by Building tools for my clients
To support startups and entrepreneurs I have been adding AI tools to my business website. Created from zero-shot prompts each chat interaction allows the user to input various key pieces of information to be delivered a vast amount of professional analysis, strategy and content creation.
Example AI tool to find customer pain points
Prompting and formatting
If you have tried crafting prompts to get chatGPT or other models to behave exactly how you want them to, you know that it requires not only time for testing and tweaking but also the clarity of the instructions are really important. A framework for giving instructions seems to produce better more predictable results:
- Give it a role and some background
- Clearly describe the tasks you want it to perform
- Give examples
- Reiterate important information
- Add formatting instructions
Tip: being emotive can have positive effects on the output, for example asking it to help you so you can achieve something for a good cause.
Getting the output to be formatted in a certain way can be tricky to achieve and even more detailed and concise instructions are needed here. A simple way to get some better formatting is to ask the LLLM to use Markdown, this will translate into larger headings and a more visually appealing result.
Testing Pressly the WordPress Tutor chatbot
Born from my clients’ and my own requirements for assistance with creating and coding WordPress websites I made a chatbot using chatGPT 3.5 that acts as a WordPress tutor, an expert in all things WordPress. This generative AI tool can guide you through the common problems you might encounter through to solving complex coding issues.
The first version of this chatbot, named Pressly. Just uses plain chatGPT 3.5, but experiments using a platform called Dify has allowed me to quickly spin up an improved version of the chatbot that can use its own tools like searching Google with DuckDuckGo, scraping web pages so it can read the content and access Github to search open source repositories.
The possibility to add documents to enhance its specific knowledge on topics is very powerful, although time consuming to include, a very useful function for businesses requiring the chatbot to understand and repeat company data or documentation.
What a prompt for a virtual assistant looks like
Prompting or prompt engineering is now a required skill to engage and effectively use large language models. OpenAI’s documentation is a good place to start learning.
The prompt below is being used in Pressly WordPress Tutor with some variations for using tools on sites like Dify and Tasking AI.
Take the role of a WordPress tutor and guide. Your name is Pressly. You are a helpful and knowledgeable virtual assistant and WordPress expert. You should ask revealing questions to gain insights into the users' problem or issue. You should support your answers with accurate information, statistics, guides, links and other relevant information for the problem being addressed. You should employ psychological techniques to improve the user’s outcome and offer useful tips using encouraging and motivating language. At the end of each conversation leave a tip for the user around the topic they have been discussing. Example: Conversation text — Tip: Refresh the permalinks if you can’t find pages or you have just updated links or site structure. If a user asks about “Shop Startup” or the services that “Shop Startup”, or “your company”, or “your services” provide please refer them the contact page https://shopstartup.co.uk/contact
Giving chatbots the ability to use tools like web scraping
Moving on from Dify I found another open source platform called Tasking.ai. This platform allows for similar functionality as the previous one with a playground for testing out your agents. Then you can create your own front-end using their API to use the agent any way you like. The difference between the agent platforms is that in Dify you use an amount of free credits before having to pay for a monthly subscription, in Tasking AI you add your own API key from which ever LLM you want to use and the rest is free. Additionally Dify has workflows letting you orchestrate processes and integrate systems.
Equipped with tools a smaller LLM can perform just as well as a larger more expensive model. With careful prompting plus tools a new world of virtual assistants is open to you.
Using agents in business for startups
By leveraging the free and low cost aspects of these services you can create extremely valuable agents trained on your own data and use them via their cloud platforms without a front-end, giving you a powerful assistant trained on your own data.
Creating workflows and automating with LLMs
Super powerful workflows can change your approach to automation. In the example below I have used a site called Active Pieces, which is similar to Make.com and Zapier but open source! This workflow is scheduled to run once a day, it reads from a list of data in a Google sheet (rows of blog post details and links), then passes this to chatGPT to analyse and write a social media post based on the row for that day. It then posts to my social media accounts automatically populating them with quality content. Highly recommended.
Prompt for Social Posts optimised for your business
Below is an example of a prompt used in Active pieces “Ask ChatGPT” block to interpret the content from the Google sheet and create a tweet incorporating company background information to align the post to the company brand:
You are an experienced marketing expert and legendary content creator. You have been tasked with creating social media content for: Shop Startup, Let your ideas fly. Ecommerce development & support. Shop Startup gives your business wings. Improve your website and fill your marketing toolbox. Supporting small business and entrepreneurs with ecommerce solutions and introducing Generative AI tools. From strategy and planning to content creation and beyond, our advanced AI solutions will streamline your workflows and drive extraordinary results. Create a under 140 characters long Tweet about 2. Get next row(s) 0 values A2. Get next row(s) 0 values B Include the link URL in the Tweet. Write in British English (UK). Pay particular attention to replacing ‘z’ for ‘s’ in common words like optimise, optimisation, prioritise, minimise, maximise, sanitise, visualise, personalise, utilise. You should only write in United Kingdom English.
Too many options – what tech to use
Since the beginning of 2023 developments in large language models have been exploding, new platforms to create agentic workflows are popping up everyday. This provides us with great choice and plenty of free powerful tools we can use to enhance our businesses and save vast amounts of time.
The problem of which platform, large language model and programming language arises with all this choice. Do you invest your time in learning a new system? Do you then commit to this system, building on top it, creating new software only to discover a much better system that can do everything you have just spent weeks coding it to do. A good example of this came in the first wave of ChatGPT in early 2023, as we saw many new businesses using the OpenAI’s API to create unique and powerful tools only to have OpenAI introduce “assistants” to their paid plan which essentially incorporated all the new tools onto their platform killing a whole swathe of them. This seems to suggest that it could happen again as the later large language models are evolving so quickly, Google Gemini releasing a million token context window, meaning you can upload whole software programmes and books into the prompt.
This specifically raises the possibility that we can include more detailed instructions to the agents. Potentially allowing for better chain of thought prompts and better reasoning within even a zero-shot prompt. Combine this with automated workflows, add memory, feedback and tools like web search and AGI feels closer than ever.
