Prompts
The instruction layer: what you tell the model, and whether you can edit it.
Every AI tool runs on instructions — the one you type and the hidden system prompt its maker wrote. Once a prompt sits inside a repeating process, wording choices compound into real money: a structured, cached prompt can run the same workload 2-10x cheaper than a naive one. And if a vendor hides the prompt layer, you can't fix its mistakes — you can only file tickets.
You need this when
- The same instruction runs hundreds of times a day inside a process
- Output quality swings between runs and nobody can say why
- You're re-paying to send the same rules on every single call
You can skip it when
- One-off drafts — just say what you want, show one good example, correct it once
- Hunting magic phrases — instead state the job, the boundaries and the output shape
- 500-word mega-prompts — cut to the rules that change behaviour, then test on real cases
The long version — open when you want the depth
What it is — in one coffee-break
A prompt is the written instruction you give an AI model — and prompt engineering is the craft of writing instructions that produce reliable results instead of lucky ones. Every AI tool you touch has prompts inside it: the visible one you type, and usually a hidden "system prompt" the tool's maker wrote to set the rules, tone and boundaries before you ever say a word.
The discipline has split cleanly in two. Casual prompting — asking a chatbot for a draft email — barely needs skill anymore; the models got good at reading intent. Production prompting is different: when the same instruction runs thousands of times inside a business process, small wording choices compound into real money and real quality differences. Practitioner guidance in 2026 treats prompts like code: versioned, tested, and structured deliberately.
When you actually need it (and when you don't)
You don't need prompt engineering to use ChatGPT for a one-off task — just say what you want, give an example of good output, and correct it once. You DO need it the moment a prompt becomes part of a repeating process: the assistant that drafts your customer replies, the workflow that summarizes invoices, the agent that triages your inbox. There, three habits pay for themselves.
First, split stable from changing. Rules, persona and format instructions go in the system prompt; the specific task goes in the user message. This isn't just tidiness — providers cache the stable part, which can cut costs by up to 90% and latency by 85% on repeated calls. A timestamp accidentally placed in your system prompt silently breaks that cache on every single call.
Second, define what good looks like. Strong prompts aren't magic phrases — they state the job, the boundaries, and the shape of a good answer, then get tested against real examples. Third, keep them short. Debugging a 500-word prompt when quality suddenly drops is miserable; change one sentence and three other behaviours shift.
How to recognize good vs bad implementations
On our Periodic Table audits, the Prompts element (Pr) separates tools that give you real instruction control from tools that hide it. Claude scores 8.5 — system prompts are a first-class, documented surface with caching built in. ChatGPT scores 8 for similar reasons. At the other end, DeepSeek's hosted app scores 6: capable model, thin prompt tooling around it. The tell in any tool you evaluate: can you see, edit and version the instructions it runs on? If the vendor hides the prompt layer entirely, you can't fix its mistakes — you can only file tickets.
What this costs
Prompt engineering itself is free — it's craft, not software. The money shows up indirectly: a well-cached, well-structured prompt can make the same workload 2–10× cheaper than a naive one, because you stop re-paying for the same instructions on every call. For a small business running a few hundred AI calls a day, that's the difference between a rounding error and a bill you notice.
Where to see it scored
Every audit passport scores Pr with a written reason: Claude (8.5), ChatGPT (8), Gemini (8.5), n8n (7). For the system that feeds prompts their surrounding information, read the next lesson: Context.
Flashcards
Check yourself
1. Where do persona, rules and output format belong?
2. Across the 62 tools audited, which scores highest on Prompts (Pr)?
3. You add a live timestamp to your system prompt. What happens?
4. What is the best single test of a vendor's prompt layer?
5. What does prompt engineering itself cost?
Cheat sheet
- Stable rules in the system prompt; the specific task in the user message.
- Caching the stable part cuts cost up to 90% and latency up to 85% on repeat calls.
- A timestamp in your system prompt silently breaks the cache on every call.
- On Claude, cache reads cost 0.1x base input price; cache writes cost 1.25x.
- Short prompts debug fast. Edit one line of a long one and three behaviours shift.
- Vendor test: can you see, edit and version the instructions it runs on?
Who actually does this well
| Best on this element | Score | Why it scored that |
|---|---|---|
| Decagon | 9 | Agent Operating Procedures are a natural-language-plus-code instruction layer, compiled into executable agent logic, Git-versioned, with a co-pilot (Duet) that auto-generates and r |
| Bland | 8.5 | Prompting is decomposed to the level it should be: per-node dialogue prompts, loop-condition prompts and variable-extraction prompts each run independently, sitting under a pathway |
| Claude / Anthropic | 8.5 | Console prompt generator, improver and shareable prompt library plus prompt caching and strong prompting docs; the full prompt-engineering loop lives in one place. |
| Gemini / Google | 8.5 | AI Studio prompt tooling, system instructions, context caching and an automated prompt optimizer that tunes prompts from failure analysis on the agent platform. |
| GitHub Copilot | 8.5 | Repository-level custom instructions, path-scoped instructions.md, prompt files, reusable Agent Skills and .agent.md custom agents make instruction design version-controlled and re |
| Langfuse | 8.5 | Prompt management is a first-class product, not a side feature: versioning, release labels, composability, server and client caching, unlimited prompt fetching on every tier includ |
| Retell AI | 8.5 | Full prompt-level control — single-prompt agents or node-based flows, direct system-prompt editing, plus a built-in AI copilot that helps build, test and refine the prompt with you |
| Arize Phoenix | 8 | Prompt management is a real product surface: versioning, tagging, a playground that compares models side by side, span replay to re-run a traced call with a changed prompt, prompts |
And the other end of the same column:
| Weakest | Score | Why it scored that |
|---|---|---|
| Ambience Healthcare | 3.5 | Clinicians get specialty-tuned templates and note structure that follows their own documentation style, and Ambience tunes templates per health system during implementation. There |
| Paradox (Olivia) | 3.5 | Admins configure knockout questions, FAQ content and Olivia's tone, name and avatar, but there is no exposed prompt surface, no system-instruction control and no prompt versioning. |
| Cornerstone OnDemand | 3 | There is no prompt surface. Admins configure which agents are switched on and for whom; nobody outside Cornerstone writes or versions a system prompt, and no prompt template, playg |
| Guild | 2 | No prompt or instruction surface exists for either the employer or the learner. Whatever Guild's AI is told, it is told inside a managed service that publishes nothing about it. |
Scored on 142 of 142 audited tools. Every score links to the full audit and its reasoning.