Auto-translation used
Read the original

NEW INDUSTRIAL ENGINEERING TECHNIQUES FOR GPT-5

Who will need it: developers, product and data specialists. Below are practical techniques for designing promts for GPT-5 level models. Focus on reproducibility and quality, without commerce.

  • Deep reasoning "out of the box" — you can give complete scenarios instead of dozens of micro-projects.
  • Built—in self-check - the model is able to criticize and refine its own response.
  • Multi-role and multi-format — one query → multiple roles and result formats at once.

The model first chooses a strategy, then solves the problem.

You are an expert in industrial engineering. Suggest 3 strategies for solving the problem (approach, pros/cons, risks). Choose the optimal one and only then complete the task.

Self-check according to quality criteria.

Form a response. Then evaluate it according to 3 criteria: completeness, accuracy, clarity (0-10). If any < 9, refine it and show the final version.

Suggest ≥3 alternative strategies. For each one: steps, time, risks. Compare and justify the choice of one.

Stage 1: Market analyst → short research. Stage 2: Productologist → value hypotheses. Stage 3: Copywriter → landing page abstracts. Give the result in a single package.

At the end of the answer, indicate the confidence (0-100%). If < 80%— specify what to specify and where to get the data.

Give me a brief concept (3-4 sentences). Ask 5 clarifying questions. After my answers, a detailed plan.

Remember the initial parameters of the project (the list). Use them in further answers, even if I don't repeat them.

Based on one idea:

  1. blog article (800-1000 words),
  2. LinkedIn post (120-180 words),
  3. Reels scenario < 60 sec (hook–value–inform. CTA).

Model the discussion: marketer, analyst, designer. Everyone gives a position and objections. In the final, there is an agreed plan and metrics for success.

Solve the problem. Then reformulate the query in a different way and check if the result matches. If not— fix it and describe what you changed.

You can independently switch the role (analyst/engineer/editor) by stages. Mark the role change with a short marker.

Analyze the .csv (columns: ...). Make aggregates and describe which graphs are needed and why (Markdown is the structure of the report).

  • Problem statement: from strict instructions to meta-promts and strategy selection by the model.
  • Thinking: "think step by step" → deep analysis by default + branching (ToT).
  • Quality: manual check → self-critique/self-refinement in one request.
  • Roles: we prescribed each → dynamic role change within the same scenario.
  • Verification: separate promt → chain-of-verification in the same run.
  • Formats: manually converted → multi-format from a single idea.
  • The scenario is formulated as: goal → quality criteria → output format.
  • At the end, there is a self-test and a confidence assessment.
  • If necessary, branch out and choose the best option.
  • Confidential data is not transferred; anonymization is on the team's side.
  • Figures and facts are validated by external sources/internal data.
  • Google, "Prompt Engineering" (whitepaper, February 2025): temperature/top-p/top-k settings, techniques and best practices.
  • Research on Tree-of-Thoughts, ReAct, and Self-Consistency: how reasoning methods improve the quality of inference.
  • The company's own regulations on data security and quality.

Note: the examples of promts are educational. Validate on your datasets and policies before selling.

ReceptionIt was in GPT-4oIt became in GPT-5
Meta-PromptingIt was necessary to describe in detail the steps of the solution and the approach in one draft. The model often performed immediately, without choosing a strategy.She can suggest 2-3 strategies herself, evaluate the pros/cons, and choose the optimal one before executing.
Self-Critique / Self-RefinementA separate audit was required to verify and finalize the response.Self-check and improvement in one query: the model analyzes and refines according to criteria.
Tree-of-Thoughts (ToT)I had to manually request a list of alternatives and compare them separately.Builds branches, evaluates options, and justifies the choice automatically.
Multi-Role PromptingIt was necessary to write several separate scripts for each role.In one scenario, the model switches roles, passing results between stages.
Confidence ScoringConfidence assessments were rarely made and without explanation.Indicates the percentage of confidence + what to clarify if < 80%.
Iterative DeepeningStep-by-step clarification through a series of dialogues.He makes a short plan, asks clarifying questions, and immediately gives out a detailed study after the answers.
Memory EmulationThe "memorization" of the context was limited: it was often necessary to repeat the introductory ones.It can keep the project parameters within the session and refer to them without repeating.
Style Transfer + Multi-Format OutputIt was necessary to request each format separately (article, post, script).One promt → multiple formats at once, while maintaining a single style.
Multi-Agent SimulationI had to describe the role dialog manually.The model itself simulates a team discussion with arguments and a final plan.
Chain-of-VerificationCross-validation was done through separate requests.Query reformulation and verification are built into the same startup.
Dynamic Role AdjustmentIt was necessary to rigidly prescribe the roles and stages in advance.The model decides when to change the role and marks the change with a marker.
Embedded Tool UseThe analysis of the files was fragmentary, with no structural conclusions.It can analyze .csv/.json, make aggregates, offer graphs and justify the choice.

Are there any practical cases of adapting promts to GPT-5 (without NDA)? Briefly describe the context, what has been changed and which quality/speed metrics have improved.

530
avatar

Сергей Конев

Aug. 11, 2025 09:27

Comments 0

Login to leave a comment