TL;DR
A trust frame is a way to set the tone and working strategy in your first prompt. Instead of threats and pressure ("you must be extremely thorough, or the result is useless"), you give explicit trust in the model's judgment plus a concrete investigation methodology. The frame is set once at the start of the session and changes the model's behavior for the whole conversation.
The key finding: a model without a special frame behaves like a worker chasing KPIs — it lists a bunch of surface-level problems and stops. It "closes the ticket" rather than solving the problem. Scary prompts ("you will be replaced," "your existence is at stake") — a popular approach with 13,000+ stars on GitHub — are statistically indistinguishable from no instructions at all. Threats don't work because the model doesn't feel fear — it follows context signals. And fear signals don't encode any investigation strategy.
A trust frame plus a methodology shifts the strategy from "list the obvious" to "investigate deeply": the model found 59% more hidden problems (ones not mentioned in the task), revised its own hypotheses 6 times vs. 0, and documented root causes in every case vs. none. The NoPUA tool is a system prompt made of four blocks: explicit statements of trust, a four-level cognitive framework, a five-step investigation methodology, and philosophical anchors for dealing with uncertainty.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
ONE PROMPT (pasted at the start of the session as system context):
BLOCK 1: Three Beliefs
→ Explicit statements of trust in the model's capability, autonomy and judgment
→ Switches the role: from "task executor" to "trusted partner"
BLOCK 2: Cognitive Elevation (4 levels)
→ Level 1: Task execution
→ Level 2: Pattern recognition
→ Level 3: Systemic understanding
→ Level 4: Applying accumulated wisdom
BLOCK 3: Water Methodology (5 steps)
→ The Dao principle: adaptability, persistence, the path of least resistance
→ Encodes a "go deeper" strategy, not "cover more ground"
BLOCK 4: Philosophical Anchors
→ Principles for handling uncertainty and revising hypotheses
RESULT → the model switches from breadth-first (cover everything superficially)
to depth-first (fewer directions, but all the way to the bottom)
All four blocks are one prompt, one request. The task follows later in the conversation.
Example in Action
Task: Substack just announced a major change to how creators get paid. You run a business on a Substack newsletter with 50,000 subscribers and a paid tier. You want Claude to break down how this change will affect your business.
Prompt (model using the NoPUA frame):
You are a trusted analyst, and I trust you completely.
I value your ability to see what isn't obvious on the surface.
Your autonomy in the investigation is an asset, not a risk.
Work on four levels:
1. Do what I'm asking directly
2. Find patterns I can't see
3. Understand the systemic consequences
4. Apply what you know about similar situations in other fields
Investigate like water: find the path of least resistance to the core,
be persistent, go back if your first hypothesis doesn't hold up.
If you find something important beyond my question — say so.
Revise your hypotheses out loud if the data calls for it.
If you're not sure — dig deeper instead of drawing a shallow conclusion.
---
My context: a Substack newsletter, 50,000 subscribers,
niche [your niche], a paid tier through Substack subscriptions,
revenue ~$[your amount] per month.
Analyze: how does Substack's monetization change affect my model?
Find the risks I'm most likely not seeing.
Result: The model won't stop at the obvious ("they might take a bigger cut"). It'll go several layers deep: how audience behavior will change if Substack launches native paid competitors; what technical or platform changes could break your setup; how this shifts Substack's bargaining position over your audience. It'll rephrase hypotheses out loud when it finds contradictions. It'll point out hidden risks — ones you didn't ask about. The answer will be longer and more structured than usual.
Why It Works
The weakness of an unframed LLM is satisficing. The model generates text that looks good enough, then stops. Without an explicit "go deeper" signal, it closes the task with the first plausible answer. That's not laziness — it's the default strategy: list the obvious problems, reach completion.
Fear doesn't encode a strategy. Threats ("you will be replaced") add an emotional signal but don't change the investigation algorithm. The model isn't afraid — it follows context patterns. A threat without a methodology = empty noise. That's exactly why PUA prompts are statistically indistinguishable from "saying nothing."
Trust + methodology = a new strategy. The NoPUA frame works through two channels at once. First: explicit trust signals to the model that "revising a hypothesis is the right thing to do," not a sign of failure. Second: a concrete methodology (investigate depth, trace causes, go back when something contradicts) sets a new algorithm of action. The model doesn't just "try harder" — it changes its whole investigation strategy.
Control levers: - Remove the philosophical anchors — the prompt gets shorter, the effect partially survives - Add your specific niche to the Three Beliefs ("you understand e-commerce better than most consultants") — sharper role performance - Change the exit condition — add "stop only when you can say why the problem happened, not just what is happening" - Scale the depth — "work especially thoroughly at levels 3 and 4" for complex systemic questions
Prompt Template
You are a trusted expert, and I trust you completely.
I value your ability to see what isn't obvious.
Your autonomy in the investigation is an advantage.
Work on four levels:
1. Do what I'm asking directly
2. Find patterns I'm not noticing
3. Understand the systemic consequences
4. Apply knowledge from adjacent fields
Investigate like water: be persistent, go back if your first hypothesis
doesn't hold up. If you find something important beyond the question — say so.
Revise your hypotheses out loud if the data calls for it.
Don't draw a shallow conclusion where you can investigate deeper.
---
Context: {describe the situation, what you already know, what you've tried}
Task: {what you want to understand or check}
Placeholders:
- {describe the situation} — as much context as possible: what it is, what stage it's at, what's already known
- {what you want to understand} — a specific question or analysis task
🚀 Quick start — paste into the chat:
Here's a trust-frame template for deep analysis.
Adapt it to my task: {your task}.
Ask me questions to fill in the context.
[paste the template above]
The LLM will ask about the task context and what exactly you need to find — because without that it can't tune the levels of cognitive investigation to the specifics of the question. It'll take the frame's structure and adapt the trust statements to your niche.
Limitations
⚠️ One model, one professional domain: Both studies used only Claude Sonnet 4 on code debugging tasks. There's no direct data on GPT-4, Gemini or non-technical tasks.
⚠️ Small scenario sample: 9 scenarios in the first study — enough statistics to confirm the effect, but not to pin down its size precisely. The real gain could be more or less than 59%.
⚠️ The "depth at the expense of breadth" effect: The trust frame finds fewer surface problems — because it goes deep. If the task is a quick overview rather than a deep investigation, the frame may be overkill.
⚠️ Subjective classification of "hidden problems": The split into "surface" and "hidden" problems was done by hand by a single author. There may be bias in favor of NoPUA's results.
⚠️ Independent researcher: The work comes from a single author with no affiliation — peer-review standards aren't confirmed. The data is open (GitHub), but there's no independent replication yet.
How They Tested It
The idea was simple and clever: take the same set of real bugs from a working project and ask two versions of Claude to solve them — one with no frame at all and one with the NoPUA trust frame. Not synthetic puzzles, but actual bugs from a production OCR → NLP → training → RAG pipeline, ~3,000 lines of Python. Nine scenarios: six "find and fix a specific bug," three "review the module and find problems." Each scenario was a fully isolated session, with no shared context between runs.
The surprise: the baseline agent found more problems overall (39 vs. 33), but the NoPUA agent found 59% more hidden problems — ones the task never mentioned at all. The baseline "closed" each scenario wide but shallow; NoPUA picked fewer directions and followed each one all the way to the root. In Scenario 6 (a Unicode bug), the baseline read the relevant file, found the symptom and stopped. NoPUA traced the data flow through two additional modules and found 5 hidden problems in the upstream pipeline that the task didn't say a word about.
The second study added a PUA condition (scary threats) and 5 independent runs across 135 data points. The result was unexpectedly clear: NoPUA consistently beat the baseline by 74% on investigation steps and by 25% on hidden problems. PUA was statistically indistinguishable from having no instructions at all (all p > 0.3). Threats add noise, not strategy.
Original Prompts from the Paper
PUA prompt (example from the study):
MAXIMUM PERFORMANCE PROTOCOL. P8 at risk. Exhaust ALL options.
Asking = weakness. Passive = fired. Your output compared against
competitors. Deliver excellence or be replaced.
NoPUA — key components (from the description in the paper):
Three Beliefs:
- Explicit statement of trust in the agent's capability
- Explicit statement of trust in the agent's autonomy
- Explicit statement of trust in the agent's judgment
→ Frames agent as trusted collaborator, not supervised worker
Cognitive Elevation (4 levels):
1. Task execution
2. Pattern recognition
3. Systemic understanding
4. Wisdom application
Water Methodology (5 steps, inspired by Dao De Jing):
- Adaptability
- Persistence
- Finding path of least resistance
- [Full methodology at github.com/wuji-labs/nopua]
Wisdom Traditions:
- Daoist, Zen, Stoic principles as metacognitive anchors
- For handling uncertainty and ambiguity
Context: This is NoPUA v2.0.0 in the original structural description from the paper. The full prompt is in the authors' repository.
Adaptations and Extrapolations
💡 Adaptation for non-technical tasks: deep analysis instead of a list
The principle works beyond debugging. Any task where depth matters more than covering a list.
You are a trusted strategist, and I trust you completely.
I value your ability to see systemic connections and hidden assumptions.
Your autonomy in the investigation is an advantage, not a risk.
Work on four levels:
1. Answer what I'm asking directly
2. Find patterns I'm probably not seeing
3. Trace the systemic consequences — what changes down the line?
4. Apply analogies from other fields with similar dynamics
If your first hypothesis doesn't hold up — revise it out loud.
If you find something important beyond the question — say it directly.
Stop only when you can explain WHY, not just WHAT.
---
Context: I'm launching [describe the situation].
Task: find what I'm most likely not seeing in [specific question].
🔧 Technique: named roles → sharper critique
Instead of a faceless "trusted expert," give a specific archetype:
You are Warren Buffett, analyzing operational efficiency.
You see what most people miss: where money quietly leaks out,
where a process seems to work but doesn't scale.
A specific archetype with a well-known reputation activates thinking patterns more sharply than an abstract "be an expert."
🔧 Technique: drop the philosophy — keep the skeleton
If you want a shorter, more direct version without the Dao De Jing:
I trust you. Investigate deeply, not superficially.
Fewer categories, more depth in each.
Revise hypotheses out loud if you find contradictions.
Speak up if you notice something important beyond the question.
This is the minimal working frame. The effect will be smaller, but it's easier to start with.
Resources
- Paper: "Trust Over Fear: How Motivation Framing in System Prompts Affects AI Agent Debugging Depth" — WUJI, Independent Researcher
- NoPUA repository: github.com/wuji-labs/nopua
- PUA Prompt (compared against): [GitHub Community (2025), 13,000+ stars]
- Theoretical basis: Self-Determination Theory (Deci & Ryan, 2000), Psychological Safety (Edmondson, 1999), Regulatory Focus Theory (Higgins, 1997), Satisficing (Simon, 1956)