Anapoly AI Labs

Tag: ai tools

  • The pace is accelerating

    Gas Town would have been science fiction 18 months ago.

    Like so much of that genre, it has become reality. Steve Yegge created a real system architecture that has demonstrated it is capable of directing, harnessing, and controlling the software development work of 30 agents. The human overseer has 30 terminal windows open to monitor what is happening and enters commands at great speed throughout the process; hence his advice not to try it unless you are a top echelon software engineer not prone to insanity.

    But although it is chaotic, it works, and we can be sure that this demonstrator (the fourth version Yegge has produced over the past year) will soon evolve into a better version of a multi-agent software factory.

    As for the crazy Moltbook experiment, that uses OpenClaw. This is an open-source platform allowing users to run autonomous AI agents locally that integrate with messaging apps to execute tasks and interact with other agents on a network.

    Google just released Universal Commerce Protocol UCP), an open-source standard to power agentic commerce, providing a unified language for AI agents to discover products and execute secure transactions with businesses.

    Both OpenClaw and UCP utilize the Model Context Protocol (MCP) to enable agents to do things in the real world. OpenClaw agents can use MCP-based skills to potentially transact with merchants adopting the UCP standard.

    This is science fiction becoming reality before our eyes.

  • An AI power-user’s perspective

    Transparency label: AI-assisted

    In his YouTube channel, Dylan Davis shares his thoughts on how we can make good use of AI. This infographic (produced by NotebookLM using only Dylan’s video as its source) shows five ways power users do so.

    Importantly, the fifth habit he highlights is not letting AI do all the thinking for us. If we delegate all our reading and summarising to AI, we risk losing the ability to focus for extended periods. We need to read a sufficient number of long-form documents ourselves, in order to retain that abiliy. Think of attention span as a muscle: we must exercise it by doing some hard work, or it will wither away.

    In another video, Dylan explains how we can create a GPT (ChatGPT) or a Gem (Gemini)to generate professional Nano Banana Pro image prompts. I used NotebookLM to write a standard operating procedure for this, based entirely on source material in the video. Users act as Creative Directors providing context, while the Gem serves as a Technical Prompter, transforming inputs into optimized JSON code to ensure consistent, high-quality visual outputs from Nano Banana Pro.

    JSON (JavaScript Object Notation): a lightweight data-interchange format based on a subset of the JavaScript Programming Language. It is easy for humans to read and write. It is easy for machines to parse and generate.

  • Banana Nano Pro prompt generator

    Banana Nano Pro Prompt Generator

    Source material created by Dylan Davis and made freely available through his YouTube channel. Accessed: 17 December 2025

    Transparency label: AI-generated

    Objective: To consistently generate professional-grade image prompts for Nano Banana Pro by outsourcing the technical prompting to a custom AI “Gem”.

    Role: The User acts as the “Creative Director” (providing intent), while the Gem acts as the “Technical Prompter” (formatting for the model).

    Tools Required

    • Google Gemini Advanced (to create and use Gems)
    • Nano Banana Pro (Image Generation Model)
    • Reference Images (Optional but recommended)

    Phase 1: One-Time Setup (Creating the Gem)

    Note: This step only needs to be done once to build the tool.

    1. Create a New Gem: Open Gemini and start a new Gem (custom AI assistant).
    2. Define the Instructions: In the instructions section for the Gem, paste the content of file 2025-12-16_Prompt_Base-Prompt-for-Nano-Banana-Pro-Prompt. The file contains directives that cover the following logic derived from the source:
      • Role Definition: You are an expert at prompting the Nano Banana Pro image model.
      • Task: Your job is to take five specific user inputs and synthesize them into high-quality image prompts.
      • Input Guardrails: If the user does not provide all five inputs (Purpose, Audience, Subject, Brand Rules, Reference Image), ask them to provide the missing information before generating prompts.
      • Output Styles: Generate three distinct prompt variations:
        • Literal
        • Creative
        • Premium
      • Output Format: The final output must be in JSON format (using labels and descriptions within brackets). This structure resonates most effectively with Nano Banana Pro.
      • Research Logic: Use internal reasoning to infer aspect ratios and styles based on the “Purpose” provided (e.g., if the purpose is “Instagram,” infer a vertical aspect ratio).

    Phase 2: Input Data (The 5-Step Context)

    When you are ready to generate an image, open your custom Gem and provide the following five inputs. You do not need to worry about technical jargon; just use natural language.

    1. Purpose: Why are you creating this image?
      • Examples: Instagram ad, YouTube thumbnail, Pitch deck, Landing page hero section.
      • Why: This dictates aspect ratio and composition.
    2. Audience: Who is this image for?
      • Examples: Luxury buyers, Tech executives, Gen Z consumers.
    3. Subject: What is the primary focus?
      • Examples: A matte black ceramic mug, a specific person, a character.
    4. Brand Rules: What are the aesthetic guidelines?
      • Examples: Warm tones, minimal styling, specific brand colors, or fonts.
    5. Reference Image (Yes/No): Are you providing a reference image to the image model?
      • Context: Mention if you are inserting a product into a scene, using a sketch, or mimicking a competitor’s ad composition.

    Phase 3: Generation & Execution

    1. Run the Gem: Submit the five inputs. The Gem will “think” (reason) and output three blocks of code (Prompts A, B, and C) in JSON format.
    2. Copy the Prompt: Select “Copy Code” on the variation you prefer (e.g., Prompt A).
    3. Paste into Nano Banana Pro: Paste the copied JSON prompt into the Nano Banana Pro input field.
    4. Upload Reference Image: If you specified a reference image in Phase 2 (e.g., a photo of your product), upload it now. Nano Banana Pro excels at taking existing references and inserting them into new scenes.
    5. Generate: Run the model to produce the initial images.

    Phase 4: Refinement & Editing

    Do not start from scratch if the image is close to perfect. Nano Banana Pro allows for high-quality minor edits without re-rolling the prompt.

    1. Analyze the Result: If the image is 90% there but needs tweaks (e.g., background is wrong, text is misplaced), keep the image open.
    2. Conversational Edits: Type instructions like a creative director to adjust specific elements.
      • Example: “Make the background warmer,” “Move the text to the left,” or “Zoom out and place this mug in a coffee shop”.
    3. Final Polish (Watermarks): If the model leaves a watermark:
      • Download the image.
      • Upload it to a tool like Canva.
      • Use the “Magic Eraser” tool to remove the watermark.

    Summary checklist for every run:

    • Did I provide all 5 inputs (Purpose, Audience, Subject, Brand, Reference)?
    • Did I copy the JSON output?
    • Did I attach my reference image in Nano Banana Pro?

  • Methodology & tool-kits

    Transparancy label:human-only

    Yesterday’s discussion crystallised some ideas that I have been teasing at recently. Namely that we could develop our labs as a place where we:

    (a) develop a methodology and its associated tool-kits for the implementation of AI in an enterprise; and

    (b) use these in a synthetic environment to help business people gain confidence in moving forward with AI.

    Our value to clients would derive from the relevance and usefulness of the methodology and tool-kits.

    Methodology: a comprehensive system of principles, processes, and rules that governs a specific discipline to ensure consistency and repeatability. It serves as a structural framework defining what must be done and who is responsible. It is applied through specialized tool-kits consisting of templates, software, and techniques that enable practitioners to apply the theoretical standards to real-world tasks.

    To pick up on the point Dennis emphasised, an early part of the methodology might be analysis of how an enterprise actually works (rather than what the non-existent, incomplete or out of date documentation says). The process for this would involve eliciting information from people in the enterprise. A tool-kit would enable AI to be configured for that purpose within the specific context of that enterprise. 

  • Using local-only AI in a micro-enterprise

    Transparency label: AI-assisted

    I’ve added a briefing note in the Resources area that sets out an approach for a micro-enterprise to run AI entirely on local machines. This creates a system that is private, predictable, comparatively inexpensive, and easy to expand in small steps.

    The note explains how a three-layer model for controlling the behaviour of an AI ( business context at the top, governance rules in the middle, and task instructions at the bottom) can still apply even when everything sits on a laptop, and how a small software component (the orchestrator) can keep the whole arrangement predictable and safe.

    If you’re curious about “local-only AI” for a micro-enterprise, or wondering what it might look like in practice, you should find this a useful starting point.

    Read the note: Using Local-Only AI in a Micro-Enterprise

  • Precision Content Prompt Pack

    Precision Content Prompt Pack

    Version: 01, 1 August 2025
    Authors: Alec Fearon and ChatGPT-4o.

    Transparency label: AI-assisted

    Purpose

    This is a six-step process for converting a source document into a structured format that is easy for an LLM to understand. The format is based on the Darwin Information Typing Architecture (DITA) and ideas developed by Precision Content. It has the following content types:

    • Reference (what something is)
    • Concept (how to think about it)
    • Principle (why it works)
    • Process (how it unfolds)
    • Task (how to do it)

    The six steps are carried out in sequence, one at a time. They clean, segment, type, clarify, and re-package the original source material. There is human review at the end of each step.

    To use

    First, open a chat in ChatGPT and upload this file into the chat; it is in Markdown (.md) because that is easy for LLMs to read.
    Note: you can drag & drop the file into the chat or use the upload button.

    Tell ChatGPT: “Please create a new canvas from this markdown file so we can work together using the precision content prompt pack.” ChatGPT will:

    • Read the file
    • Create a canvas
    • Use each ## heading to split the file into separate cards
    • Preserve formatting and headings

    Then upload the source file into the chat. Tell ChatGPT: “Please convert the uploaded file [filename] into precision content using the steps defined in the canvas Anapoly AI Labs Precision Content Prompt Pack. Begin Step 0.”

    ChatGPT will extract the content of the file and clean it as per Step 0 – Pre-Processing. It will paste the cleaned material into the “📄 Source Document” card for you to review. That sets you up to proceed with the following steps. The output of each step is put into the “Work Area – Output by Step” card in the canvas. Edit the output of each step as necessary before proceeding to the next step.

    The final output is put into the card “Review Notes / Final Output / Glossary”. You can tell ChatGPT to export it from there as a file for download. If it is to be used as reference material, filetype .md is recommended.

    Step 0 – Pre-Processing

    Purpose: Clean the raw input before analysis.

    Prompt:

    Clean the following document for structured analysis. Remove:

    • Repeated headers/footers
    • Navigation links, timestamps, metadata
    • Formatting glitches (e.g. broken paragraphs)

    Retain all meaningful content exactly as written. Do not summarise, interpret, or reword.

    Step 1 – Segmenting the Document

    Purpose: Divide into discrete, meaningful segments.

    Prompt:

    Break this cleaned document into a numbered list of coherent segments. Each segment should reflect a single topic, paragraph, or unit of meaning.

    Format:
    [1] [text]
    [2] [text]

    Step 2 – Typing the Segments

    Purpose: Label each segment by information type.

    Types:

    • Reference – what something is
    • Concept – how to think about it
    • Principle – why it works
    • Process – how it unfolds
    • Task – how to do it

    Prompt:

    For each segment, assign the most relevant type. Include a short justification.

    Format:
    [1] Type: [type] – [reason]

    Step 3 – Rewriting for Precision

    Purpose: Convert to structured, plain-language modules.

    Prompt:

    Rewrite each segment according to its type:

    • Use short declarative sentences
    • Bullet points for steps or lists
    • Avoid vagueness or repetition

    Step 4 – Grouping by Type

    Purpose: Reorganise output by information type.

    Prompt:

    Sort all rewritten segments under clear headings:

    • 🗂 Reference
    • 🧠 Concept
    • ⚖️ Principle
    • 🔄 Process
    • 🔧 Task

    Preserve segment numbers.

    Step 5 – Structured Output Bundle

    Purpose: Package the content for reuse.

    Prompt:

    Format output with markdown or minimal HTML.
    Include metadata at the top:

    Title: [your title]
Source: [file name or link]
Date: [today's date]
Content type: Precision Content

    Step 6 – Glossary Generation

    Purpose: Extract and define key terms.

    Prompt:

    Identify important terms in the text and define each using only information in the document.

    Format:
    Term: [definition]

    📄 Source Document

    [Paste the cleaned or raw source text here after Step 0.]

    Work Area – Output by Step

    Use this section to draft segmented content, types, rewrites, and grouped outputs.

    Review Notes / Final Output / Glossary

    Use this area for human commentary, final outputs, or glossary results.

  • Collaboration in ChatGPT?

    transparency label: human only

    There are reports that:

    OpenAI has been quietly developing collaboration features for ChatGPT that would let multiple users work together on documents and chat about projects, a direct assault on Microsoft’s core productivity business. The designs have been in development for nearly a year, with OpenAI’s Canvas feature serving as a first step toward full document collaboration tools.

    Source: The Information via BRXND Dispatch

    This would be a change towards something like simultaneous collaboration with colleagues on a shared document in Microsoft Office. At present, a ChatGPT Team account allows more than one person in the Team account to work in a project space and to take part in the chats within that project, but only one person at a time, as I understand it.

  • First thoughts on a lab framework

    transparency label: Human-only

    A few hours spent with ChatGPT-o3 resulting in good first draft of a framework for thinking about our labs. It covers

    • types of lab
    • the roles of people involved with the labs
    • the core technical configuration of a lab
    • assets needed to launch, operate, and archive a lab
    • a naming convention for these assets

    No doubt the framework will need to be tweaked and added to as our ideas mature.

    The chat with o3 was a valuable mind-clearing exercise for me, and I was impressed by how much more “intellectual” it is compared to the 4o model. Like many intellectuals, it also displayed a lack of common sense on occasions, especially when I asked for simple formatting corrections to the canvas we were editing together. The 4o model is much more agile in that respect.

    During the chat, when the flow with ChatGPT didn’t feel right, I hopped back and forth to consult with Perplexity and NotebookLM. Their outputs provided interestingly and usefully different perspectives that helped to clear the logjam.

    A decision arising from my joint AI consultation process was the choice of Google Workspace for the office productivity suite within our labs. This will allow for much better collaboration when using office tools with personal licences than would be the case with Microsoft Office 365. Given the ad hoc nature of labs and the cost constaints we have, this is an important consideration.

  • Drafting Anapoly’s first Lab Setup with ChatGPT

    Transparency label: AI‑heavy (ChatGPT model o3 produced primary content; Alec Fearon curated and lightly edited)

    Purpose of experiment

    Use ChatGPT as a thinking partner to create a first draft of an acclimatisation lab setup for Anapoly AI Labs and to compare output quality between two models (o3 and 4o).

    Author and date

    Alec Fearon, 24 June 2025

    Participants

    • Alec Fearon – experiment lead
    • ChatGPT model o3 – reasoning model
    • ChatGPT model 4o – comparison model

    Lab configuration and setup

    Interaction took place entirely in the ChatGPT Document Workbench. The first prompt was duplicated into two branches, each tied to a specific model. All files needed for reference (conceptual framework, lab note structure) were pre‑uploaded in the project space.

    Preamble

    Alec wanted a concise, critical first draft to stimulate team discussion. The exercise also served as a live test of whether o3’s “reasoning” advantage produced materially better drafts than the newer 4o model.

    (more…)

  • That was the moment …

    transparency label: Human only

    It hit me that generative AI is the first kind of technology that can tell you how to use itself. You ask it what to do, and it explains the tool, the technique, the reasoning—it teaches you. And that flipped something for me. It stopped being a support tool and became more like a co-founder.

    A quote from BRXND Dispatch, a SubStack newsletter by Noah Brier, which featured Craig Hepburn, former Chief Digital Officer at Art Basel.