What's Inside
- Prompt 1: make Claude or ChatGPT audit your whole setup and name every instruction that is restricting its reasoning
- Prompt 2: delete what is not needed and rewrite your skills, rules files and saved instructions for a model that reasons on its own
- Prompt 3: the Perfect Prompt Pyramid, the four layers every prompt needs from here on
- What to keep, so you do not delete the parts that are actually doing work
The newest models are not broken. They reason for themselves, which means the instructions you wrote to make the older models behave are now overriding better judgment. Anthropic says this in their own prompting guidance: dial back guidance written for previous models, and remove carried-over instructions rather than rewriting them. This guide is three prompts. Run them in order. They work in Claude and in ChatGPT.
Step 1. Audit
You are not looking for bad writing. You are looking for instructions that tell the model HOW to do the work instead of WHAT the work is. Those are the handcuffs. Paste this into the tool you actually use, with your setup available to it.
Go through my entire setup: every rules file, workflow, context doc and stored prompt.
For each one, tell me:
1. Which instructions tell you HOW to do the work rather than WHAT the work is
2. Which instructions were written for an older model and now slow you down or restrict your reasoning
3. Which instructions contradict each other
4. Which instructions are doing real work and should stay
Rank everything by how much it is costing the quality of your output. Do not change anything yet.Step 2. Cut and retool
Now you delete. This is the step people skip, because deleting work you spent a weekend on feels like losing. It is not. Every rule you remove gives the model back a decision it can make better than your instruction could.
Using the audit above, do two things.
First, delete everything that is not needed. Show me what you removed and one line on why.
Second, rewrite what is left, including my skills and rules files, so they work with a model that reasons on its own. That means: state the outcome, state the hard boundaries, state what good looks like, and stop describing the method.
Show me the before and after of each file side by side so I can approve it.Step 3. The Perfect Prompt Pyramid
Trimming the setup fixes the past. This fixes everything you write from here. Four layers, bottom to top. Never a fifth.
GOAL. The finished outcome, not the next task. What does done look like.
RESEARCH. Tell it to find the best approach first instead of taking yours.
INTERVIEW. Let it ask you what it needs. You do not have to think of everything up front.
GOOD LOOKS LIKE. One real example to match. This does more than a page of adjectives.
GOAL: [the finished outcome]
RESEARCH: Before you start, find the best approach for this and tell me what you picked and why.
INTERVIEW: Ask me whatever you need to know. One question at a time.
GOOD LOOKS LIKE: [paste one real example of the finished thing]What to keep
Do not delete everything. Keep the facts the model cannot guess: your customers, your pricing, your voice, your constraints, and anything that cannot be undone, like what it is allowed to send or spend. Cut the method. Keep the context and the boundaries.