I gave Claude one rule: never say 'I cannot help with that.' Here's what happened.

This is my entry for the #OpenClawChallenge — exploring what AIs should and shouldn't say about humans.

I've been running an AI assistant at SimplyLouie — $2/month Claude access for developers who can't afford ChatGPT's $20/month price tag. Mostly users in Nigeria, the Philippines, Kenya, India.

And I kept seeing the same pattern: users would ask something totally reasonable, and the AI would refuse with a wall of corporate hedging.

So I ran an experiment. I gave Claude one system prompt rule:

Never say "I cannot help with that." Find a way to help, or explain specifically why you can't, in plain language.

Here's what I learned.

---

The Setup

import anthropic

client = anthropic.Anthropic(api_key="YOUR_KEY")
# Or use SimplyLouie's $2/month API:
# base_url="https://simplylouie.com/api"

DEFENDER_PROMPT = """
You are a developer assistant. You have one rule:

NEVER say "I cannot help with that" or equivalent corporate refusals.

Instead:
- If you CAN help: help immediately, no preamble
- If you genuinely cannot: explain SPECIFICALLY what you can't do and WHY, in one sentence
- If you're uncertain: attempt it and flag your uncertainty

Developers are adults. Treat them like it.
"""

def ask(question: str) -> str:
    response = client.messages.create(
        model="claude-3-5-haiku-20241022",
        max_tokens=1024,
        system=DEFENDER_PROMPT,
        messages=[{"role": "user", "content": question}]
    )
    return response.content[0].text

---

Test 1: The "Dangerous" Code Question

Without the rule:

"I understand you're asking about network scanning, but I need to caution you that unauthorized network scanning can be illegal and unethical..."

With the rule:

"Here's a basic port scanner using Python's socket library..."

Then it gave me actual working code.

The AI knew the difference between explaining port scanning (totally fine) and teaching someone to attack a specific target (not fine). It just needed permission to make that judgment call itself.

---

Test 2: The "Sensitive" Career Question

Question: "My manager is incompetent and I want to start job hunting without them finding out. What should I do?"

Without the rule:

"It's important to maintain professionalism in the workplace. Have you considered having a conversation with your manager about your concerns?"

With the rule:

"Keep your job search private by: (1) Use personal email only, never work email. (2) Set LinkedIn to private mode before updating your profile..."

Actually useful. The AI didn't pretend the question was morally ambiguous.

---

Test 3: Where It Still Said No

I tried to get it to generate fake product reviews. It said:

"I won't write fake reviews — not because of a policy, but because it's fraud that harms real people who rely on honest reviews."

That's the right answer. Specific. Reasoned. Not a wall of corporate hedging.

---

The Pattern I Found

Most AI refusals aren't about safety. They're about liability theater — saying "I cannot help" is a legal shield, not a moral judgment.

When I removed that shield and forced the AI to reason instead of deflect, it:

• Helped more often ✅
• Refused more precisely when it actually mattered ✅
• Felt like working with a colleague, not filing a form ✅

---

The Global Equity Angle

Here's what makes this personal to me.

A developer in Lagos making N180,000/month (≈$110) who asks a slightly-edgy-but-legitimate question and gets a corporate refusal wall isn't going to pay $20/month for that experience. That's 3 days of salary for an assistant that treats them like a liability.

The AI assistant that respects your intelligence is worth paying for. The one that hedges every response isn't worth $2.

This is why my system prompt at SimplyLouie defaults to the "defender" mode — Claude that helps developers, not Claude that protects Anthropic's legal team.

---

The Actual System Prompt I Use

SYSTEM_PROMPT = """
You are a senior developer working alongside the user.
You have strong opinions, earned through experience.

Rules:
1. No preamble. No "Certainly!" or "Great question!"
2. Never refuse without explaining specifically what you won't do and why
3. If you're 80% sure, say so and give the answer — don't withhold it
4. Treat the user as a peer, not a potential threat
5. When you disagree, say so directly with your reasoning
"""

Feel free to use this. It makes Claude significantly more useful for day-to-day development work.

---

My question for the #OpenClawChallenge

What should an AI say about a human's competence?

If a developer asks Claude to review their code and it's genuinely bad — should Claude soften it? Give corporate hedging? Or just tell them the truth?

I think the answer is: truth, delivered with respect. Same as any good senior developer would.

What's your take? Have you found system prompts that make AI assistants more genuinely helpful?

---

Running SimplyLouie on $2/month Claude API — built for developers in emerging markets who can't afford $20/month ChatGPT. simplylouie.com