Think²: Structured Self-Correction Through Three Phases of Thinking in a Single Prompt

Everything runs in a single prompt, all three phases in one request:

PHASE 1: PLANNING
→ Classify the task type
→ List the known facts, formulas, constraints
→ Predict the shape of the answer (type, range, structure)

PHASE 2: MONITORING
→ Solve step by step, explicitly tracking state
→ At every transition: "Is this step's logic sound? Any contradictions?"
→ When in doubt — stop and trace the source

PHASE 3: EVALUATION
→ Check the answer against the predictions from Phase 1
→ Check that every constraint from the task is satisfied
→ Lock in the final conclusion

⚠️ Where the method shines: multi-step tasks with clear constraints where there's a risk of a confident mistake — diagnostics, analysis, logical decisions, risk assessment. Where it's weak: simple questions, creative tasks without hard criteria, tasks for weak models.

Task: You got a startup offer at 20% below your current salary, but with a 0.5% stake and options. The company is at Series A, valued at $80 million. You need to decide: accept or not.

Prompt:

Help me make a decision about a job offer. Follow three phases strictly:

TASK: I've been offered a job at a startup. Details:
- Current salary: $160,000/year
- Offer: $128,000/year (-20%)
- Equity stake: 0.5%
- Stage: Series A
- Current company valuation: $80 million
- Cliff: 1 year, vesting: 4 years
- Sector: EdTech, B2B

---

PHASE 1 — PLANNING:
Before analyzing — write down:
1. Which factors are critical for this decision
2. What data we have and what's missing
3. What format the final conclusion should take
4. Predict: which pitfalls people usually miss in a decision like this

PHASE 2 — MONITORING:
Analyze step by step. After each step write explicitly:
"✓ This step is complete" or "⚠️ Caution needed here: [reason]"
Steps:
- Financial losses over the vesting period
- Equity value scenarios (×3, ×10, failure)
- Risks that are hard to quantify
- Opportunity costs

PHASE 3 — EVALUATION:
Check: does your conclusion account for ALL the critical factors from Phase 1?
Is there any contradiction between the data and the recommendation?
Give a final conclusion: [accept / decline / negotiate terms] + the main argument.

Result: The model will work through three explicitly labeled blocks. In Phase 1 it will list the key variables on its own — exit probability, equity liquidity, dilution terms. In Phase 2 you'll see each step clearly marked, with warnings. In Phase 3 the model will reconcile the final conclusion with what it itself called critical — and if there's a contradiction, it will say so.

The LLM's weakness: the model generates text linearly — token by token. It doesn't "go back" to check whether what it wrote at the end matches what it promised at the start. Chain-of-Thought provides steps but not checkpoints — moments where the model must stop and explicitly compare the current state with the target.

The LLM's strength: the model follows explicit instructions inside the prompt well. If it says "check the logic after each step," it will check. The structure of the answer format becomes the structure of the thinking process.

How the method uses this: the three phases create forced stops. Planning pins down the expected shape of the answer before solving even starts. Monitoring turns every transition into a checkpoint. Evaluation closes the loop: the model can't just finish — it has to reconcile the ending with the beginning. That's what psychologists call the regulatory cycle.

Control levers: - Phase 1 detail → more items in the constraints list = deeper planning, but a longer answer - Monitoring markers (✓ / ⚠️) → remove them if you want clean output without bookkeeping labels - The Phase 3 condition → replace "check for contradictions" with "give a confidence percentage for each point" — and you'll get a calibrated assessment - Detail of the Phase 1 prediction → ask it to predict the shape of the error, not just the shape of the answer — the model gets sharper at catching its own weak spots

Solve the task strictly in three phases:

TASK: {task description with concrete data and constraints}

---

PHASE 1 — PLANNING:
Before you start solving, write down:
1. The task type and key variables
2. What's already known, what's missing
3. A prediction: what shape/type the final answer should have
4. Where people usually go wrong on tasks like this

PHASE 2 — MONITORING:
Solve step by step. After each step write explicitly:
"✓ Checked: [what you checked]" or "⚠️ Risk: [what raises doubt]"
Steps: {list 3-5 logical stages for your task}

PHASE 3 — EVALUATION:
- Does the answer match the predicted shape from Phase 1?
- Are all the task's constraints satisfied?
- Are there any internal contradictions?
Final conclusion: {the format of the answer you need}

Placeholders: - {task description} — a concrete situation with numbers, conditions, context. The more precise, the better the diagnosis - {list the steps} — 3-5 logical stages for solving your task. You can simply write "figure them out yourself based on the task" - {answer format} — for example: "recommendation + main argument" or "list of risks in descending order of importance"

🚀 Quick start — paste into the chat:

Here's a template for the Think² method (three-phase thinking).
Adapt it to my task: {your task}.
Ask clarifying questions if you need data to fill in the fields.

[paste the template above]

The LLM will ask which steps are needed in Phase 2 and what format the final conclusion should take — because those are the elements that depend on the specific task and determine the quality of the diagnosis.

⚠️ Model dependence: On weak models without a built-in reasoning mode, the method can worsen results — especially on math and code. The structure adds overhead that an underpowered model can't handle. For ChatGPT-4, Claude Sonnet and above it works well.

⚠️ Not for simple tasks: If the task takes a single action, three phases are overkill. The method pays off only where there's a risk of a confident mistake: complex logic, analysis with conflicting data, multi-step decisions.

⚠️ Don't expect miracles on math: On pure computation (high-school algebra, simple arithmetic) the gain is negligible or negative. The main strength is diagnosing logical errors and raising perceived trustworthiness, not number crunching.

⚠️ MetaController is just a concept: The idea of automatically routing tasks between a "fast" and a "slow" mode is elegant, but in the study it showed serious failures — the model mistook hard tasks for easy ones based on surface features. The manual fix: apply the phases yourself when you sense a task is non-trivial.

The IIIT Hyderabad team ran a clean experiment: they took two fundamentally different kinds of models — Llama-3-8B (a standard instruction-tuned model) and Qwen-3-8B (with a built-in reasoning mode) — and ran them through six benchmarks, each deliberately targeting a different phase: GSM8K for planning, CRUXEval for execution monitoring, CorrectBench for error diagnosis, TruthfulQA for resisting hallucinations.

The most interesting result was unexpected: structure only helps a model that already knows how to think. Qwen-3, which has built-in thinking tokens, absorbed the three phases as natural scaffolding and posted the best results almost everywhere. Llama-3, which wasn't trained on reasoning chains, started drowning in the structure — on GSM8K it dropped from 80% to 69%. That's a key insight: a prompt's meta-structure doesn't replace a model's ability to reason, it amplifies an ability that already exists.

Separately, they ran a live human evaluation: 580 answer pairs, college-educated annotators, a fully blind design (no one knew who wrote what). They rated not the correctness of the answer but trust in the reasoning. 84% preferred the Ann Brown approach — because the model explicitly acknowledges uncertainty where it exists instead of making confident mistakes. CoT often delivers a confident answer even when it's wrong — the three phases forcibly align the "confidence of the voice" with actual correctness.

The authors moved the detailed prompts for each benchmark into Appendix B, which isn't available in the public version of the paper. Below is the exact description of the phase mechanics from the main text:

Phase 1 (Planning):

"The model first classifies the problem structure and identifies stable knowledge (e.g., formulas, constraints, rules) before generating solution steps. This separates constraint retrieval from execution, forcing strategic roadmap construction and prediction of the answer's structural form (e.g., value range or type), thereby reducing premature token-level commitment."

Phase 2 (Monitoring):

"The model explicitly tracks intermediate reasoning states (e.g., variable changes or logical transitions) and performs debugging checks when ambiguity or complexity arises. This transforms generation into controlled processing, enabling real-time detection of inconsistencies and revision of heuristic decisions."

Phase 3 (Evaluation):

"The model checks whether the solution satisfies predictions and structural commitments established during Planning. Closing the regulatory loop enables detection of internal contradictions and hallucinated premises beyond surface-level correctness."

Context: These are the exact descriptions of the mechanics from Section 4 (Mind Over Matter Framework) — the foundation the entire method is built on.

💡 Adaptation: Phase 1 as a "Where Will the Model Go Wrong" Detector

Instead of standard planning, ask the model in Phase 1 to explicitly predict its own weak spots:

PHASE 1 — PLANNING:
Before you begin:
1. What data do you need that isn't in the task?
2. At what point in the reasoning are you most likely to make a mistake?
3. Which answer would be "too obvious," and why might it be a trap?

The model flags the risks itself before it runs into them. Especially effective for tasks with "traps" — non-obvious conditions in contracts, case studies, technical specs.

🔧 Technique: Phase 3 with Numerical Confidence → a Calibrated Assessment

Replace the standard Evaluation with:

PHASE 3 — EVALUATION:
For each key claim in the answer, assign a confidence level (0-100%):
- [Claim 1]: __% — because [justification]
- [Claim 2]: __% — because [justification]
If confidence is < 70% — flag it as "needs verification."

Useful for legal, medical and financial questions — where the cost of a mistake is high and you need to know exactly where not to trust the answer.

🔧 Technique: Two-Pass — Solve First, Then Clean Phases

Some tasks are better solved in two passes:

First, solve the task as usual.

Now apply the three phases RETROSPECTIVELY to your solution:
PHASE 1: What should you have known before solving?
PHASE 2: Where in your reasoning were there transitions without an explicit check?
PHASE 3: Does your answer satisfy all the conditions of the task?

This helps uncover missed conditions after the first answer — less structural overhead, more speed.

Paper: Think²: Grounded Metacognitive Reasoning in Large Language Models

Authors: Abraham Paul Elenjical, Vivek Hruday Kavuri, Vasudeva Varma — IIIT Hyderabad

Theoretical foundation: Ann Brown's Theory of Metacognitive Regulation (Brown, 1987); Kahneman's Dual-Process Theory (Thinking Fast and Slow, 2011)

Baselines: Chain-of-Thought (Wei et al., 2022); Metacognitive Prompting (Wang & Zhao, 2024)

Benchmarks: GSM8K, CRUXEval, MBPP, AIME, CorrectBench, TruthfulQA