Generative AI is powerful, but anyone who has used it for more than a quick answer knows the pitfalls: it can hallucinate facts, lose the thread in long conversations, or give outputs that are shallow or off-target. On top of that, organizations have to consider privacy, intellectual property, and ethics—what data the model can see, who owns the outputs, and how to reduce harmful bias. The problem we’re solving for is steering: helping models stay accurate, relevant, safe, and aligned with human goals. Early on, the instinct was to fine-tune models directly, but today we have a richer set of tools that let us guide behavior more flexibly.

Fine-tuning: rewiring the brain

In the beginning, the default method for customization was fine-tuning. This meant retraining the model’s weights with specialized examples until it “spoke” like your domain or organization. While powerful, it ended up being costly, brittle, and static—more like building a machine that only knows one trick than teaching a student who can keep learning and adapting. In other words, once fine-tuned, the model is stuck with what you gave it, and updating it means starting over. Imagine taking a generalist AI, and training it only on sock-related info—eventually, the overabundance of sock info will displace the more generalized information.

Today, we’ve mostly moved away from fine-tuning an entire model because it locks you into a static version of a fast-moving field. Instead, lighter, more dynamic methods are usually better. Still, fine-tuning can be the right choice in scenarios where compliance requires outputs in a very specific format, where internet access isn’t possible, or where latency must be minimized by baking knowledge directly into the model.

Pros: Deep customization, especially for highly specialized domains which don’t change a lot.
Cons: Expensive, slow to update, and risks erasing general abilities.
Think about: Domain adaptation, catastrophic forgetting, and parameter-efficient fine-tuning.

A female professional trains a robot at a whiteboard, where she is writing clearly-numbered instructions for the robot.

Prompts and instructions: making clear requests

At the simplest level, you can just ask differently. Prompt design and system instructions steer the model toward certain behaviors. It’s like giving your team member a job description before they start. Clear instructions—“always use formal tone,” “answer as if you were a travel agent”—can shape output dramatically without changing the model at all.

Beyond single prompts, you can chain prompts together into a sequence or multi-step process. For example, one prompt might extract key facts, another might analyze them, and a third might format the final answer. This multi-step approach makes outputs more reliable and mirrors how people break work into phases rather than doing it all at once.

Pros: Fast and flexible, no retraining needed; chaining improves reliability.
Cons: Fragile; small changes in wording can lead to big swings in results; multi-step chains require discipline and testing.
Think about: Prompt libraries, system message design, instruction-tuned models, and prompt chaining workflows.

A black female librarian helps a robot find the perfect book.

Retrieval-augmented generation (RAG): the librarian whispering in their ear

Instead of forcing the team member to memorize every fact, you can give them access to a reference library. Retrieval-Augmented Generation (RAG) lets the model pull in documents, databases, or knowledge graphs at runtime. Imagine your team member always has a librarian who fetches the right file before they answer. This keeps answers current, reduces hallucination, and preserves privacy when you control what sources they can access.

Pros: Keeps answers accurate and up-to-date.
Cons: Dependent on the quality and freshness of the external knowledge base.
Think about: Vector databases, embeddings, and data governance.

A robot fits an interchangeable camera lens to one of their eyes.

Lightweight fine-tuning: changeable lenses

Sometimes you still want the model to adopt a very specific style or domain fluency. Lightweight fine-tuning methods like LoRA let you add small “adapters” without retraining everything. It’s like mounting a special-purpose lens onto a camera: you still use the same body, but the view is customized. This is useful for brand voice, compliance language, or niche vocabularies.

Pros: Customizes behavior cost-effectively.
Cons: Still requires technical know-how and may lock outputs into a narrow style.
Think about: LoRA, adapters, model checkpoints, and compliance use cases.

A team of colorful robots, each with their own specialized tool (a calculator, a paintbrush, a typewriter) sit around a conference table.

Orchestration and agents: a whole team, not just one machine worker

Why make one model do everything? Orchestration frameworks coordinate multiple models and tools—summarizers, reasoners, calculators, databases. Instead of a single machine worker, it's more like a team of junior coworkers—managed by a leader who decides who does what. It’s more resilient and more scalable, though it adds complexity.

Pros: Specialization across tasks improves reliability.
Cons: Complex to design, monitor, and debug. Complete automation requires a lot of guardrails and testing to be trust-able.
Think about: Agent frameworks, Model Context Protocols (MCPs), workflow automation.

A robot walks down a fence-lined path, encountering a fork in the road. A red arrow points one way, and a green arrow points the other way.

Guardrails: rules, guidelines and decision principles

Just as an office sets rules for what team members can and cannot do, guardrails put boundaries around AI behavior. They can block sensitive topics, enforce safety standards, or ensure outputs align with company policy. Think of them as HR guidelines: they don’t make the team member smarter, but they prevent costly mistakes.

Pros: Improves safety, trust, and compliance.
Cons: Can over-restrict or block legitimate use cases if poorly designed.
Think about: Content filters, red-teaming, bias audits, and governance frameworks.

A robot complete a paper form, referring to their handwritten notes, surrounded by sticky notes.

Memory and long context: a collection of notes

Extended context windows and memory layers help AI “remember” across longer conversations or sessions. Instead of rewriting the model with every new fact, you hand the team member a binder of notes they can consult anytime. This is cheaper and more flexible than fine-tuning, though it requires careful curation to avoid information overload.

Pros: Enables continuity and personalization.
Cons: Risk of clutter, outdated notes, or sensitive data leakage.
Think about: Extended context windows, session memory, and secure data handling.

A conveyor belt of chat bubbles, green and red. Humans and robots sort them together.

Pre- and post-processing: filtering what goes into or comes out of the generative AI model

Another way to steer AI is by controlling inputs before they reach the model, and outputs before they reach the user. Pre-processing can strip out sensitive data like personally identifiable information (PII) or filter out hate speech before it ever enters the model. Post-processing can tag or flag certain types of content—for example, highlighting factual claims so a human or another system can verify them.

Pros: Reduces risks without retraining the model.
Cons: Adds overhead and may miss edge cases if filters are poorly designed.
Think about: Input sanitization, moderation APIs, truth-checking workflows, and layered review systems.

Putting it all together:

Today, steering generative AI isn’t about finding a single method. It’s about creating an ecosystem:

Prompts for flexibility.
RAG for reliable knowledge.
Fine-tuning for narrow cases.
Orchestration and agents for complex workflows.
Guardrails for safety.
Memory for continuity.
Pre- and post-processing for filtering and verification.

Each plays a role. Together, they turn a raw model into something useful, safe, and aligned with human goals—less like a wild brain and more like a well-run workplace.

A lightweight fine-tuning technique that adds small modules to a model rather than retraining the whole thing.

"We used tools like Adam and LoRA to create company-specific adapters for general models, so we don't have to retrain an entire model or create our own."

A program or system powered by artificial intelligence (AI) designed to perform tasks autonomously by perceiving its environment, processing data, and making decisions to achieve specific goals. Examples include virtual assistants like chatbots, recommendation engines, and AI-driven workflow managers.

"An AI agent powered by ChatGPT could plan a two-week trip to South Africa, handling everything from booking flights and accommodations to a visit to Nelson Mandela's former home, all while updating the user in real-time through natural language conversations."

When a fine-tuned model loses important general knowledge it had before.

"After we fine-tuned the model too narrowly, it forgot how to handle basic grammar—classic catastrophic forgetting."

The amount of information a model can “hold in mind” at one time.

"Because the context window was limited, the chatbot forgot details from earlier in the conversation."

Numerical representations of text or other data that make it easy to compare similarity.

"We used embeddings to quickly find documents related to the customer’s support ticket."

A control or safeguard placed around AI systems to limit harmful or undesired outputs.

"The platform added AI guardrails to block inappropriate language in customer chats."

In the context of AI, a hallucination refers to a situation where an artificial intelligence system generates or perceives information that is not based on real or accurate data. It is an erroneous or false perception or output produced by the AI system.

"ChatGPT thought that in addition to owning Causeit, Inc. that I am a screenwriter for Marvel movies. So I said, 'where're my royalty payments?'"

A structured way of linking facts and entities to show relationships.

"As we built our movie chatbot, we created a knowledge graph of all the movies available at our library, cross-referenced in many different ways, like what kind of story they told."

A standard that lets AI models talk to external tools, APIs, or other models in a safe, structured way—giving them a plug-and-play connector to extend their abilities without retraining.

"Using the MCP, the chatbot pulled information directly from the CRM system without needing retraining."

A fine-tuning method for generative AI models, like Adam or LoRA, that fine-tune only parts of a model.

"PEFT let us adjust the model for legal terminology without changing the underlying model, for a lot less compute cost and review time."

The process of adjusting a model’s parameters and/or underlying data to minimize errors or maximize performance on a task.

"Optimizing our model by removing out-of-date map information reduced errors in predicting delivery times."

Data that can identify an individual, such as names or phone numbers.

"We stripped PII like phone numbers before feeding customer chats to the AI."

Sequencing a number of prompts for an AI tool, like a generative AI chat system, in order to accomplish a given goal.

"To get this to work, we need to do some basic prompt-chaining—first, feed it our source news article and then ask the tool to 'rewrite the article in a more casual tone' and then 'create a metaphor for the technology shift mentioned' and then 'translate into Spanish.'"

Prompt engineering (also known as prompt crafting or prompt design) refers to the process of designing and refining prompts or instructions given to an AI language model to get the desired responses (such as by adding context or specifying outut formats).

"We had to do some prompt engineering with specific instructions to make sure ChatGPT returned consistent summaries of our articles."

A strategy for 'grounding' AI tools in accurate facts by combining generative and retrieval models. RAG uses a pre-trained model to retrieve relevant information from vetted and verified documents (a knowledge graph) and then uses a generative model to output content based in fact.

"Our chatbot was confidently giving incorrect answers to our patients. Since we're in medicine, that's not acceptable, even for casual use, so we used a RAG strategy so that the chatbot retrieve facts from medical texts but then makes them easier to read for lay users."

A simplified mathematical fingerprint of information (like a word, image, or sound) that a computer can compare with others to find similarities.

“We converted product descriptions into vectors so the AI could recommend similar items to shoppers.”

Humans who deliberately attempt to trick or negatively influence a machine system to increase its quality, accuracy, security, or consistency. Red teams in AI work to lessen the number of offensive, inaccurate, biased, and/or undesirable results users experience. In cybersecurity, red teams are sometimes part of 'white hat' hacking for testing.

"The red team attempted to trick the AI tool into saying racist things so that developers could spot ways to prevent the model from coming to those false conclusions."

Being able to fine-tune, adjust, correct, or otherwise guide an AI system to operate more in line with the expectations and ethics of its owner or user. Traditional analytical algorithms are generally easier to steer—and explain—than machine learning and generative AIs.

"While we've had success steering AI for content recommendations to our users when it was based on tags, we're concerned about adding generative AI to that because it will be harder to adjust."

The practice of ensuring that generative AI tools return results that are accurate ('grounded' in facts) rather than just those which are statistically probable or pleasing to a user.

"OpenAI's ChatGPT can now provide citations of its sources so that users can understand the context of its answers."

Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla at volutpat diam ut venenatis tellus in. Nunc sed velit dignissim sodales ut eu sem integer vitae. Sagittis nisl rhoncus mattis rhoncus urna neque viverra justo. Donec et odio pellentesque diam volutpat commodo. Platea dictumst quisque sagittis purus sit amet volutpat consequat mauris. Fringilla est ullamcorper eget nulla facilisi etiam dignissim. Et tortor at risus viverra adipiscing at in tellus integer. Proin sagittis nisl rhoncus mattis rhoncus urna neque. Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.

Banner

Inline Block

CaptionNunc congue nisi vitae suscipit tellus. Sagittis nisl rhoncus mattis rhoncus urna neque viverra. Tempus quam pellentesque nec nam. Egestas fringilla phasellus faucibus scelerisque eleifend donec pretium vulputate sapien. Gravida

‍

Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Mauris sit amet massa vitae. Sed felis eget velit aliquet sagittis. Sem integer vitae justo eget magna fermentum. Nunc congue nisi vitae suscipit tellus. Sagittis nisl rhoncus mattis rhoncus urna neque viverra. Tempus quam pellentesque nec nam. Egestas fringilla phasellus faucibus scelerisque eleifend donec pretium vulputate sapien. Gravida in fermentum et sollicitudin ac orci. Ultricies mi eget mauris pharetra et. Maecenas accumsan lacus vel facilisis volutpat. Mauris pharetra et ultrices neque ornare aenean. Urna cursus eget nunc scelerisque viverra. Mi proin sed libero enim sed faucibus turpis in. Lobortis feugiat vivamus at augue eget arcu dictum.

Left inline no wrap

Left Wrap (after first ¶)

Luctus accumsan tortor posuere ac ut consequat. Tristique senectus et netus et malesuada fames. Magna ac placerat vestibulum lectus mauris ultrices eros. Faucibus a pellentesque sit amet porttitor eget dolor morbi. Mi in nulla posuere sollicitudin aliquam ultrices sagittis orci. Mus mauris vitae ultricies leo integer. Morbi tristique senectus et netus et. Commodo sed egestas egestas fringilla. Congue nisi vitae suscipit tellus mauris a diam maecenas sed. In pellentesque massa placerat duis ultricies. Nulla facilisi cras fermentum odio eu feugiat pretium. Neque gravida in fermentum et sollicitudin ac orci phasellus. Elementum sagittis vitae et leo duis. Dignissim cras tincidunt lobortis feugiat.

Right Wrap (after first ¶)

Style Tests

Bold Text looks like this.

Italic text looks like this.

Ordered lists

Causeit offers AI & Digital Fluency solutions to clients in the form of ‘mindset transformation’ packages consisting of content and live time combined to bring new thinking into existing firms. Current and near-future offerings include:

Level one
1. Level two
2. Level two Sem viverra aliquet eget sit amet. Enim neque volutpat ac tincidunt vitae semper. Duis ut diam quam nulla porttitor massa id. Enim sit amet venenatis urna. Lacus vestibulum sed arcu non odio euismod lacinia. Et malesuada fames ac turpis egestas integer. Nulla facilisi cras fermentum odio eu feugiat. Ac turpis egestas integer eget aliquet nibh praesent tristique magna. Mattis rhoncus urna neque viverra justo. Urna nec tincidunt praesent semper feugiat. Magna etiam tempor orci eu lobortis elementum nibh.
  1. Level three
  2. Level three
    1. Level four
    2. Level four
      1. Level five
      2. Level five
      3. Sem viverra aliquet eget sit amet. Enim neque volutpat ac tincidunt vitae semper. Duis ut diam quam nulla porttitor massa id. Enim sit amet venenatis urna. Lacus vestibulum sed arcu non odio euismod lacinia. Et malesuada fames ac turpis egestas integer. Nulla facilisi cras fermentum odio eu feugiat. Ac turpis egestas integer eget aliquet nibh praesent tristique magna. Mattis rhoncus urna neque viverra justo. Urna nec tincidunt praesent semper feugiat. Magna etiam tempor orci eu lobortis elementum nibh.
      4. ‍
3. Level two
Level one
1. ‍

Unordered lists

Level one Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvinar
Level one Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvinar
- Level two Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvina
- Level two Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvina
  - Level three Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvina
    - Level four Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvina
      - Level five
        Level two Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvina.

‍

Text with ^{a superscript}Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla at volutpat diam ut venenatis tellus in. Nunc sed velit dignissim sodales ut eu sem integer vitae. Sagittis nisl rhoncus mattis rhoncus urna neque viverra justo. Donec et odio pellentesque diam volutpat commodo. Platea dictumst quisque sagittis purus sit amet volutpat consequat mauris. Fringilla est ullamcorper eget nulla facilisi etiam dignissim. Et tortor at risus viverra adipiscing at in tellus integer. Proin sagittis nisl rhoncus mattis rhoncus urna neque. Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla

^‍and_{text with a subscript}Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla at volutpat diam ut venenatis tellus in. Nunc sed velit dignissim sodales ut eu sem integer vitae. Sagittis nisl rhoncus mattis rhoncus urna neque viverra justo. Donec et odio pellentesque diam volutpat commodo. Platea dictumst quisque sagittis purus sit amet volutpat consequat mauris. Fringilla est ullamcorper eget nulla facilisi etiam dignissim. Et tortor at risus viverra adipiscing at in tellus integer. Proin sagittis nisl rhoncus mattis rhoncus urna neque. Vitae sapien pellentesque habitant morbi tristique senectus et netus. At imperdiet dui accumsan sit amet nulla facilisi. Lacus vel facilisis volutpat est velit egestas dui id ornare. Aliquam purus sit amet luctus venenatis.Pulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. NullaPulvinar mattis nunc sed blandit libero. Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla

and text with a link

Quote Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla at volutpat diam ut venenatis tellus in. Nunc sed velit dignissim sodales ut eu sem integer vitae. Sagittis nisl rhoncus mattis rhoncus urna neque viverra justo. Donec et odio pellentesque diam volutpat commodo. Platea dictumst quisque sagittis purus sit amet volutpat consequat mauris. Fringilla est ullamcorper eget nulla facilisi etiam dignissim. Et tortor at risus viverra adipiscing at in tellus integer. Proin sagittis nisl rhoncus mattis rhoncus urna neque. Vitae sapien pellentesque

Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla at volutpat diam ut venenatis tellus in. Nunc sed velit dignissim sodales ut eu sem integer vitae. Sagittis nisl rhoncus mattis rhoncus urna neque viverra justo. Donec et odio pellentesque diam volutpat commodo. Platea dictumst quisque sagittis purus sit amet volutpat consequat mauris. Fringilla est ullamcorper eget nulla facilisi etiam dignissim. Et tortor at risus viverra adipiscing at in tellus integer. Proin sagittis nisl rhoncus mattis rhoncus urna neque. Vitae sapien pellentesque

Code Block Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla at volutpat diam ut venenatis tellus in. Nunc sed velit dignissim sodales ut eu sem integer vitae. Sagittis nisl rhoncus mattis rhoncus urna neque viverra justo. Donec et odio pellentesque diam volutpat commodo. Platea dictumst quisque sagittis purus sit amet volutpat consequat mauris. Fringilla est ullamcorper eget nulla facilisi etiam dignissim. Et tortor at risus viverra adipiscing at in tellus integer. Proin sagittis nisl rhoncus mattis rhoncus urna neque. Vitae sapien pellentesque

Rich media:

Video: Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla at volutpat diam ut venenatis tellus in. Nunc sed velit dignissim sodales ut eu sem integer vitae. Sagittis nisl rhoncus mattis rhoncus urna neque viverra justo. Donec et odio pellentesque diam volutpat commodo. Platea dictumst quisque sagittis purus sit amet volutpat consequat mauris. Fringilla est ullamcorper eget nulla facilisi etiam dignissim. Et tortor at risus viverra adipiscing at in tellus integer. Proin sagittis nisl rhoncus mattis rhoncus urna neque. Vitae sapien pellentesque

Banner

Inline Block Center

Inline Block Left

Wrap Left

‍

Wrap Right

Size Override Center Full Width (does not work)

Embed: Porttitor rhoncus dolor purus non enim praesent elementum facilisis leo. Nulla at volutpat diam ut venenatis tellus in. Nunc sed velit dignissim sodales ut eu sem integer vitae. Sagittis nisl rhoncus mattis rhoncus urna neque viverra justo. Donec et odio pellentesque diam volutpat commodo. Platea dictumst quisque sagittis purus sit amet volutpat consequat mauris. Fringilla est ullamcorper eget nulla facilisi etiam dignissim. Et tortor at risus viverra adipiscing at in tellus integer. Proin sagittis nisl rhoncus mattis rhoncus urna neque. Vitae sapien pellentesque

‍

Heading 1

Heading 2

Heading 3

Heading 4

Heading 5

Heading 6