Troubleshooting & Iteration

🔄 You’ve built templates (Lesson 6) and installed them (Lesson 5). But here’s the reality: custom instructions don’t work perfectly on the first try. They need debugging, just like code. Let’s learn how.

The Five Most Common Problems

Problem 1: Instructions Are Too Vague

Symptom: The AI’s behavior doesn’t noticeably change after installing instructions.

Diagnosis: Read your instructions and ask: “Would a human colleague know exactly what to do with these?” If the answer is no, the AI doesn’t either.

Examples of vague vs. specific:

The fix: Replace every adjective with a measurable behavior. If you can’t measure it, the AI can’t reliably follow it.

✅ Quick Check: Your instruction says “write clearly.” The AI produces technically correct but dense academic prose. What went wrong? (“Write clearly” means different things to different audiences. Replace with specifics: “Use short sentences (under 20 words). One idea per paragraph. Grade 8 reading level.”)

Problem 2: Conflicting Instructions

Symptom: The AI follows some instructions but ignores others, seemingly at random.

Diagnosis: Look for contradictions in your instruction set:

• “Be concise” + “Be thorough” → which takes priority?
• “Always use bullet points” + your request for a cover letter → format conflict
• “Never assume” + “Anticipate what I need” → behavioral conflict

The fix: Use priority stacking from Lesson 4. Make it explicit which rules override which:

“Default to concise (under 200 words). When I ask for detailed analysis, switch to thorough — ignore the length limit for those requests.”

Problem 3: Instruction Drift in Long Conversations

Symptom: The AI follows instructions well for the first 5-10 messages, then gradually stops.

Cause: As the conversation grows, your instructions (at the very top of the context) get farther from the AI’s immediate attention. AI models have a recency bias — they attend more to recent messages.

The fix:

• Reinforce periodically: Every 10-15 messages, restate key rules: “Remember: keep it concise, bullet points.”
• Start new conversations for new topics rather than using one endless thread
• Use Claude Projects — their instructions stay prominently attached throughout

Problem 4: Over-Constrained Instructions

Symptom: The AI produces robotic, formulaic responses. Every response looks identical regardless of the question.

Diagnosis: You’ve written too many rigid rules with no flexibility.

The fix: Add breathing room:

• Change “always” to “by default”
• Add explicit exceptions for creative or unusual tasks
• Remove rules that don’t meaningfully change behavior (“be helpful” — the AI already tries)

Problem 5: Wrong Level of Specificity

Symptom: Instructions work great for one type of task but terrible for others.

Cause: Your instructions are optimized for a single use case but you’re asking the AI to do multiple types of work.

The fix: Either use conditional logic (Lesson 4) to handle multiple task types, or create separate instruction sets for different workflows.

The Troubleshooting Protocol

When instructions aren’t working, follow this systematic process:

Step 1: Isolate — Test each instruction individually. Which specific rules are being followed? Which aren’t?

Step 2: Check for conflicts — Read all instructions together. Do any rules contradict each other?

Step 3: Measure — Replace subjective instructions (“be concise”) with measurable ones (“under 100 words”).

Step 4: Simplify — Remove instructions that don’t change behavior. Less is often more.

Step 5: Test — Run three different prompts and verify the AI follows instructions across all of them. One test isn’t enough.

✅ Quick Check: Your instruction says “Be critical when reviewing my work.” But the AI still starts every review with “Great work!” What’s the fix? (Be more specific: “When reviewing my work, skip all positive feedback. Lead with the problems. List issues in order: critical → important → minor. Do not use phrases like ‘great work’ or ’nice job.’”)

The Iteration Cycle

Custom instructions aren’t write-once. Here’s a healthy iteration schedule:

Day 1: Install your initial instructions using RISEN Day 3: Note what’s working and what’s not Week 1: Review and adjust — fix specific problems Month 1: Major revision — you’ll have a clear picture of what you actually need Quarterly: Check if your role, tools, or workflow has changed

Each iteration makes your instructions tighter and more effective. Think of it like code: the first version works, but the tenth version is clean.

Security: What Not to Put in Instructions

A quick note on security. Custom instructions can be extracted by other users in some contexts (shared Custom GPTs, API system prompts). Don’t include:

• Passwords, API keys, or tokens
• Confidential company information
• Personal data (home address, SSN)
• Trade secrets

OWASP ranks prompt injection as the #1 AI vulnerability — 73% of production AI deployments are affected. Keep sensitive data out of instructions entirely.

Key Takeaways

• Most instruction failures come from vagueness — replace subjective descriptions with measurable behaviors
• Conflicting instructions cause inconsistent behavior — use priority stacking to resolve conflicts
• Long conversations cause drift — reinforce key rules periodically or start fresh threads
• Over-constrained instructions produce robotic output — add flexibility with “by default” and exceptions
• Follow the 5-step troubleshooting protocol: Isolate → Check conflicts → Measure → Simplify → Test

Up Next

Final lesson. In the Capstone, you’ll build your personal instruction library — a collection of tested, refined instruction sets organized by workflow. Plus a maintenance plan to keep them current as AI platforms evolve.