Beyond Prompting
How working with AI moved from writing prompts to engineering systems
Traditional prompting
The reasoning shift
New prompting rules
Context engineering
Harness engineering
Loop engineering
Agentic AI
The SCRIBE framework
Six habits that made early chat models useful. The human did the thinking; the prompt carried it.
Specificity
Say exactly what you want: length, format, audience and scope.
Context
Give the background: who, why, and what has already happened.
Role
Assign a persona to set expertise, register and point of view.
Instructions
Spell out the steps the model should follow, in order.
Banter
Iterate in conversation: push back, refine, ask for another version.
Evaluate
Check the output for accuracy, tone and fit before you use it.
SCRIBE in practice: a typical 2023 prompt
You are a senior management consultant with 20 years’ experience in GCC retail.
Our client, a Dubai retail group, has seen mall footfall fall 12% since Q1.
Write a one-page memo for the CEO, maximum 400 words, in British English.
Step 1: list five likely causes. Step 2: rank them. Step 3: recommend three actions. Think step by step.
NEXT TURN“Make it more concise and add a cost estimate for each action.”
HUMAN CHECKVerify the figures, the tone and the logic before sending.
Why it worked then
- No internal planning. Step lists stood in for reasoning the model could not do on its own.
- Weak intent recognition. Personas and detail pulled the model towards the right register.
- One shot, no questions. Banter was the only way to correct a wrong first attempt.
- Errors were common. Human evaluation was the last and only line of defence.
Reasoning models made the scaffolding redundant
They plan before they answer
Built-in reasoning, with an adjustable effort setting, replaces “think step by step”.
They infer intent, and ask
Given a goal, they work out the method, and ask questions when the brief is unclear.
They follow plain prose
Shouting, stacked bans and rigid recipes now create conflicts, not compliance.
Evidence: Meincke et al., Wharton Prompting Science Reports 2–4. Forced chain-of-thought adds time for little gain on reasoning models; tips and threats show no reliable effect; expert personas do not improve factual accuracy.
What happened to each letter of SCRIBE
The same request, rewritten for a reasoning model
Five rules for prompting a reasoning model
Describe the goal, not the method
State the deliverable and what “good” looks like. Let the model plan the route.
Give the reason behind each constraint
“It goes to procurement, so keep it formal” lets the model generalise correctly.
Define boundaries, not lists of bans
Scope, length, format and audience, in plain prose. No capital-letter warnings.
Drop the thinking scaffolds
Remove “think step by step”. Set the reasoning effort instead.
Invite questions before work begins
“Ask me what you need first” turns guesswork into a confirmed brief.
What good looks like in 2026
- Show, don’t describe. Attach real examples, templates and rubrics instead of adjectives.
- Keep the prompt short. Put the substance into context: files, data, past work.
- Use personas for perspective only. “Review this as a sceptical CFO” still helps.
- Choose the right model and effort level. Quick tasks suit a fast model with little reasoning; hard trade-offs need a deeper model and more effort.
- Test against several real cases, not one lucky run. Change one thing at a time.
Control what the model sees, not only what you ask
A clear result
The deliverable, audience, depth and edge cases, stated up front.
Context that changes the answer
Non-obvious facts: past failures, conventions, trade-offs and the intent behind limits.
Real examples
Actual documents, data, templates and rubrics, not adjectives.
Permission to question
Let the model flag gaps and risks before it starts work.
The harness is the platform the model works inside
Let the model find its own context
Give it tools to search, read and run, rather than pre-stuffing every request with retrieved chunks.
Model + harness is what you buy
Benchmarks measure the pair. A better harness often beats a bigger model in a poor one.
Hooks turn policy into code
Block destructive commands before they run; check every change after it lands.
Loops: the system iterates until the goal is met
A clear goal
Success criteria, quality bar and scope, defined before the first run.
Specialist agents
Researcher, builder and verifier, each with a focused role.
Memory
Files that record what was tried, so no effort is repeated.
Independent verification
Tests, checkers or a separate verifier, never the builder marking its own work.
Stopping conditions
Done when criteria pass, with token caps and time-outs as a backstop.
A loop without a stopping rule is a bill without a ceiling.
Agentic AI: from one model to a digital workforce
Sub-agentsisolated work, summary back
Agent teamshared task list, peers talk
Choose the lightest pattern that does the job
From prompt writer to system designer
Prompt
What you ask
Context
What it sees
Harness
What it can do
Loop
How it improves
© 2026 Galal & Karawi Management Consulting, DMCC. All rights reserved.
Parts of this presentation may have been drafted, proofread or formatted with the assistance of AI tools, under the direction and review of G&K consultants.