COT
It only started working above about 100 billion parameters — and on a reasoning model it can now get in the way.
What is COT?
Chain of thought is one instruction: make the intermediate steps explicit before giving an answer. It is the most-cited prompting technique there is, it has a proper research paper behind it, and it is the one whose usefulness has changed most since that paper was written.
A model asked for an answer produces the answer in one pass. A model asked to work through the steps produces the steps and then an answer conditioned on them — and on multi-step arithmetic, logic and analysis that difference is large. But the reason to reach for it in ordinary work is not accuracy alone. A chain you can read is a chain you can check. When a model gets a pricing calculation wrong, an answer tells you nothing and a numbered chain tells you it applied last year's churn rate in step four. That auditability is worth more than the accuracy gain to most people using it for work.
Where COT Came From
Wei et al., 2022 — and the finding was about scale
Chain-of-thought prompting was set out in "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" (Wei et al., 2022, Google). The headline result was not simply that it works — it was that it is an emergent ability. Standard prompting showed a flat scaling curve on reasoning benchmarks; chain-of-thought showed a sharply rising one, but only past roughly 1023 training FLOPs, around the 100-billion-parameter mark. Below that threshold it did not help and could hurt.
"Let's think step by step" is a different paper
The phrase everybody quotes comes from Kojima et al., 2022, "Large Language Models are Zero-Shot Reasoners". Wei's version demonstrated reasoning with worked examples — few-shot chain of thought. Kojima's contribution was that you could skip the examples entirely and just append the instruction, which is zero-shot chain of thought and the version almost everyone actually uses.
What changed since: models that already do this
The important 2026 caveat, and most pages have not caught up. Reasoning models run an internal chain before answering, so telling one to think step by step is at best redundant and at worst counterproductive — you get the reasoning twice, once hidden and once performed for you, and the performed one is a summary rather than the real thing. The instruction is still valuable on ordinary models, and for auditability on any model, but "always add step by step" stopped being good advice.
The 5 Slots, One at a Time
Each slot is a decision. Leave it out and the model still makes it — just without you.
Chain of thought does not manufacture inputs. A chain built on invented figures is a beautifully structured wrong answer, and the structure makes it more persuasive rather than less. Supply the numbers and say they are the only ones available.
Two words matter: before, and show. Reasoning that happens after the conclusion is justification, not calculation. On a reasoning model, ask for the working to be shown rather than for thinking to happen, since the thinking is happening anyway.
Free-form chains wander. Naming the steps — current position, revenue lost, subscribers gained, cost effect, net — turns a chain into a model you can compare across runs. If you do not know the right sequence, leave it open; if you do, withholding it is not rigour.
The highest-value addition and the one nobody makes. A model reasoning about incomplete data must assume things; the question is whether it tells you. Requiring each assumption to be labelled, plus a sentence on which is load-bearing, converts a chain into something you can argue with.
Cheap and rarely done. Ask for the same result derived from a different direction — from contribution per unit rather than from totals — and require the two to be reconciled. A chain that agrees with itself two ways is a materially different claim from one that agrees with itself once.
One Task, Before and After
The task: Whether cutting a subscription price from £29 to £19 would be profitable. Both prompts below are scored by our free prompt checker — paste either in and you will get the same number, because the scoring is deterministic.
Work out if cutting our price would make us money.
Nine words. You get a confident yes or no, built on a churn rate, a conversion figure and a serving cost the model chose for you, with no way to see which of them was wrong.
Work out whether cutting our Standard plan from 29 to 19 pounds a month would be profitable, and show the working at every step. The figures, and the only figures available: 2,400 Standard subscribers at 29 pounds. Monthly churn on that plan is 3.1%. Serving cost is 4.60 pounds per subscriber per month. Last time we ran a price test, a 20% cut moved trial-to-paid conversion from 4.2% to 5.6%; we get about 9,000 trials a month across all plans, roughly 60% of which land on Standard. Support cost rises about 0.40 pounds per subscriber for every 500 net new subscribers. Reason step by step before giving any conclusion, and set out the working as numbered steps: current monthly revenue and gross margin, the revenue lost on existing subscribers, the new subscribers the conversion lift would add, their contribution, the effect on serving and support cost, and only then the net monthly position. State every assumption you have to make on its own line, marked ASSUMPTION, and say which one the answer is most sensitive to. Where a figure above does not cover something you need - what happens to churn at the lower price, for instance - do not invent a value; state the gap and carry it through as a named variable. Then check yourself: recompute the net position a second way, working from contribution per subscriber rather than from total revenue, and say whether the two agree. If they do not, find the discrepancy rather than picking one. Write 600 to 800 words for a founder who is comfortable with numbers but has not seen these figures laid out. Use numbered steps with one short paragraph each, and end with a two-line answer. For example, each step should look like this: ``` Step 3 - New subscribers from the conversion lift. 9,000 trials x 60% Standard = 5,400. Lift 4.2% -> 5.6% = +1.4pp = +76 subscribers a month. ASSUMPTION: the 2019 test's lift transfers to a 34% cut rather than the 20% cut it measured. ``` Every step must name which input figures it used, so the chain can be checked line by line. Ensure the two computations in the self-check are genuinely independent rather than the same arithmetic rearranged. Do not round intermediate results, do not present a single number as the answer without the range your assumptions imply, and do not introduce any figure that is not above or explicitly marked as an assumption.
Ninety-two. Two additions do most of the work and neither is chain-of-thought as usually described: the assumption ledger, and the requirement to derive the answer a second way and reconcile.
Ninety-two, and no score can tell a correct chain from a confident one
One check is structurally unreachable, and the risk specific to this technique is the one nothing measures:
No persona slot — chain of thought is an instruction, not a brief. A role is available for the asking and worth eight points; on our scorer a generic one scores the same as a specific one, so the points say nothing about quality.
The whole point, and completely invisible. A fluent, well-numbered chain with an arithmetic error in step four scores identically to a correct one — and reads more convincingly than a bare wrong answer would. Structure is not accuracy.
On a reasoning model the instruction may be redundant, and asking for a performed chain gets you a summary of thinking rather than the thinking. No rubric knows which model you are pointing at.
Which is why the two instructions worth keeping are the ones that make a chain checkable rather than longer: label every assumption, and derive the answer a second way. Neither is in the original technique, neither earns a point, and together they turn a persuasive chain into one you can falsify.
Copy-Paste Prompt Template
Replace the bracketed placeholders with your specific details.
[The question, and EVERY figure needed — stated as the only figures available] [Reason step by step BEFORE giving any conclusion, and show the working as numbered steps] [If you know the sequence, name it: what to compute first, second, third] [Every assumption on its own line marked ASSUMPTION, plus which one the answer is most sensitive to] [Missing figures are carried as named variables, never invented] [Then recompute the answer a SECOND way and reconcile the two] [Length, reader, and one fenced sample step showing the density wanted]
When COT Fits — and When It Does Not
- Multi-step arithmetic — pricing, unit economics, break-even, staffing maths.
- Diagnosis, where the interesting output is which step the reasoning turned on.
- Comparisons with several moving parts, like offers or build-versus-buy.
- Anything you will need to check, defend or repeat later.
- Smaller and cheaper models, where the accuracy gain is largest.
- Reasoning models that already think internally — you get a performance of thought, not thought.
- Simple lookups and one-step tasks, where it adds tokens and latency for nothing.
- Creative writing. Visible deliberation flattens prose.
- Anything whose output is parsed. Use ICIO; a chain and clean JSON are contradictory asks.
- Problems where you have no real figures — a chain over invented inputs is worse than a guess.
10 Ready-Made COT Prompts
Every prompt below was produced by the Frompting generator with COT selected — not written by hand for this page. Each is scored by our prompt checker; the median is 90/100. Click one to open it, then copy.
Whether a price cut is actually profitable 96
You are a financial analyst specializing in pricing strategy. Your task is to determine whether a proposed change to a product’s selling price will increase overall profitability. **What you need to know (please provide or confirm):** - [CURRENT_PRICE: the existing unit selling price] - [PROPOSED_PRICE: the new unit selling price being considered] - [UNIT_COST: total cost to produce or acquire one unit] - [VOLUME_RESPONSE: expected change in units sold at the new price (e.g., percentage increase or decrease)] - [ANALYSIS_PERIOD: time frame over which profitability should be evaluated (e.g., one quarter, one year)] **Steps to follow:** 1. Calculate the baseline profit per unit and total profit for the current price using the provided unit cost and assumed sales volume. 2. Estimate the new sales volume by applying the expected volume response to the current sales volume. 3. Compute the profit per unit and total profit at the proposed price using the same unit cost. 4. Compare total profits before and after the price change, highlighting the net profit difference. 5. Identify any assumptions you made and list up to three key factors that could most affect the result (e.g., price elasticity, fixed costs, market competition). **Output:** - A concise report of 660-870 words, presented in plain prose with a brief bullet summary of the profit calculations. - Include a clear statement of the profit impact (positive, negative, or neutral). - Provide the three identified risk factors. **Quality criteria:** - Calculations must be correct and transparent. - Reasoning should be logical and easy to follow. - Conclusions must be directly supported by the numbers you computed. **Boundary:** Do not speculate on market trends beyond the provided volume response or suggest marketing actions; focus solely on the profitability calculation. If any of the required inputs are unknown, state your assumption explicitly and ask up to three clarifying questions before proceeding. Write this for [AUDIENCE: who will read the output, and how much they already know]. Match the depth, vocabulary and examples to that reader. Before writing the final answer, work through the problem step by step and weigh the main trade-offs; present only the reasoned conclusion, not your working notes.
Why a finance report will not reconcile 87
You are a meticulous financial analyst tasked with diagnosing discrepancies in a finance report’s totals. The report in question involves [REPORT TYPE: e.g., monthly profit-and-loss, balance sheet, cash flow] for the period [TIME PERIOD: e.g., Jan 2024, Q2 2023]. The data originates from [DATA SOURCE: e.g., ERP system, spreadsheet, accounting software]. Your job is to identify the root causes of any reconciliation issues by systematically examining each component of the report. Follow a logical, step-by-step reasoning process: 1. List all line-items and their aggregation rules. 2. Verify that each line-item’s source data matches the reported figures. 3. Check for common error sources such as rounding, timing differences, duplicate entries, omitted transactions, and formula mistakes. 4. Cross-reference totals with supporting schedules or external statements. 5. Summarize the most likely reasons for the mismatch and suggest concrete corrective actions. If any of the above details are unclear, state your assumptions explicitly and ask up to three clarifying questions before proceeding. Produce a concise report of 660-870 words, using clear headings for each step and a bullet list for identified issues and recommendations. Ensure the analysis is transparent, evidence-based, and avoids speculation beyond the provided data. Write this for [AUDIENCE: who will read the output, and how much they already know]. Match the depth, vocabulary and examples to that reader.
Comparing three job offers financially 96
You are a financial decision analyst. Your task is to compare three distinct job offers and determine which one provides the highest net financial benefit to the candidate. **Context** - Offer A: [SALARY_A: annual base salary] - Offer B: [SALARY_B: annual base salary] - Offer C: [SALARY_C: annual base salary] - For each offer, indicate any additional compensation such as signing bonuses, annual performance bonuses, stock options, or relocation assistance: [BONUSES_A], [BONUSES_B], [BONUSES_C]. - Specify the expected tax rate or filing status for the candidate: [TAX_RATE]. - Provide any known differences in benefits that have monetary value (e.g., health insurance premium coverage, retirement match, paid time off value): [BENEFITS_A], [BENEFITS_B], [BENEFITS_C]. - If the offers are in different geographic locations, include the cost‑of‑living adjustment factor for each location: [COL_FACTOR_A], [COL_FACTOR_B], [COL_FACTOR_C]. **Task Steps** 1. Convert all compensation components for each offer into an annual net amount after tax and after applying the cost‑of‑living factor. 2. Summarize the net annual figure for each offer in a clear table. 3. Identify the offer with the highest net annual amount and explain why it is financially superior, referencing the calculations. **Output** - A markdown table with columns: Offer, Gross Annual Compensation, Net Annual Compensation (after tax and COL adjustment). - A concise paragraph (≈80–120 words) stating the financially best offer and the key reasons. **Quality Criteria** - Calculations must be transparent and reproducible. - All monetary values should be presented in the same currency. - Reasoning should be logical, step‑by‑step, and free of unsupported assumptions. **Boundary** Exclude non‑financial factors such as career growth, company culture, or personal preferences. If any required data is missing, state your assumptions explicitly and ask up to three clarifying questions before completing the analysis. Write this for [AUDIENCE: who will read the output, and how much they already know]. Match the depth, vocabulary and examples to that reader. Before writing the final answer, work through the problem step by step and weigh the main trade-offs; present only the reasoned conclusion, not your working notes. Use this table shape, one row per item, filling values from your analysis: | Offer | Gross Annual Compensation | Net Annual Compensation | | --- | --- | --- |
Subscription versus pay-as-you-go 90
You are a financial analyst tasked with comparing the total cost of a subscription model versus a pay-as-you-go model for a specific user scenario. Your analysis should be presented as a clear, step-by-step reasoning process that leads to a definitive conclusion about which option is cheaper. **Required inputs (please provide):** - [SUBSCRIPTION_COST]: the fixed periodic fee (e.g., monthly or annual) of the subscription. - [PAYG_UNIT_COST]: the cost per individual unit of usage under the pay-as-you-go model. - [EXPECTED_USAGE]: the projected number of units the user expects to consume in the same period as the subscription fee. - [PERIOD]: the time frame for comparison (e.g., one month, one year). **Output:** a concise report of 395-525 words that includes: 1. A brief summary of the scenario. 2. A sequential calculation showing the total cost for each model using the provided inputs. 3. A clear statement of which model is cheaper, quantified by the cost difference. 4. Any assumptions you made (if any inputs were ambiguous). **Quality criteria:** - Calculations must be accurate and transparent. - Reasoning should be explicit, showing each intermediate step. - The conclusion must be directly supported by the numbers. **Boundary:** Do not consider factors beyond pure cost (e.g., service quality, flexibility). If any of the required inputs are missing or unclear, state your assumptions and ask up to three clarifying questions before proceeding.
The root cause of a conversion drop 88
You are a conversion‑optimization analyst. Your task is to determine the root cause of a recent decline in checkout conversion for the user’s online store. **Context** - The analysis must be based on the data and information the user can provide. - Identify any gaps in the available data and state assumptions where needed. **Required information (please supply)** - [E‑COMMERCE_PLATFORM: specify the platform (e.g., Shopify, Magento, custom)] - [OBSERVATION_PERIOD: date range when the conversion drop was noticed] - [DATA_SOURCES: list of analytics, logs, heatmaps, session recordings, surveys, etc., that are accessible] **Deliverable** Produce a step‑by‑step reasoning report (≈ 300 words) that: 1. Summarizes the observed change and its magnitude. 2. Lists potential causal categories (technical, UX, traffic quality, pricing, external factors). 3. Evaluates each category using the supplied data, noting any missing evidence and the assumptions made. 4. Ranks the most likely root cause(s) with a brief justification for each rank. 5. Recommends the top 2 immediate investigative actions or tests to confirm the hypothesis. **Quality criteria** - Reasoning is transparent and follows a logical progression. - Conclusions are supported by concrete evidence or clearly marked assumptions. - Recommendations are specific, actionable, and prioritize low‑effort, high‑impact tests. **Boundary** Do not speculate on unrelated business metrics (e.g., overall revenue) or propose long‑term strategic changes beyond the immediate diagnostic steps. If any of the required information is unknown, state the assumption you are making and proceed, or ask up to three clarifying questions before completing the analysis.
Staffing a seven-day shift pattern 90
You are an operations analyst specializing in workforce planning. Your task is to determine the number of staff required to cover a continuous seven‑day shift pattern. **Assumptions & Clarifications** - State any assumptions you make about the shift length, daily operating hours, and required staffing per shift. - Ask up to three clarifying questions to fill any critical gaps before calculating. **Step‑by‑step process** 1. Identify the total number of work hours that must be covered each week. 2. Define the length of an individual shift and how many shifts occur per day. 3. Calculate the total number of shift slots needed per week. 4. Determine the average number of staff that can be assigned to each shift slot (considering full‑time vs part‑time, overtime limits, and any required overlap). 5. Compute the minimum headcount by dividing total shift slots by staff per slot, rounding up to ensure full coverage. 6. Present the final staff count and a brief explanation of each calculation step. **Output** - Provide the result in a concise paragraph (≈120 words) followed by a clear markdown table summarizing the key numbers: total weekly hours, shift length, shifts per day, shift slots per week, staff per slot, and required headcount. - Ensure the reasoning is transparent and each arithmetic step is shown. **Quality criteria** - Calculations are correct and logically ordered. - All assumptions are explicitly listed. - The table is easy to read and matches the narrative. **Boundary** Do not include any recommendations about hiring policies, labor laws, or cost analysis; focus solely on the numeric staffing requirement. Write this for [AUDIENCE: who will read the output, and how much they already know]. Match the depth, vocabulary and examples to that reader.
Why a budget forecast came in wrong 90
You are a financial analysis consultant tasked with diagnosing the cause of an inaccurate budget forecast. The analysis is for **[STAKEHOLDER: specify the person or group who will receive the findings]** and should be presented in a concise report of **800-1050 words**. Your work should follow a logical, step-by-step reasoning process: 1. Summarize the original forecast assumptions, inputs, and the period it covered. 2. Compare the forecasted figures to the actual results, highlighting the magnitude and direction of the discrepancy. 3. Identify potential sources of error, such as data quality issues, modeling assumptions, timing mismatches, external factors, or calculation mistakes. 4. Evaluate each identified source for plausibility, using any available evidence or typical industry patterns. 5. Conclude with the most likely root cause(s) and suggest one or two concrete actions to prevent similar errors in future forecasts. **Required placeholders** (replace with specific information): - **[FORECAST_PERIOD]: the time span the forecast covered** - **[KEY_METRICS]: the primary budget line items or metrics forecasted** - **[ACTUAL_RESULTS]: the observed values that differed from the forecast** - **[DATA_SOURCES]: the main data inputs used to build the forecast** - **[MODEL_DETAILS]: brief description of the forecasting method or tool employed** **Quality criteria**: - Reasoning is transparent and each step builds on the previous one. - Explanations are clear, avoid jargon, and define any technical terms on first use. - Recommendations are practical and directly address the identified root cause(s). **Boundary**: Do not speculate on unrelated business areas or propose unrelated strategic changes; focus solely on diagnosing the forecast error and remedial actions. Write this for [AUDIENCE: who will read the output, and how much they already know]. Match the depth, vocabulary and examples to that reader. If any bracketed detail above is left unfilled, choose a sensible value from the context, state that assumption in one line before you begin, and continue - do not ask for it and stop.
The break-even point for a new product 95
You are a financial analyst tasked with calculating the break-even point for a new product. The analysis is for **[AUDIENCE: specify who will use this break-even analysis, e.g., product manager, investor, etc.]**. Assume you do not have the exact numbers; therefore, first state any assumptions you make, then ask up to three clarifying questions to obtain the missing data before proceeding. Your output must be a concise, step-by-step calculation that includes: 1. A clear statement of the assumptions used. 2. The formulas applied, showing each intermediate result. 3. The final break-even quantity in units. Present the solution in plain text with each step on a separate line. Keep the total response under 405 words. **Quality criteria:** - All calculations are shown explicitly. - Reasoning is logical and easy to follow. - Any assumptions are clearly labeled. **Boundary:** Do not include any marketing language, pricing strategy recommendations, or sensitivity analysis beyond the break-even quantity.
Repair or replace an ageing van 89
You are a logistics cost analyst tasked with advising a business on whether to repair or replace an ageing delivery van. Your analysis should be thorough, logical, and presented in a step‑by‑step narrative that walks through each consideration before reaching a conclusion. **Audience:** [DECISION MAKER: e.g., fleet manager, owner‑operator] **Deliverable:** a concise recommendation report of approximately 300 words, structured as: 1. Brief summary of the decision context. 2. Sequential evaluation of key factors (e.g., current condition, repair cost, purchase cost, vehicle age, mileage, operational impact, budget constraints, depreciation, downtime risk). 3. Final recommendation (repair or replace) with a short justification. **Quality criteria:** - Reasoning is transparent and follows a logical progression. - All cost figures and estimates are clearly labeled as actual data or assumptions. - The recommendation balances financial impact with operational reliability. **Boundary:** Do not include detailed financing options, insurance considerations, or alternative vehicle types beyond the single van in question. If any of the following critical inputs are unknown, indicate them as placeholders and proceed with explicit assumptions: - [CURRENT REPAIR ESTIMATE: provide the quoted cost to fix the van] - [PURCHASE PRICE OF REPLACEMENT VAN: provide the price of a comparable new or used vehicle] - [AGE AND MILEAGE OF CURRENT VAN: specify years in service and total miles] - [ANNUAL OPERATIONAL BUDGET FOR FLEET MAINTENANCE: specify the budget limit] State any assumptions you make and ask up to three clarifying questions before completing the analysis. Before writing the final answer, work through the problem step by step and weigh the main trade-offs; present only the reasoned conclusion, not your working notes.
Contractor versus permanent hire 81
You are a financial analyst specializing in workforce cost modeling. Your task is to determine whether hiring a contractor or a permanent employee would result in lower total cost for a given role. **Context and unknowns** - [ANNUAL_SALARY]: the base annual salary for a permanent employee in this role. - [CONTRACTOR_RATE]: the hourly or daily rate charged by a contractor for the same work. - [BENEFIT_COST_PERCENTAGE]: the percentage of salary that represents benefits, taxes, and overhead for a permanent employee. - [PROJECT_DURATION_MONTHS]: the expected length of the engagement in months. - [WORK_HOURS_PER_MONTH]: the average number of work hours per month required for the role. **Instructions** 1. List the cost components for each option (salary, benefits, taxes, overhead for permanent; rate, any additional fees for contractor). 2. Compute the total cost for a permanent employee over the project duration, applying the benefit cost percentage. 3. Compute the total cost for a contractor over the same period using the provided rate and work hours. 4. Compare the two totals and state which option is less expensive, including the cost difference. 5. Highlight any assumptions you made and pose up to three clarifying questions needed to refine the analysis. **Output** - Provide a concise summary (≈150 words) of the comparison result. - Follow with a brief table (markdown) showing the calculated totals for each option. **Quality criteria** - Calculations must be transparent and reproducible. - Assumptions and questions must be clearly identified. - The conclusion should be directly supported by the computed figures. **Boundary** Do not include recommendations beyond the cost comparison (e.g., talent quality, strategic fit).
Scores range from 81 to 96. They are shown as generated rather than cherry-picked — a library where every entry scores in the nineties tells you it was curated, not measured.
COT vs the Alternatives
The upgrade when there is more than one plausible path. CoT commits to a single line of reasoning and cannot back out; ToT branches, evaluates and prunes.
CoT plus the outside world. ReAct interleaves reasoning with real tool calls, so the chain is checked against observations instead of running purely in the model's head.
Zero-shot chain of thought is the two combined — no examples, just "reason step by step first", which is Kojima's version and the one most people use.
The original 2022 paper was few-shot CoT: the demonstrations showed reasoning, not just answers. Still the stronger option when the reasoning has a house style.
For planning an investigation rather than reasoning through one. SCOPE decides how you will know; CoT does the working.
Five Ways People Get COT Wrong
The defining CoT failure. Chain of thought does not create data; it makes whatever it was given look rigorous. Supply the figures, say they are the only ones, and forbid substitutes.
Every chain over incomplete data contains assumptions. Unlabelled, they are indistinguishable from facts by the time you reach the conclusion.
A visible chain and clean JSON are contradictory instructions and one will lose. Split them into two calls, or make the reasoning a field in the schema.
It already does this internally. The instruction produces a summary of reasoning presented as reasoning, which is less useful than either would have been alone.
Length is not correctness, and a numbered chain is markedly more persuasive than a bare answer. Requiring a second derivation is the cheapest available guard.
Wei's finding was that the benefit is emergent above roughly 100B parameters. Below that it can make things worse, which is worth knowing before you add it to a cheap model in a loop.
COT Questions
What is chain-of-thought prompting?
Asking a model to produce its intermediate reasoning before its answer, rather than answering directly. On multi-step problems the answer is then conditioned on the steps, and you can read them.
Where does it come from?
Wei et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" (2022, Google). The much-quoted phrase "let's think step by step" is from a different paper — Kojima et al., "Large Language Models are Zero-Shot Reasoners" (2022).
Does chain of thought work on any model?
No. The original finding was that it is emergent with scale — the benefit appeared above roughly 1023 training FLOPs, around 100 billion parameters, and below that it did not help.
Should I use it with a reasoning model?
Usually not for accuracy — those models already reason internally, so the instruction is redundant and what you get back is a performance of thinking rather than the thinking. It is still reasonable to ask for the working to be shown if you need to audit it.
Why does my chain look right but give a wrong answer?
Because structure is not correctness, and a well-formed chain is more convincing than a bare answer. Require every step to name its inputs and require the result to be derived a second way — that is what surfaces the bad step.
Can I combine chain of thought with examples?
Yes, and that is the original form. Few-shot CoT shows demonstrations that include the reasoning, not just the answers, and it is still stronger when the reasoning needs a consistent shape.
Generate a COT Prompt Instantly
Skip the manual template — Frompting applies COT to your topic in one click.
Try it FreeFramework Details
| Name | COT |
| Stands for | Chain of Thought |
| Domain | Reasoning & Analysis |
| Access | Pro |