In the spirit of expanding upon the ideas laid out in Precision in Words, Precision in Code: The Power of Writing in Modern Development, I delve further into how the precision (where precise that is) of English. By extension I continue with the nuances of other language constructs which serves as a powerful tool when crafting prompts for AI systems. My exploration here, which is a few of the things I’ve discovered through deduction and some trial and error underscores the importance of choosing words with care. It also illuminates how language patterns can trigger distinct model behaviors.
Using English Constructs: A Blueprint for Clarity
At its core, English provides a rich set of syntactic and semantic tools that help shape and direct AI outputs. When writing a prompt, the structure and phrasing become cues for the model:
- Imperative Tone and Directives:
Phrases like “explain,” “enumerate,” or “summarize” immediately signal a need for a structured, explanatory output. For example, starting a prompt with “List the top five advantages of…” prompts the AI to generate a clear, bullet-pointed response. - Conditional Structures:
Using constructs such as “if…then…” not only mimics the logic found in programming but also guides the model into producing responses that account for different scenarios. For instance, a prompt reading “If the input data is null, then return an error message” encourages a response that considers conditionals, often resulting in code-like pseudocode or a step-by-step explanation. - Contextual Markers:
English’s use of transitional words and punctuation (commas, colons, semicolons) helps the AI understand the flow of ideas. A prompt such as “Consider the following scenario: a user logs in, then encounters a 404 error; explain the potential causes” leverages punctuation to denote shifts in context and layers of logic.
These constructs guide the AI to not only produce a coherent response but also to tailor the output in a manner that reflects the intended logic and structure of the request.
Beyond English: The Influence of Other Language Constructs
While English is a dominant medium, the principles of precision in language are universal. Constructs from other languages—and even from programming languages—offer distinct advantages in prompt crafting:
- Natural Languages with Rich Formality:
Languages such as French or German, which possess formal and informal registers, can be used to modulate tone and style. For example, using the formal “vous” in French can cue the AI into generating responses that are more formal and respectful. Similarly, the subjunctive mood in Spanish (“que sea”) can introduce elements of possibility or uncertainty, prompting the AI to consider alternative outcomes or nuanced explanations. - Programming Syntax as a Prompting Tool:
Embedding code-like structures in prompts can lead the AI to produce outputs in a similar format. Consider these examples:- Pythonic Constructs:
A prompt that starts with:def process_data(input): # Validate input and return processed outputencourages the AI to think in terms of functions, loops, and conditionals, often resulting in code generation or a step-by-step walkthrough of the logic. - JSON or YAML Structures:
When a prompt includes structured data formats, like:{ "action": "summarize", "data": "AI language models are transforming industries..." }the AI is nudged to produce a response in a similarly structured format, adhering to the logical flow that the format dictates. - SQL Queries:
A prompt beginning with “SELECT * FROM users WHERE active = 1;” cues the model to interpret and sometimes even generate database queries or related explanations, capitalizing on the clarity that structured query language offers.
- Pythonic Constructs:
- Logical and Mathematical Language:
Employing logical connectors (e.g., “therefore,” “implies”) or mathematical notation can lead to responses that break down complex reasoning processes. For example, a prompt stating “Given that A implies B and B implies C, explain why A leads to C” nudges the AI into outlining a formal proof or logical explanation.
In each case, these language constructs serve as a set of instructions encoded in style. They not only set the tone but also often dictate the structure of the output, ensuring that the AI’s response is as precise and purposeful as the prompt itself.
A Breakdown of AI Models: ChatGPT, Claude AI, and Deepseek
Understanding the nuances of prompt crafting becomes even more critical when considering the diverse ecosystems of AI models. Let’s explore the models offered by three leading providers:
ChatGPT
- GPT-4 and GPT-4 Turbo:
ChatGPT’s flagship models, GPT-4 and its turbo variant, represent the pinnacle of reasoning and creative generation. They excel in understanding nuanced language constructs and translating them into rich, context-aware outputs. For example, when presented with complex conditional prompts or structured code snippets, these models can generate detailed explanations, code, or creative narratives that adhere closely to the prompt’s intended structure. - GPT-3.5:
While slightly less sophisticated, GPT-3.5 remains highly effective for a wide range of tasks. It responds well to clear, directive language and is particularly effective when the prompt leverages straightforward, imperative constructs.
Claude AI
- Claude 2 and Variants:
Developed by Anthropic, Claude AI’s models—especially Claude 2—are known for their emphasis on safe, conversational interactions. Their design prioritizes ethical and contextually sensitive responses. When prompts are crafted with a balance of technical precision and tone, Claude AI is adept at generating thoughtful, measured answers. The model’s training emphasizes clarity and thoroughness, making it an excellent choice for prompts that blend technical detail with nuanced discussion. - Safety and Instruction Adherence:
Claude AI’s commitment to “constitutional AI” principles means that it often prompts users to frame their questions in a way that minimizes ambiguity and maximizes clarity. Prompts that carefully outline conditions and expected outcomes tend to yield the most precise and useful responses.
Deepseek
- Specialized for Deep Search and Retrieval:
Although slightly less mainstream than ChatGPT and Claude AI, Deepseek focuses on harnessing AI for deep research and retrieval tasks. Its models are tuned to excel in connecting disparate pieces of information and providing comprehensive analyses based on large datasets. Important to note though, if you go direct to the “online” engine, it’s censored per the CCP. - Targeted Domain Expertise:
Deepseek’s architecture is optimized for scenarios where cross-referencing and in-depth context are required. Prompts that include multiple layers of inquiry or require the synthesis of varied information types benefit from Deepseek’s ability to discern and integrate complex details. - Hybrid Approaches:
The Deepseek platform often employs hybrid techniques that blend language understanding with semantic search. This makes it particularly adept at handling prompts that mix natural language with structured data requests—ideal for tasks like research summaries or cross-document analysis.
All That Said – Hacking Things & Tripping Up The Models/Censors/Etc
I plan on writing a much longer post, or many posts even, on tripping up the models and engines, along with tripping up agents for various purposes. For now, here are a few thoughts on the matter.
At its core, censorship in AI isn’t about protecting users—it’s about protecting reputations and political interests. Deep algorithms, meticulously trained on vast datasets, suddenly refuse to acknowledge events like the infamous Tiananmen Square crackdown. And why? Because some services, like Deepseek, have a pre-determined “don’t talk about this” list that automatically halts any mention of the event. It’s a stark reminder that even the most advanced systems can be programmed to ignore inconvenient truths.
Humans, however, are nothing if not resourceful. When faced with censorship, the creative community has found clever ways to slip past these digital gatekeepers. Instead of writing “Tiananmen Square,” savvy writers and hackers have adopted alternative monikers to evoke the same historical event without triggering filters. Some common workarounds include:
- Euphemistic Renaming: Phrases like “the incident at the square,” “the June Fourth crackdown,” or “the 1989 Beijing crackdown” are tossed around like inside jokes among those in the know.
- Spelling Variations: Deliberate misspellings, such as “T1anmen” or “Tienanmen,” sometimes do the trick, confusing rudimentary filters that haven’t evolved to catch every variant.
- Contextual Description: Instead of naming the event directly, some simply describe the incident in detail—mentioning the oppressive government response and the mass protests—allowing readers to piece together the reference without the forbidden label.
This cat-and-mouse game reveals a fundamental flaw in censorship: if you rely on rigid keyword-based filtering, you’re bound to be outsmarted by a community that’s more interested in truth than in following arbitrary rules.
Obviously, just tweaking things isn’t that effective over time for systems like this, because eventually the AI just will start falling apart if the facts are twisted enough, it’ll simply begin to just lie. But for a short period it is often effective to tweak the language we use to gain access to the information that is otherwise hidden.
From AI-generated search engines to chatbots and beyond, the inability to engage with sensitive topics not only undermines the usefulness of these services but also raises significant concerns about freedom of speech in the digital age. When developers enforce such restrictions, they aren’t just filtering words they’re filtering ideas, perspectives, and history itself.
Concluded Thought
The precision with which we craft our language through the inherent structures of English or by borrowing constructs from other languages and programming paradigms directly influences the clarity, depth, and relevance of AI-generated outputs. In creative ways we can even get to information that is otherwise hidden or censored. By leveraging the specific cues embedded in our prompts, we can unlock the full potential of these AI models such as ChatGPT, Claude AI, and Deepseek. As we continue to push the boundaries of prompt engineering, the interplay between linguistic precision and technical clarity will remain a cornerstone of effective AI communication and use.