Cognitive Artifacts
Cognitive artifacts are encapsulated, shareable structures of meaning and reasoning — the building blocks of extended cognition. They are how humans package thought into forms that can be stored, transmitted, reused, and built upon. In augmented intelligence, they supply the units of thought.
The one idea to keep: the structure you put around a request to an AI is part of the request. A cognitive artifact — a schema, template, or filled form — turns a vague ask into a defined shape, and a defined shape gives you something you can verify field by field instead of re-reading prose to guess whether it is right.
What Are Cognitive Artifacts?
A map is a cognitive artifact. So is a checklist, a mathematical equation, a database schema, a decision tree, and a well-structured prompt template. What they share is this: each one takes complex knowledge, compresses it into a defined structure, and makes it usable by anyone who understands the format.
The term comes from cognitive science, where it describes any external object that aids or extends human cognition. A calendar extends your memory. A spreadsheet extends your ability to calculate. A diagram extends your ability to reason about relationships. These are not just tools — they are extensions of your mind, distributed across physical and digital objects.
In the context of augmented intelligence, cognitive artifacts take on a specific and critical role: they are the units of exchange between human and machine cognition. When you give an AI a well-structured prompt, you are handing it a cognitive artifact. When the AI returns a structured output — a JSON object, a decision matrix, a formatted analysis — it is handing one back. The quality of the collaboration depends directly on the quality of these artifacts.
This is why freeform text, despite being natural and easy to produce, is often the worst medium for human-AI collaboration. It is the least structured cognitive artifact available — and structure is what makes artifacts powerful.
What is a cognitive artifact?
An external object that aids or extends human cognition by compressing complex knowledge into a defined, reusable structure.
In augmented intelligence, what specific role do cognitive artifacts play?
They are the units of exchange between human and machine cognition — a well-structured prompt and a structured AI output are each cognitive artifacts.
Why is freeform text often the worst medium for human-AI collaboration?
It is the least structured cognitive artifact available, and structure is what makes artifacts powerful.
Worked example: the same task, two artifacts
Suppose you need an AI to assess one job applicant against a role. The task is identical in both cases. The only thing that changes is the artifact you hand over — and that changes whether you can trust the answer.
Freeform text request
The AI replies with three confident paragraphs: "a strong candidate with solid experience and good communication skills." It reads well. But which requirements did it actually check? Did it confirm the years of experience, or infer them? You cannot tell, so to be sure you re-read the whole CV yourself — and the AI's answer saved you nothing.
Structured artifact
The AI returns a table. "5+ years backend → unclear, evidence: 'worked on backend systems,' no dates given." Now the gap is visible. You verify each row against its quoted evidence instead of trusting the prose, and an empty evidence field flags exactly what the AI could not support.
Notice what the human added (shown in highlight): not more words, but a shape — fields, allowed values, and a demand for evidence tied to a source. That structure is the cognitive artifact. It is what makes the second answer checkable and the first merely persuasive.
Architected Data Objects
The three pillars framework introduces Architected Data Objects (ADOs) as the third pillar of augmented intelligence. ADOs are cognitive artifacts with a specific set of properties: they are standardised, efficient, and designed to encapsulate truth into shareable, universally interpretable units.
Think of the difference between a conversation about project status and a project dashboard. The conversation is rich, nuanced, and ephemeral — once it ends, the information degrades with every retelling. The dashboard is structured, persistent, and consistent — anyone who looks at it sees the same information, interpreted through the same format.
Both convey information. But the dashboard is an ADO. It compresses complexity into a defined structure. It can be shared without loss. It can be processed by machines as well as humans. It can be versioned, compared, and audited. The conversation cannot do any of these things.
ADOs are not limited to dashboards. They include:
- Schemas and templates that define the shape of information before it is created.
- Structured outputs from AI — JSON, XML, tables, typed objects — that encode information in machine-readable form.
- Knowledge atoms — the smallest self-contained units of knowledge that can be independently verified, stored, and composed with other units.
- Canonical formats — agreed-upon structures for common types of information (meeting notes, project briefs, technical specifications) that reduce interpretation overhead.
The ADO principle: Every time you convert loose information into a defined structure, you create something more durable, shareable, and computationally useful than the original. This conversion is not overhead — it is where much of the value in augmented intelligence is created.
What three properties define an Architected Data Object (ADO)?
ADOs are standardised, efficient, and designed to encapsulate truth into shareable, universally interpretable units.
What is a knowledge atom?
The smallest self-contained unit of knowledge that can be independently verified, stored, and composed with other units.
Loose Text vs Structured Knowledge
The default output of most AI interactions is prose — paragraphs of text that read like a knowledgeable person speaking. This prose is easy to consume but difficult to use. It degrades with every step of reuse.
Consider asking AI to analyse a market opportunity. It produces three paragraphs of fluent text about market size, competitive landscape, and growth potential. This feels useful in the moment. But try to use it:
- Can you compare it directly to an analysis of a different market? Not without re-reading and manually extracting the comparable data points.
- Can you feed it into another AI system for further processing? Only as raw text, losing the implicit structure.
- Can you track how the analysis changes over time? Only by comparing paragraphs, which is imprecise and labour-intensive.
- Can another person quickly find the specific data point they need? Only by reading through the prose to locate it.
Now consider the same analysis as a structured artifact: a defined schema with fields for market size (with units and date), competitor list (with market share percentages), growth rate (with confidence interval), key risks (enumerated), and data sources (cited). Every piece of information has a defined location, format, and relationship to other pieces.
The structured version is comparable, composable, machine-readable, versionable, and auditable. It took slightly more effort to create — you had to define the schema and ask the AI to populate it — but the return on that effort compounds with every subsequent use.
This is the fundamental argument for cognitive artifacts in AI work: a small upfront investment in structure yields large returns in reusability, reliability, and collaborative value.
What is the trade-off of a structured artifact versus loose prose?
It takes slightly more upfront effort to define the schema, but the return compounds with every subsequent use through reusability and reliability.
Schemas, Templates, and Formats as Cognitive Artifacts
Schemas are perhaps the most powerful type of cognitive artifact because they are generative — they do not just capture one piece of knowledge, they define the shape of an entire category of knowledge. A schema for "meeting notes" determines what information gets captured from every meeting. A schema for "bug report" determines what context accompanies every issue. The schema itself encodes organisational wisdom about what matters.
When you work with AI, schemas serve a dual purpose. First, they constrain the AI's output into a useful form. Instead of receiving a wall of text, you receive data that fits a defined structure. Second, they force you to think clearly about what you actually need. Defining a schema requires you to decide: what fields matter? What format should each take? What is required versus optional? This thinking is valuable independent of the AI.
Templates are schemas with defaults — partially completed artifacts that encode best practices. A meeting notes template that pre-populates "Decisions," "Action items," and "Open questions" sections ensures these categories are addressed every time. A project brief template that includes "Success criteria" and "Known risks" prevents these from being overlooked.
Formats are the most granular level: the specific conventions for representing individual data points. Dates as ISO 8601. Currency with explicit codes. Names in "family, given" order for sorting. These feel like bureaucratic details, but they are the connective tissue that allows cognitive artifacts to be reliably composed, compared, and processed.
In practice: Before asking AI to generate content, define the output format. Before building a knowledge base, design the schemas. Before starting a repeated process, create the templates. These artifacts are the infrastructure of effective augmented intelligence — they make every subsequent interaction more efficient and reliable.
Why is a schema considered the most powerful type of cognitive artifact?
It is generative — it does not capture just one piece of knowledge but defines the shape of an entire category of knowledge.
How does a template differ from a plain schema?
A template is a schema with defaults — a partially completed artifact that encodes best practices, such as pre-populated sections.
Why Structured Output from AI Is More Valuable
There is a straightforward reason why structured AI output outperforms freeform text: structured output is verifiable.
When AI returns prose, verifying it requires reading every sentence and assessing each claim — a slow, cognitively demanding process that most people skip. When AI returns structured data, you can verify it field by field. You can write automated checks. You can compare it against known values. You can spot missing data by looking for empty fields rather than re-reading paragraphs to notice what was omitted.
Structured output is also composable. Individual artifacts can be combined into larger structures. A collection of structured meeting notes can be aggregated into a project timeline. A set of structured analyses can be compared in a table. A library of structured templates can be assembled into a complete workflow. Prose does not compose — it accumulates.
Finally, structured output creates a feedback loop. When AI fills a schema and gets a field wrong, you can identify exactly which field failed and why. This gives you precise information for improving your prompt, adjusting your schema, or flagging a domain where the AI is unreliable. When prose contains an error, you often cannot isolate the cause because the error is entangled with surrounding text.
This does not mean prose is never appropriate. Natural language remains the best interface for exploration, brainstorming, and communication between humans. But when you need AI output that you will build on, share, store, or act on — structure it. The cognitive artifact you create will serve you far better than the paragraph that was easier to produce.
Why is structured AI output easier to verify than prose?
You can check it field by field, write automated checks, and spot missing data as empty fields, rather than re-reading every sentence to assess each claim.
The trap: "Structure is just bureaucratic overhead that slows me down."
It feels faster to type a sentence than to define a schema, so structure looks like a tax on getting started. But that framing only counts the cost of writing the artifact, never the cost of trusting the result. Freeform prose pushes the real work — checking each claim, extracting the data point you need, comparing it to last week's answer — onto every future reader, every time. The structure is not extra work added to the task; it is the verification work done once, up front, in a form the AI and the next person can both reuse. You are not choosing between effort and no effort. You are choosing where the effort lands.
Why isn't defining structure just bureaucratic overhead?
Skipping structure does not remove the work — it pushes the cost of verifying, extracting, and comparing onto every future reader. Structure is that work done once, up front, in a reusable form.
Building a Practice of Cognitive Artifacts
Working with cognitive artifacts is a practice, not a one-time decision. Here is how to develop it:
- Start with your most repeated AI interactions. Identify the tasks you do most often with AI. For each one, design a simple output schema. Use it consistently and refine it based on what works.
- Build a personal template library. Collect the schemas, formats, and prompt templates that work well for you. Store them in your knowledge management system. Over time, this library becomes a powerful productivity tool — each template encodes lessons learned from dozens of past interactions.
- Share and standardise within teams. Cognitive artifacts become more powerful when shared. When a team agrees on a standard schema for a project brief, every brief produced by any team member is immediately comparable and composable. Standardisation is not bureaucracy — it is collective intelligence infrastructure.
- Iterate on structure, not just content. When an AI output is not useful, the problem might not be the AI's capability — it might be the artifact's structure. Revise your schemas based on actual use. Add fields you keep needing. Remove fields you never check. The artifact should evolve with your understanding of the domain.
Try it on your own work
Pick one task you keep handing an AI as freeform text — a summary you request, a draft you ask it to review, an analysis you keep re-running. Turn it into a structured artifact:
- Name the fields you actually act on. Look at the last prose answer you got and ask what you did with it. Each thing you extracted or checked is a field. Write them down — that list is your first schema.
- Demand evidence, not just answers. For every field, add a paired field that forces the AI to show its work: a quote, a source, a confidence level, or "unclear" as an allowed value. This is what makes the output verifiable instead of merely fluent.
- Run the task through the schema and refine it. Ask the AI to fill the structure instead of writing prose. Where a field comes back empty or wrong, you have found exactly what to fix — in the prompt, the schema, or your own assumptions. Keep the version that worked.
Save that schema where you will find it again. The second time you use it costs almost nothing, and it carries forward every lesson the first run taught you.
You've answered 0 recall card(s) on this page. They'll resurface right before you'd forget.
Continue Learning
Cognitive artifacts supply the units of thought. Explore the full learning hub or see how these concepts apply in practice.