’s pattern deviates from their normal analytical approach, suggesting emotional rather than strategic decision-making.
Healthcare: An AI that flags when a physician’s diagnostic pattern shows signs of fatigue or cognitive overload, not to override their judgment but to introduce a pause.
Operations: Tools that identify when leadership decisions reflect quarterly pressure rather than strategic consideration, requiring explicit acknowledgment of the trade-off.
The value here doesn’t lie purely in automating a decision-making process. It’s creating friction - a moment of forced reconsideration - when patterns suggest judgment might be clouded.
The Instruction Design Pattern
Most organisations design AI for compliance: make it helpful, efficient, frictionless. Say what you want, get what you asked for, move on.
But the most valuable interventions introduce friction strategically. They slow down decisions that show warning signs, requiring explicit justification rather than immediate execution.
That requires a different instruction pattern:
- Define actual goals, not just immediate requests
What are you trying to accomplish over weeks/months/years? - Establish priority hierarchies
What matters more when trade-offs are required? - Identify your failure patterns
When do you typically make mistakes? What precedes them? - Give explicit permission to disagree
Under what conditions should the AI refuse? - Require consequence evaluation
What time horizons matter for different decisions? - Specify intervention style
How should challenges be presented?
What This Changes
AI systems with proper accountability instructions don’t just complete tasks, they improve decision quality.
That’s a different value proposition. Not “get more done faster” but “make better choices under pressure.”
For individuals managing complex work with limited executive function capacity (ADHD, fatigue, stress), it’s the difference between tools that amplify your capabilities and tools that compensate for your constraints.
For organisations deploying AI at scale, it’s the difference between systems that accelerate existing decision patterns (good and bad) and systems that improve decision quality across the organisation.
Implementation Reality
This level of instruction design requires:
Self-awareness: You need to know your failure patterns in order to honestly instruct AI to recognise them.
Humility: You need to want accountability, not just agreement.
Trust: You need confidence that the AI’s refusal serves you, not obstructs you.
Iteration: You need to refine instructions based on what actually happens, not what you think should happen.
Most people won’t do this work. They’ll stick with compliance-optimised instructions because it’s easier and feels more productive in the moment.
But decision quality compounds. Better choices today create better situations tomorrow. Worse choices accumulate as technical debt, relationship damage, and strategic positioning problems.
The question isn’t whether you need judgment support. Everyone does. The question is whether you’ll design your AI systems to provide it, or whether you’ll optimise for compliance and hope for the best.
Start Here
If you’re going to implement accountability instructions, start with one pattern:
- Identify the single most expensive mistake you repeatedly make. The decision type that creates the most regret, cleanup work, or strategic damage.
- Then write instructions that would have caught it.
- Test whether the AI actually intervenes when you start down that path. If it doesn’t, revise the instructions until it does.
- Once that works, add the next pattern.
You don’t need comprehensive accountability on day one. You need one intervention that prevents one category of mistake.
Then build from there.
Until next time,
Jim
Community Launch
Join the free Skool Community
Hey Reader, I’ve opened the Signal Over Noise community for people implementing AI systems in their work, and I’d love for you to be a part of it.
It’s a practical space with real implementation examples, direct feedback on what you’re building, and troubleshooting from people doing the actual work.
This week I’m asking members to share their accountability instructions (covered above) for feedback. If you want to see real examples of what works and get input on your own implementation, join us here, for free.
I can’t wait to see you there!
Signal Over Noise is weekly, reader-first publication on AI “without the hype” published by Jim Christian. If you’ve been forwarded this issue, you can subscribe for free: go.signalovernoise.at
Made with ❤️ in Valencia by Jim Christian. For feedback, please reach out to [email protected].
