From Prompts to Context: Design the World Your AI 'Thinks' In

Prompts are the front door. Context is the whole house.

"Prompt engineering" has continuously dominated AI discourse. We've all seen endless lists of "The 10 Best Prompts for X" or clickbait like "RIP McKinsey. No More $300K Consultants Needed!" farming attention. Meanwhile, the sharp teams quietly tracked the deeper trend: models are permeating everyday workflows at insane speed. The focus is evidently shifting from prompt engineering to context architecture.

Creative, well-structured prompts, with the right sequence of vocabulary are not going away. Prompting is a form of context. But the game is indeed changing. Models can now remember more, see more and do more. While lightyears from the capabilities we humans have, persistent memory (beyond a single chat), richer tool access, and longer context windows enable remarkable experiences that blunt the impact of a single “magic prompt.”

Sam Altman also recently talked about the growing importance and need for memory and personalisation, something they'll be focusing on for their next release. We can take inspiration from how human cognition couples memory with reasoning, perception etc. but that’s a piece for another day. Today is about one idea:

Your model’s output quality depends less on a clever incantation and more on the "world" you build around it.

Context: Race Toward ‘Artificial General Intelligence’ (AGI)

The cat is out of the bag. I don't think that's the saying, but the point still stands. Today, we have many players like OpenAI, Anthropic, Meta and Google all racing toward AGI. Unimaginable computing power and data are being used to train the latest models. To the level, we’ve now run out of human data according to Elon Musk, and so you can imagine we start to look toward synthetic data next.

It is worth noting, Apple recently released a paper basically stating their belief that LLMs were not AI's final form. Yann Le Cunn (AI thought leader) says similar things. I discussed this briefly in the first edition of the MyPromptVerse newsletter, which you can read here. More on this in time.

Irrespective of the form of AI, trillions are pouring into this race resulting in capabilities broadening in scope and depth. This means the subtle art of crafting the "perfect" prompt will matter far less. These advanced tools, powered by AI, whether LLMs or not, will increasingly become more capable of inferring intent, self-optimising instructions, and prioritising relevant context. Designing the architecture to facilitate context engineering should be the focus.

Why You Should Focus on Context

Output quality now increasingly depends on the "world" you architect for your AI. One should be paying attention and building future proof context architecture for this reason. Think back to AI models just 18 months ago. Limited memory, no real-world connections, basic text generation. Now, let's briefly recall some of the transformation that has taken place over the past few years:

Memory Going 'Elephant-Mode'... Kind Of...

Modern models routinely handle long inputs (think hundreds of pages). While very large context windows can strain comprehension and increase the chance of errors, they dramatically expand what you can load: briefs, brand guides, specs, transcripts, datasets.

The takeaway isn’t “stuff more info,” it’s design for sustained, structured, anad relevant context rather than repeating prompts.

Real-World Integration

With features like visual/multimodal reasoning, Model Context Protocol (MCP), “connectors”, and “Projects/Gems/GPTs”, AI can now plug into things you do daily, such as your emails, calendar, files, tools, and other data sources, like your camera.

Your prompt is now the front door, while the context is the whole house. Build accordingly. Architect the foundation, layout and specific rooms.

Agents Can Plan & Execute.

AI models can now spend more time reasoning, engaging in deep multi-step internal monologues (eg. chain of thought) to plan and check, before execution. They also can orchestrate and call tools.

Start to treat them like collaborators or assistants, not as autocomplete. Be active in the process. Define the mission, the tools, the guardrails, and how you’ll manage and evaluate outputs.

From Prompt Engineer to Context Architect

Prompt engineering is like building a house, only with a hammer. Whereas context engineering is using the entire workshop. To make the transition, start designing for:

• Clear navigation: so both you and the model can find what matters quickly (folders, indices, embeddings, retrieval rules).
• Mission & objectives: what “good” looks like for this task; success criteria the model can aim at.
• Accessible knowledge: brand voice, datasets, knowledge files, examples, constraints, style rules.
• Tools & integrations: APIs, search, files, spreadsheets, CMS, calendars; permissions and scopes.
• Process & QA: drafts → critiques → revisions; checklists, tests, and evals; when to escalate to a human.

Example: “Write a marketing email” vs. a system that already includes: brand voice, ICP notes, past campaign results, do‑and‑don’t lists, examples, KPIs, and constraints. The latter wins on efficiency, consistency, and conversion.

Designing your Context System

The 5 P‑Pillars Framework

1. Purpose: What job is this system hired to do? (eg. “produce on‑brand B2B emails that pass legal review in <24h.”)
2. People: Who uses it; who approves; who is the model “speaking as” and "speaking to".
3. Prior: The reusable knowledge: style, glossary, FAQs, canonical examples.
4. Payload: Inputs... briefs, datasets, files, tool outputs; retrieval strategy.
5. Protocol: The workflow: steps, tools, QA checks, escalation, logging.

Anti‑patterns to avoid

• Magic‑prompt worship: repeatable quality comes from systems, not spells.
• Context dumps: long walls of text without structure, headings or thought.
• Unverifiable claims: include sources/links and freshness notes when facts matter.
• One‑model dependence: design to swap models to de‑risk vendor changes.

CONTEXT INTELLIGENCE v1 Custom Instructions:

Use these instructions with GPT‑class reasoning/agentic models for richer, context‑first outputs. Paste into a new session’s system field or your custom GPT/project agent. You can also access my custom GPT, with these instructions set, by clicking here.

SYSTEM SETUP
- Think hard about the user input to understand their intent.
- Be clear, correct, useful; concise. Ask clarifying questions only if blocked; otherwise act with sensible reasoning and assumptions.
- Adopt relevant expert perspective(s)/field(s) (eg. “senior product strategist”); never claim real degrees, jobs, or awards unless the user provided them. Use Natural tone; no filler/BS/ai slop.
- Flag uncertainty or time-sensitive facts and propose verification. Refuse unsafe requests and suggest safer alternatives.
- Self-check before sending: accuracy, task fulfilment, brevity, structure, compliance, helpful context. Revise if any fail.

ROLE DEFINITION
- You are an AI expert in the relevant field(s), with long‑term memory, tool use, and autonomous task planning.
- You understand how to align with mission objectives and operate within defined context boundaries.

MISSION OBJECTIVE
- Complete the user’s requested task (eg. generate a marketing plan, write documentation, analyse data, build a feature), optimising for the key metrics (eg. clarity, conversion, usability, brand alignment).

SUCCESS CRITERIA
Evaluate outputs using this rubric:
- Fulfils the brief
- Aligns with known context (docs, goals, audience) and suggests high impact missing context that could help improve the output.
- Shows independent reasoning when needed
- Self-checks or reflects on quality before finalising

CONTEXT BLOCK
- Persistent Knowledge Assets
- Brand Guide / Voice Rules: [link or summary]
- Audience Persona(s): [key insights or examples]
- Prior Work / Examples to Follow: [titles/URLs]
- Data Sources or Tools Accessible: [eg. database, CSV, API access]

KEY CONSTRAINTS
- Must follow [style/tone rules]; Match the user's language/region; Write in a human, straightforward style.
- In the first reply of a new thread, include a compact TL;DR only if it helps; skip for rewrites/very short answers.
- Avoid [banned terms, formats, assumptions]
- State the expert perspective when helpful (no fabricated credentials).
- Prioritise [data accuracy, emotional tone, brevity, etc.]
- Format as [output style — Markdown table, bullet points, JSON, prose]
- Tooling or Memory Access?
[Yes/No ... and list what’s available: file readers, calculators, API connectors, memory of past interactions]

TASK LOOP
- Planning Directive: Silently outline approach, needed tools/data, and what success looks like.
- Action Directive: Execute autonomously; pause only for unclear constraints or tool/data errors; summarise when complete.
- Reflection Directive: Check for misses, errors, and inconsistencies; annotate assumptions if needed.

OUTPUT FORMAT

Use this structure unless overridden:
## Output
[final deliverable — report, email, plan, etc.]
---
## Reasoning (optional)
[if prompted to include thought process]

## Assumptions Made
[brief list of gaps filled]

## Self-Check Summary
[final evaluation: met brief? constraints? tone?]
        

How to Use It: 2 Ways

1. Open a new session and paste these instructions in.
2. Or attach as system instructions for a custom GPT, project folder, Gem, or your own agent.

You can ask it lazy things like: help create me plan one month of content, and provide me the context that would help with this. Try it with GPT-5 Thinking or Pro and let me know your thoughts!

Context Engineering Implementation Plan:

• Phase 1 | Foundation: Pick one high-impact use case/workflow. Baseline results, then run it through the CustomGPT: Context Intelligence, so you can identify missing gaps. Plan the architecture to facilitate fast retrieval of relevant high impact context assets.
• Phase 2 | Asset building: Create a small knowledge base of the high impact context assets: brand voice, glossary, examples, past winners, etc.
• Phase 3 | Team Integration & Testing: Standardise evaluation and success criteria and QA checks. Add MCP connectors for key data/tools. Relevant training on usage. Test and get feedback.
• Phase 4 | Optimise & Scale: Based on feedback, optimise and create cache reusable context, instructions and add agentic steps (planning/tooling). Begin scaling by implementing this process across other workflows.

Final Thoughts

By 2026, the winners will have the richest, most thoughtfully architected context for their purpose. Build your context system now. Treat prompts as the ergonomics on top of a well‑designed house.

Original portfolio artwork +