Two people, the same model, the same hour. One walks away with a usable draft; the other walks away convinced AI is overhyped. The model is identical, so the difference sits somewhere else. Buyers name it the same way again and again: their people type into a natural-language tool the way they type into a search box, and get the generic answer a search-box request earns.
Prompting Skill
How Do You Write AI Prompts
That Get Good Results?
Why a search-box habit returns boilerplate, the three-part structure that fixes it, and the limit no amount of prompt craft can cross.
Built from real buyer questions in our sales meetings
Direct Answer
Write a paragraph, not a keyword string. A prompt that works carries three things a search query never does: the role you want the model to play, the context it cannot see for itself — audience, source material, constraints, the shape of the output — and then the specific ask. Iternal teaches that order in its prompting sessions and drills it inside the AI Academy, where a learner writes a prompt against a scenario and has it scored on the spot. The extra sentences cost 30 to 60 seconds.
The limit: prompt craft cannot make a wrong source right. When the underlying document is out of date or simply wrong, a well-built prompt returns the wrong answer more fluently. The training path carries its own edges. Iternal states that the academy courses do not support creating customer-specific prompts, that the advanced prompt-engineering course is hard enough that people retake it before passing, and that the pre-built workflow catalog may not hold exactly what a particular customer needs — which leaves that person editing the prompt anyway.
A score shown in a demonstration is an illustration, not a measurement. Iternal has demonstrated its AI Academy grading a thin, low-context prompt at 39%, then grading the rewrite — more context, explicit requirements — at 82%. Iternal’s own training tool scored Iternal’s own exercise, the rubric and measurement method are published nowhere, and that session the exercise misfired by showing a prompt saved from an earlier demonstration instead of its usual blank field. Treat the pair as the shape of the gain, and ask for the rubric before anyone quotes the numbers.
Prompting is a context problem; a corpus is a different problem. Better prompts fix answers that come back generic. They do nothing for an answer that is wrong because the source material was wrong, and they settle nothing about whether a claim traces back to the document it came from. For more information visit the accuracy and traceable answers page.
Why a Keyword Prompt Returns a Keyword Answer
A search engine matches your words against an index. A language model composes an answer out of whatever you handed it. Give it a handful of keywords and it fills every gap you left with the average of everything it has read, which is why professionals dismiss the result on sight: boilerplate for the average person is precisely what it is. Buyers said so in their own words, over and over. End users are comfortable with AI chat tools and still cannot talk to the AI effectively; frontline staff have low literacy for prompting; a generic prompt returns a generic response tailored to nobody in the room.
Iternal holds a firm position on where the fault sits, and it deserves reading as a stated view rather than settled fact: a bad AI output today is usually user error rather than a system error. The remedy follows from the diagnosis. Map the human process step by step and the context experienced people take for granted surfaces — who the work is for, what constrains it, the standard it is judged against. Buyers already have a name for the skill and reach for it constantly: prompt engineering.
The Three-Part Prompt: Role, Context, Then the Ask
Iternal teaches a three-part structure in its short live prompting sessions, and the structure travels because it supplies what a model is missing rather than arguing with what a model is bad at. Write the three parts in order and the prompt largely writes itself:
- Role. Tell the model who it is acting as. Act as a lawyer with thirty years of experience in this practice area narrows the register before the model writes a word.
- Context. Give it what it cannot see: the audience, the source document, the limits it must respect. Constraints written into the prompt — a small budget, an executive briefing that still carries technical detail — show up in what the model proposes.
- The ask. Say exactly what you want back. Iternal’s coaching is to type it the way you would explain it to a third grader: plain, literal, specific about the format.
The same request, written both ways:
| Habit | What gets typed | What comes back |
|---|---|---|
| Search box | make a lesson plan | A plan pitched at the average classroom, tailored to nobody. |
| Three-part prompt | Act as a high-school teacher. Plan one class for 28 students against the syllabus section below, with timings, materials and a check for understanding. | A plan shaped around the class in front of you, in the format you asked for. |
Length follows from the three parts rather than from a rule: a prompt built this way runs to a paragraph, and Iternal’s beginner courses work people up from a five-word habit toward fifteen or twenty words. Two techniques compound on top. Close the prompt by asking the model for clarifying questions before it answers, which removes the guesswork about where the model is lost. When a result goes badly wrong, copy it into a fresh chat and ask for the fix there rather than arguing with the model in place.
Formal prompt frameworks are optional against modern models, and earn their keep only when you need a fixed output structure. Iternal’s advanced course covers that case: clear demarcation, XML tags, dashed section headers.
What a Prompt Score in a Demonstration Actually Proves
Numbers travel further than the caveats attached to them, so the caveats go first. The 39%-to-82% jump quoted for AI Academy prompt scoring came out of a live Iternal demonstration: the tool graded a thin, low-context prompt, the presenter rewrote it with more context and explicit requirements, and the same tool graded the rewrite higher. Iternal’s training tool scored Iternal’s own exercise. No rubric and no measurement method accompany the figures, and the exercise itself misbehaved during that run, loading a prompt left over from an earlier demonstration in place of its usual blank field.
Direction is the useful part, and direction is all the pair shows. The AI Academy puts that movement in front of a learner within seconds: it runs the submitted prompt through a model, grades what comes back, explains what was lacking, and lets the learner rewrite and be rescored, across tiers selectable by industry and job title. Evaluate the feedback loop. The score is worth nothing as a benchmark. Three questions turn the remainder into something written down rather than something remembered from a demonstration:
Pin it down: questions for your evaluation
-
Send us the rubric behind the AI Academy prompt score: what it measures, against what task, and who defined it.Whether a score shown on stage is something your own people can reproduce.
-
Confirm in writing the current number of quick-start workflows included with our license, and which of our job roles they already cover.The catalog you actually receive, and how much prompt writing your own people will still own.
-
Can course exercises be run against our own scenarios and our own documents?Whether the training transfers to the work your teams do, or stops at generic examples.
When Not to Write the Prompt at All
The fastest fix for weak prompting removes the prompt from the user entirely. AirgapAI ships more than 2,800 quick-start workflows out of the box, surfaced as icons under the chat box and covering departments across every major industry vertical. Each icon carries a full, professionally written prompt behind the scenes, so the person clicking it supplies the input while the workflow supplies the craft.
Iternal demonstrates the point unkindly to itself: a two-word command such as make post with a document attached still returns a polished, structured social post, because the workflow carried the context nobody typed. Administrators choose which workflows each role sees, and a built-in utility writes a new one from a plain description.
The catalog has a floor, and Iternal names it: the pre-built set may not hold exactly what a given customer needs, and that customer then edits the prompt. Iternal also states that workflows help while education builds the confidence and generates the new ideas. Treat them as complements: the catalog for your common tasks, the skill for everything else.
The Review Pass for Work That Only Looks Finished
Weak prompting produces output that is obviously weak, and obviously weak output gets thrown away. The expensive failure looks finished and falls apart underneath. Buyers described it precisely: applications that look like they work from the outside while the connections between the pieces are missing; generated graphics whose copy cannot be edited afterwards; drafts a person still has to rewrite; internally assembled tools nobody keeps current or governed once they exist.
Iternal’s working rule is the simplest corrective available: never let a single prompt output be the deliverable. Five questions, asked before anything leaves your hands, catch most of what slips through.
- Are the pieces wired together? Open the parts a demonstration never touches and confirm they connect to each other.
- Can it be edited where it will need editing? Assets whose copy or styling cannot be changed become dead weight at the first revision.
- Did a person edit it? A human reviewer belongs between the draft and the recipient, the discipline buyers described for AI-assisted proposals.
- Does it read like a person wrote it? Iternal strips the tell-tale punctuation out of AI-drafted email; the same pass applies to anything customer-facing.
- Who owns it next month? Name the person accountable for keeping it current, or accept that it expires quietly.
One confidently broken deliverable teaches an organization more about AI than ten good ones.
Answered elsewhere
- Whether an answer is grounded in the right documents and can be traced back to its source — see the accuracy and traceable answers page.
- What a role-by-role training program should cover, and who needs which level — see the training curriculum page.
- Getting people to change how they work once the tools land — see the adoption and rollout page.
- Plain-language definitions of the AI and deployment terminology — see the plain-language glossary.
- Producing design and web assets faster with AI — see design and web production.
- A step-by-step walkthrough of prompt structure with worked examples — see how to write a prompt.
FAQ
FAQ: Writing Prompts That Work
Because a short prompt leaves the model to guess, and its guess is the average of everything it has read. A handful of keywords typed the way you would type them into a search engine returns boilerplate aimed at the average person. Supply a role, the context the model cannot see, and a specific ask, and the answer narrows to your situation.
A paragraph — long enough to carry a role, the context and a specific ask. Iternal’s beginner courses work people up from a five-word habit toward fifteen or twenty words, and the habit keeps growing with skill. The extra writing costs 30 to 60 seconds, which makes it the cheapest quality improvement available to you.
Inside AirgapAI, yes: quick-start workflows and personas can be tailored by industry, department, job role or individual user, and a built-in utility writes a new workflow from a plain description. Iternal states the exception plainly — the academy courses do not support creating customer-specific prompts, so bespoke prompt work happens in the product rather than in the coursework.
Tell it what to change, in the same plain language you used to ask for the work: multi-step modification of an output is fully supported, subject to the model’s intelligence and how long the chat has run. When a result has gone badly wrong, copy it into a fresh chat and ask for the fix there, so the model stays on the correction.
For common tasks, no — AirgapAI ships more than 2,800 quick-start workflows that carry the prompting, and a deliberately bad two-word command still returns a structured result. For everything else, yes: Iternal states the catalog may not hold exactly what a customer needs, and that customer then edits the prompt.
Write One Paragraph More Than Feels Necessary
Take the last prompt somebody on your team complained about, add the role, the context and the specific ask, and run it again. Half a minute of typing is the whole experiment. A visibly better second answer means you have found your adoption lever; an equally poor one means the problem sits in the source material, which is a different job with a different fix.