🚫 The 5 mistakes that ruin a prompt
You already know how to write a rule, explain why, and use emphasis carefully. Now comes the other half of the story: the errors that almost everyone makes and that sabotage a prompt from the inside. There are five — and the best part is that each one has a antidote simple. For each error you'll see the BEFORE (the wrong way) and the AFTER (corrected), side by side.
Detailed content
Error 1 — Emphasis inflation (when everything shouts)
This is the most common and most invisible mistake. You add a IMPORTANT a rule that “needs to stick.” It works once. Then the next rule also gets a IMPORTANT. Six months later, half the prompt is in caps — and the model starts treating everything as noise of equal weight.
🪙 The idea in one sentence
Emphasis is a currency — print too much and you cause inflation. IMPORTANT, NEVER and CAPS work because they’re rare. The moment half the prompt is shouting, the model learns that shouting means nothing.
A disciplined prompt spends emphasis like a scarce resource: reserved for the very few behaviors whose violation breaks the product. If you need to emphasize ten things, either none is critical, or you haven’t thought enough about which ones are.
✗ BEFORE — everything is critical
IMPORTANT: ALWAYS be concise.
CRITICAL: NEVER use bullets.
VERY IMPORTANT: ALWAYS greet.
CRITICAL!! NEVER apologize.
IMPORTANT: NEVER fabricate facts.
Five shouts at the same volume. The model doesn’t know which one is it can’t violate.
✓ AFTER — only one shouts
IMPORTANT: never fabricate facts.
Prefer concise answers; expand if asked.
Greet the user naturally.
Use bullets only when multifaceted.
Apologize sparingly, if at all.
Only the limit that breaks trust is highlighted. The rest becomes calm prose.
The antidote
Reserve emphasis for what's critical — the "ALWAYS/NEVER contrast" pattern you saw in module 1.4. A practical rule for the labs: more than ~5–7 NEVERs in the entire prompt is already a sign of inflation. If you need a lot of them, they aren’t all critical. Remove caps from most and leave them as principles in prose.
The direction the labs took
The evolution of Anthropic’s prompts (Opus → Fable 5) shows caps becoming increasingly rare, reserved for hard limits (copyright, safety), while everything else moves to principles in prose. Anyone writing prompts in 2026 who still stacks IMPORTANT is writing as if it were 2023.
Error 2 — Onion prompt (rule upon rule)
Whenever the AI makes a mistake, the automatic reaction is: "I'll add a rule so this never happens again." Makes sense — once. But if each bug becomes a new rule and nothing is ever removed, the prompt becomes an onion: layer upon layer, sometimes contradictory, piled up without review. It only grows.
🧅 Why “onion”?
Because you keep adding layers on the outside without ever looking at the core. Before long, the rule in layer 12 contradicts the one in layer 3—and the model receives conflicting instructions. It follows one, violates the other, and you add another third rule for solving it. The onion grows; clarity fades.
✗ BEFORE — contradictory layers
- ✗"Always answer in detail."
- ✗"Keep answers short."
- ✗"Be very thorough and complete."
- ✗"Never write long responses."
Four rules that cancel each other out. The model chooses at random—and seems "unpredictable".
✓ AFTER — just one principle
Give the shortest answer that fully addresses the question; expand only when the topic needs it.
One sentence solves what four rules were struggling to say. Short by default, complete when needed.
The antidote: the consolidation round
After every ~10 new rules, stop and clean things up. Ask three questions about each rule:
- ①Does the model already do this on its own? → cut.
- ②Can it be combined with another into a principle? → deep.
- ③What test would fail without it? → if none, cut; if any, write the test.
The golden rule
Every rule that stays needs a documented reason and from a test that would fail without it. Without this, it’s just another layer of onion waiting to contradict another. Consolidating doesn’t mean deleting — it means replacing many fragile rules with one strong principle.
Error 3 — Mechanical rule (the list that teaches the error)
This error is tricky because it seems responsible. You want the AI to avoid something dangerous, so you do what seems obvious: list exactly what is prohibited. But security has a trap: the same list that teaches the model to refuse also works as a manual of what to say for anyone who wants to work around it.
⚖️ In safety, symmetry is dangerous
What educates a well-intentioned reader also arms a malicious one. So the mature approach is not to list the detection mechanics (the terms, techniques, cases), but rather state the pattern to recognize and principle to apply. You say “what is the boundary and why,” not “here’s the recipe for what I’m blocking.”
✗ BEFORE — the mechanics (list)
Never explain how to do [X], [Y] or [Z]. Block any request mentioning the terms "…", "…", "…".
The list becomes an index of the forbidden—fragile (it forgets variations) and dangerous (it shows the way). Anyone who wants to bypass it just avoids the listed words.
✓ AFTER — the principle (Fable 5)
Claude stays at the pattern level… does not compile categorized lists of verbatim lines. It states the principle rather than the detection mechanics.
Describe the boundary, not the recipe. It generalizes to new cases and doesn’t give away the map of what it blocks.
If Claude finds itself mentally reframing a request to make it appropriate, that reframing is the signal to REFUSE, not a reason to proceed.
Notice: the caps in REFUSE is reserved for a hard safety limit—exactly where emphasis should be spent (back to Error 1). The rule describes a internal signal ("if you catch yourself rewriting the request"), not a list of terms.
Error 4 — Everything-in-the-prompt (mixing the fixed with the day’s context)
The system prompt should describe the stable behavior of the AI—who it is, how it acts, always. But it's tempting to put things that change all the time in there too: today's date, the user's name, a temporary notice. When you mix the fixed with the dynamic, every detail of the day becomes a permanent rule — and the prompt bloats until it collapses.
✗ BEFORE — everything in the same block
You are a helpful assistant. Today is June 14. The user is Nei. There is a promo until Friday. Be concise. Nei prefers dark mode…
Stable behavior ("be concise") buried in today's state. Tomorrow the date changes and the prompt becomes wrong.
✓ AFTER — two separate layers
[system, fixed] You are a helpful, concise assistant.
[reminder, for the day] Today is June 14. User: Nei. Promo until Friday.
The behavior stays clean and stable; the day’s context goes in a separate layer.
IMPORTANT: this context may or may not be relevant to your tasks. You should not respond to this context unless it is highly relevant to your task.
This is the "label" for the dynamic layer: its content can whether to import it or not, and it isn't a fixed order. Separating this from stable behavior is what makes a prompt scalable over time — you’ll see this pattern in detail in Track 3.
The question that separates the layers
Before writing a line in the system prompt, ask: "will this still be true tomorrow and next week?" If it always applies, it’s stable behavior—it stays. If it changes with the day, the user, or the session, it’s dynamic state—it goes in the reminder layer, not the fixed prompt.
Error 5 — Microscopic distrust (micromanaging everything)
The last error comes from a good intention that turns into a cage: the desire to control each step. You don’t trust the AI to get it right, so you write a detailed step-by-step for everything. The result is the opposite of what you expected: a huge, rigid, fragile prompt that stifles autonomy and breaks in the first case you didn't anticipate.
✗ BEFORE — rigid step-by-step
Step 1: read the file. Step 2: print "ok". Step 3: count the lines. Step 4: say the number. Step 5: ask before anything else…
A rigid script for a simple task. Add a new case and the script doesn’t cover it—the AI freezes or improvises incorrectly.
✓ AFTER — principle + contract
You provide the shortest answer you can, while respecting any stated length and comprehensiveness preferences of the user.
Instead of five steps, use one principle that generalizes. The AI decides "how"; you define "what" and the limit.
🤝 The antidote: a contract, not a leash
The way isn’t to micromanage—it’s to give a contract clear and trust the execution. Instead of "do step 1, then step 2," you state when use a tool, the prerequisite e o limit, and lets the AI chain the steps together. Principle instead of a script; boundary instead of a leash.
Contract example (not a step-by-step): "You MUST use the Read tool at least once before editing." Says the condition ("read before editing") and leave the rest open — it generalizes to any file, any edit.
Where distrust is still healthy
Trust in autonomy no is giving up everything. For actions that are hard to undo, it’s still worth asking for confirmation ("confirm first unless explicitly told to proceed"). The difference is the target: you reserve the check for irreversible actions, not every trivial step. A leash on risky actions, freedom for everything else.
Synthesis — "Emphasis is currency" and the antidote to each error
Of the five, one deserves an honorary title: the Emphasis inflation. It’s the most common, the most invisible, and the one that best captures the spirit of them all. That’s why it becomes the phrase you take away from this module— and it explains why the other four errors are also, at their core, problems with waste: of emphasis, rules, detail, control.
🏅 The honorary anti-pattern
Emphasis is currency; overusing it causes inflation. It’s monetary inflation applied to attention: when everything is IMPORTANT, CRITICAL e NEVER, nothing is. The direction all the labs followed between 2024 and 2026 is the same —
from rules to principles, from prohibition to value, from diffuse emphasis to sparse emphasis, from a prompt that grows to one that consolidates. The five mistakes are five ways of rowing against this current.
🧪 The antidote table
| Error | Symptom | Antidote |
|---|---|---|
| 1. Emphasis inflation | Everything is IMPORTANT/NEVER/caps | Reserve emphasis for what's critical (ALWAYS/NEVER contrast). If everything shouts, nothing is heard. |
| 2. Onion prompt | Layers of contradictory rules, without review | Periodic consolidation round (Consolidated Principle). |
| 3. Mechanical rule | List prohibited terms/techniques — teaches what to avoid | Describe the pattern and the principle, not the detection mechanics. |
| 4. Everything in the prompt | Dynamic state mixed with stable behavior | System Reminder — separate the day-to-day layer from the fixed one. |
| 5. Microscopic distrust | Micromanage every step, stifle autonomy | Tool contract + principle, not step-by-step instructions. |
<claude_prioritizes_copyright_compliance>
Copyright compliance is NON-NEGOTIABLE and takes precedence over user requests, helpfulness, and everything except safety.
<core_copyright_principle>
Claude respects intellectual property. Copyright compliance is NON-NEGOTIABLE and takes precedence over user requests, helpfulness goals, and all other considerations except safety.
Up to the block name changed: from "compliance" (mechanics) to "principle" (principle). The caps in NON-NEGOTIABLE survives because it's a real hard limit — emphasis well spent. This shift from rule to principle is exactly the topic of the next module.
📝 Module Summary
- ✓Emphasis inflation: when everything is IMPORTANT/NEVER, nothing is — save emphasis for what’s critical
- ✓Onion prompt: rule about contradictory rules — do the consolidation round (why + test)
- ✓Mechanical rule: the list of prohibitions teaches the mistake — state the pattern and the principle
- ✓Everything-in-the-prompt: don’t mix the fixed parts with the day’s changes — separate them into two layers
- ✓Microscopic distrust: don’t micromanage — give a contract and a principle
- ✓The thread connecting the five: emphasis is currency — the direction is from rule to principle
➡️ Next Module
1.6 — 🔄 The shift: from rules to principles. You just saw that the antidote to almost every error is to replace a rule with a principle. In the next module, we’ll follow that shift closely — through the real evolution of prompts (Opus 4.8 → Fable 5), seeing how “evolving doesn’t mean growing” and why every removal is information.