← Home / Method

The method

How an Onuris prompt is built.

Training an AI is not magic and it is not fine-tuning. It is a brief: the context an AI needs to do one job well, and nothing it does not. Every prompt we sell is built on the same six principles and assembled from the same twelve components.

Fig. 01 — Six principles

What a model needs to do good work.

I

Brief it like a new hire

A model is highly capable and knows nothing about your firm. It starts every conversation with no memory of the last. So each prompt carries what a new colleague would have to be told: why the work matters, who will read it and the standard expected.

II

The right context, and no more

Too little, and the AI guesses. Too much, and the signal is buried while the cost rises. Each prompt names exactly the inputs the task needs. That is what the four tiers are: four measured amounts of context.

III

Define done, not only do

Instructions say what to do. Success criteria say what must be true at the end. Models work far better against a clear target, so every Plane and Solid prompt closes with conditions the result has to meet.

IV

Room to reason

A model reasons by writing. Asked for the answer at once, it has nowhere to work. Our prompts have it go through the method in notes first, and only then write the deliverable.

V

Permission not to know

Left alone, a model fills gaps with plausible guesses and does not mention it. Every prompt tells it to state its assumptions, point out conflicts in the inputs and write “not supplied” when something is missing.

VI

Built to be checked

AI output is only useful if a person can verify it quickly. Claims are tied to named inputs, calculations are shown, and Solid prompts end with a review list and clear rules for when a person must step in.

Fig. 02 — Anatomy

Twelve components. Each tier adds more.

Every prompt is assembled from the same parts. A Point prompt uses three of them. A Solid prompt uses all twelve.

  1. 01Rolefrom Tier I · Point
  2. 02Taskfrom Tier I · Point
  3. 03Output formatfrom Tier I · Point
  4. 04Purposefrom Tier II · Line
  5. 05Audiencefrom Tier II · Line
  6. 06Inputsfrom Tier II · Line
  7. 07Rulesfrom Tier II · Line
  8. 08Methodfrom Tier III · Plane
  9. 09Examplefrom Tier III · Plane
  10. 10Success criteriafrom Tier III · Plane
  11. 11Staged passesfrom Tier IV · Solid
  12. 12Escalationfrom Tier IV · Solid

Fig. 03 — Four tiers

From a point to a solid.

More context is not always better. A quick task wants a Point; a board paper wants a Solid. Each pack gives you all four so the depth matches the work.

I · Point — the research

A single precise instruction that asks for step-by-step work. Asking a model to work step by step improves its reasoning even with no examples (Kojima et al. 2022).

II · Line — the research

Purpose, audience, delimited inputs and rules, including the UK law that applies. Specific cues steer a model toward what matters (Li et al. 2023; Anthropic and OpenAI guidance).

III · Plane — the research

An input check, a method worked through in notes, your own best examples and success criteria (Wei et al. 2022; Brown et al. 2020).

IV · Solid — the research

Three specialist viewpoints at each judgement, a stop to ask for missing data, work built in parts and an audit against UK rules (Yao et al. 2023a and 2023b; Besta et al. 2023; Bai et al. 2022).

Fig. 04 — See it assembled

Change the settings. Watch the brief change.

Pick any department, model and tier. The components and word count below the prompt are measured live.

A Department
B Your AI

C Tier

      
Words0
≈ Tokens0
Components0
Variables0

Fig. 05 — Calibration

The same brief, cut five ways.

Each model family has a structure it follows most reliably. You choose your AI; the pack arrives in its format.

01

Claude

Anthropic

  • XML-tagged sections, one per component
  • Working notes kept in a scratchpad, deliverable returned separately
  • Purpose and audience set before the task
02

ChatGPT

OpenAI

  • Markdown headings held as fixed structure
  • Told to ask before assuming anything not supplied
  • Laid out so long briefs stay easy to navigate
03

Gemini

Google

  • Persona, task, context and format labelling
  • Capitalised section labels
  • Every figure traced back to its input
04

Copilot

Microsoft

  • Goal first, then context, source and expectations
  • Works from the files and threads you name
  • Each claim linked to the file it came from
05

Open models

Llama, Mistral and others

  • Compact section markers for smaller context windows
  • Tight output guards
  • Notes first, then the deliverable, with no preamble

Fig. 06 — The evidence

Built on published research, not on hunches.

Every structural choice in an Onuris prompt follows a published finding, or the guidance the AI makers publish for their own models. The aim is output that is faster for a person to check, because checking is now where AI work slows down. When that guidance moves, the method moves with it, and every pack is rebuilt from the same source.

I

Reasoning before the answer

Researchers at Google showed in 2022 that large models reason far better when asked to write out their intermediate steps, and that steps written after the answer add nothing. A second study that year found that simply asking a model to think step by step lifted its accuracy on an arithmetic benchmark from 18% to 79%. So every Line, Plane and Solid prompt has the AI work step by step in notes before it writes a word of the deliverable.

II

Instructions and data kept apart

Anthropic and OpenAI both advise marking clearly where instructions end and data begins, and Anthropic recommends restating the immediate task near the end of a long prompt. Each input you supply sits in its own labelled block, and every brief closes by restating the job and the format, so a figure in your accounts is never read as an instruction.

III

Evidence first, your standard always

Anthropic’s guidance on avoiding invented answers is to give the model permission not to know, and to have it gather its evidence before it answers. Our prompts do both, tie every figure to a named input, and carry a slot for your firm’s own best work, because examples are among the strongest ways to steer a model.

IV

Explore before committing

In Tree of Thoughts (2023), letting a model weigh several lines of reasoning and judge its own options before committing raised GPT-4’s success on one planning puzzle from 4% to 74%. Its authors also note such deliberation is unnecessary for tasks models already do well. So Solid prompts weigh competing explanations at each key judgement, while the lighter tiers stay lean.

V

Divide, merge, refine

Graph of Thoughts (2023) found the largest gains on work that splits into parts solved separately and then merged, with feedback loops to refine the result. Solid prompts build in parts, check that the merged whole agrees with itself, and revise through scored critique until every criterion is met, or flag the gap for you.

VI

Fixed, tested wording

A 2026 study found AI writing shifts markedly with how a prompt is specified, and that naming a feature can push it into exaggerated, formulaic use. Earlier research found that for most users, prompting is trial and error. So our wording is fixed and versioned, tone comes from your own example rather than a list of style features, and we advise running important work twice and comparing. Every sentence also follows the ASD-STE100 writing rules: short, one instruction each, with nothing left vague.

VII

Stop, ask, then continue

ReAct (2023) let models interleave reasoning with actions, such as looking up a missing fact, and this cut invented facts compared with reasoning alone. In a single prompt the nearest action is to ask you. So Solid prompts check every input first, and stop to ask for a missing figure instead of guessing it.

VIII

A written standard to check against

Anthropic’s Constitutional AI (2022) had a model critique and revise its own answers against a written list of principles. Anthropic used it in training. We apply the same idea at the point of use: Solid prompts audit the draft against the department’s UK rules and rewrite any part that fails.

IX

Point attention at what matters

Directional Stimulus Prompting (2023) found that short, specific hints steered models to better results than the bare task. So every Line, Plane and Solid prompt names the UK rules and tests that matter in its department, and never lets the AI claim that work complies with a law.

Sources

  1. Wei, J. et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022.
  2. Kojima, T. et al. (2022). Large Language Models are Zero-Shot Reasoners. NeurIPS 2022.
  3. Anthropic. Prompt engineering interactive tutorial.
  4. OpenAI. Best practices for prompt engineering with the OpenAI API.
  5. OpenAI. Prompt engineering guide.
  6. Yao, S. et al. (2023a). Tree of Thoughts: Deliberate Problem Solving with Large Language Models. NeurIPS 2023.
  7. Besta, M. et al. (2023). Graph of Thoughts: Solving Elaborate Problems with Large Language Models. arXiv:2308.09687.
  8. Dang, H. et al. (2022). How to Prompt? Opportunities and Challenges of Zero- and Few-Shot Learning for Human-AI Interaction in Creative Applications of Generative Models. CHI 2022 Workshops.
  9. Crosthwaite, P., Sun, S. and Jiang, F. (2026). Prompt Sensitivity in AI Writing: A Methodological Caution for Corpus-Based Analysis of AI-Generated Language. The Modern Language Journal.
  10. Brown, T. B. et al. (2020). Language Models are Few-Shot Learners. NeurIPS 2020.
  11. Yao, S. et al. (2023b). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.
  12. Bai, Y. et al. (2022). Constitutional AI: Harmlessness from AI Feedback. Anthropic, arXiv:2212.08073.
  13. Li, Z. et al. (2023). Guiding Large Language Models via Directional Stimulus Prompting. NeurIPS 2023.
  14. Chen, F. and Stratton, J. (2026). Artificial Intelligence in the Firm: Bottlenecks in Software Production. Harvard University working paper. Summary in InfoQ.
  15. ASD (2025). ASD-STE100 Simplified Technical English, Issue 9. AeroSpace, Security and Defence Industries Association of Europe. We apply its writing rules, not its dictionary.

Onuris is independent, and is not affiliated with or endorsed by Anthropic, OpenAI, Google, Microsoft or the authors cited. The UK rules in each pack help a team work to them. They are not legal, tax or financial advice. Product names are trademarks of their owners.

Raise an army of trained agents, one department at a time.