Blog

LLM Wiki: Theory part01 | What Is Karpathy's LLM Wiki, and How Is It Different from RAG?

Karpathy dropped his LLM Wiki Gist in April 2026. Where it actually differs from RAG, what the three-layer architecture is, and when the extra layer pays off.

Harness Theory Part 12 | Patterns, anti-patterns and evolution: how to grow a harness and know when to remove it

From behaviour-to-harness derivation, deterministic-first control, Build/Buy/Compose and adoption operating modes to entropy, knowledge governance, assumption ledgers, ablation and retirement, this article develops a method for harnesses that can grow, prove their value and become safely simpler.

Harness Theory Part 10 | Quality and evaluation: how do you prove an agent has finished the task?

From task and completion contracts, quality gates, criterion-specific evidence authority and verification portfolios to runtime versus evaluation harnesses, step/trajectory/task evals, trial validity, evaluator calibration and model-harness compatibility.

Harness Theory Part 09 | Safety and Permissions: Letting Agents Act Without Letting Them Cross the Line

From agentic threat models, trust boundaries, permission precedence, human approval and sandboxing to credential binding, tenant isolation, data egress, extension supply chains and red teaming, this article explains how an agent harness confines autonomous action to authorised, isolated and auditable boundaries.

Harness Theory Part 07 | Long-running and Multi-agent Work: How Tasks Survive Context Boundaries

A practical model for fresh-context iteration, session bootstrap, progress ledgers, clean hand-offs, subagent contracts, multi-agent messaging, atomic claiming and workspace isolation.

Harness Theory Part 06 | Runtime and state: how does a harness sustain an execution world?

From initialisation, prompt assembly, model adapters and canonical run state to checkpoints, control planes, hosting, configuration versioning and protocol negotiation, this article examines who owns the truth after an agent starts work.

Harness Theory Part 04 | Context Engineering: Where Do Context, Memory, and RAG Belong?

A practical guide to context builders, progressive disclosure, managed artefacts, compaction, memory, RAG, cache invalidation, and fresh-session reconstruction.

Harness Theory Part 03 | Agent Loops and Control Flow: How Does the Model Start Doing Work?

A practical account of tool calling, agent-loop invariants, workflows and orchestration, item–turn–thread protocols, bidirectional control, stopping policy, and goal drift.

Harness Theory Part 01 | What Is a Harness, and Why Do Agents Need One?

A practical definition of an agent harness, its boundary with the model and product, harnessability, and the human steering control plane.

The Atlas of Agent Design Patterns Part 10 | Implementing Agent Patterns Without Framework Lock-In

A framework-aware but architecture-first guide to Native Code, LangChain and LangGraph, LlamaIndex Workflows, CrewAI, OpenAI Agents SDK, AutoGen, Semantic Kernel, and Microsoft Agent Framework.

The Atlas of Agent Design Patterns Part 8 | Production Agent Architectures in Practice

A production-focused guide to assembling routing, durable orchestration, tools, verification, state, memory, policy, evaluation, observability, budgets, and human control into RAG, deep-research, coding, browser, enterprise-automation, and monitoring systems.

The Atlas of Agent Design Patterns Part 7 | Context, State, Memory, and RAG

A production-focused guide to working, episodic, semantic, procedural, user, and shared memory; the boundary between memory and RAG; memory write and retrieval policies; conflict resolution; forgetting; and production memory architecture.

The Atlas of Agent Design Patterns Part 5 | Verification, Recovery, and Self-Correction

A production-focused guide to retry, parameter repair, fallback, Self-Refine, critics, verifiers, Generate-and-Test, replanning, Reflexion, human approval, idempotency, and bounded recovery.

The Atlas of Agent Design Patterns Part 4 | From Single-Path Reasoning to Trees, Graphs, MCTS, and LATS

A production-focused guide to single-path reasoning, Chain-of-Thought, self-consistency, Generate-and-Rank, beam search, Tree of Thoughts, Graph of Thoughts, MCTS, LATS, evaluator design, diversity, stopping rules, and search budgets.

The Atlas of Agent Design Patterns Part 2 | Agent Execution Paths: Direct Calls, Pipelines, Routers, State Machines, and DAGs

A practical guide to the execution structures behind production agent systems, including direct calls, fixed pipelines, routers, state machines, DAGs, event-driven triggers, human approval, and behavior trees.

The Atlas of Agent Design Patterns Part 1 | Beyond ReAct: A Six-Dimensional Map of LLM Agent Architectures

A practical engineering map for analysing LLM agent architectures across execution path, decision and planning, reasoning and search, verification and recovery, agent organisation, and state and memory.

From Pain to Venture Part 10 | Demand creation: having a product does not mean people will come.

Many products do not die because they have no value.

From Pain to Venture Part 09 | Market, competition, and positioning: you are not building in a vacuum

The earlier parts looked at problems, customers, MVPs, and business models.

From Pain to Venture Part 07 | An MVP is not a smaller product. It is a learning machine.

When people say MVP, they often mean “a cheaper version of the product”.

From Pain to Venture Part 06 | From problem to solution: do not invent the product yet, invent the hypotheses

Once a problem has been found, it is tempting to relax.

From Pain to Venture Part 04 | A pain point is not a complaint. The useful ones are important and still unmet.

Founders are easily drawn towards complaints.

From Pain to Venture Part 03 | Do not ask users what they want. Ask where they get stuck.

There is a kind of customer interview that goes wrong from the first sentence.

From Pain to Venture Part 01 | Do not start with the idea. Start by admitting where things are stuck.

Some venture ideas are not born in front of a whiteboard.

x402: Invisible Web3 03 | When Web3 Disappears, So Do the Risks

Invisible Web3 should not hide risk. It should move authority, budgets, metadata protection, facilitators, discovery, refunds and audit into the product.

x402: Invisible Web3 01 | The Most Successful Web3 May Be the One Users Never Notice

An opinion piece on invisible Web3, x402, AI agent payments, and why mature infrastructure should help users complete tasks instead of forcing them to learn the rails.

Product Rhythms in the Age of AI Part04 | If I Were Running a Product Team Today, I Wouldn’t Run Scrum the Same Way

Product Rhythms in the Age of AI Part02 | AI Won’t Replace PMs First. It Makes Some Work Cheaper Before It Does Anything Else

Product Rhythms in the Age of AI Part01 | It’s Not Agile That’s Aging. It’s Some of the Ways We Deliver

Behind the Trend 01 | How Singapore's Policies Are Turning into Business Opportunities, from Smart Nation 2.0 to NAIS 2.0

Singapore is not only writing tech policy. It is designing the infrastructure for the next round of business formation. This is what I found after reading these documents.

Local LLM Fine-Tuning, Explained Appendix B | Commands, warnings and troubleshooting quick reference

Local LLM Fine-Tuning, Explained part 09 | How to choose a mainline version: Lexi, adapter variants and the judgement that remains

what should I actually keep?

Local LLM Fine-Tuning, Explained part 08 | From adapters back to Ollama: merge, quantisation and deployment

a finished training run is not the same thing as a finished model

What Reality Corrected 05 | Interest is not movement

Near the end of a first-customer-success workshop at Draper U, the speaker drew a ladder on the whiteboard. At the bo…

What Reality Corrected 06 | A better product is not enough if switching still hurts

In another Draper U session, the speaker asked a nastier question than “why would a customer buy?” **If the customer …

What Reality Corrected 08 | Culture is not what you write on the wall. It is what your decisions repeatedly train people to do

Before Jevan Soo Lenox’s fireside at Draper U started, I had a fairly defensive posture towards the topic. Anything l…

Local LLM Fine-Tuning, Explained part 06 | How to read the training scripts: from `train_lora.py` to `train_partial_ft.py`

an experiment design document that binds together the model, the data, the tokenizer, the training method, the trainable scope and the training rhythm

Local LLM Fine-Tuning, Explained part 04 | What SFT, LoRA and full fine-tuning each actually change

SFT answers how you teach. LoRA answers how you change parameters. Full fine-tuning answers how deep you are willing to cut.

What Reality Corrected 03 | A pitch deck is not your company story. It is the thing that decides whether an investor keeps reading

At Draper U, Melanie opened her pitch workshop with a line that was more useful than any deck template. Imagine your …

Local LLM Fine-Tuning, from Modelfile and LoRA to DPO part 03 | How the toolchain splits: Hugging Face, Transformers, PEFT, TRL, and Ollama

Local LLM Fine-Tuning, from Modelfile and LoRA to DPO part 02 | How model memory actually splits: context, external memory, parameter memory, and continual learning

What Reality Corrected 01 | Before anyone replies, they need to know where to place you

The thing I kept from Rachel Konrad's two Draper U sessions was not a writing trick. It was something more awkward th…

What Reality Corrected 02 | It wasn't that you were bad at networking. Most networking events are just badly designed

The first time this really clicked for me was not in class. It was after I hosted a founder × investor mixer in the B…

LLM Application Engineering and Governance for PMs Part 3 – LLMs Do Not Do the Work Alone: How Tool Calling and RAG Connect Models Back to the Real World

The first instinct many teams have when they start building with LLMs is fairly predictable: write a better prompt, add a few examples, and see whether the model can somehow hold the whole thing toget

After the Pause | On Starting, Pivoting, and Pausing a Venture Over Two Years — Part 07: Going from founder back to candidate was harder than I expected

I assumed one thing would at least become simpler once I came back to the job market after a stretch of building a company.

After the Pause | On Starting, Pivoting, and Pausing a Venture Over Two Years — Part 06: I Did Not Give Up. I Just Needed to Stay Afloat

In the end, I did not pause the venture because I stopped believing in the problem.

LLM Application Engineering and Governance for PMs Part 1 － Beyond Prompt Engineering: What PMs Are Actually Doing When They Put LLMs into Products and Workflows

- You already use ChatGPT, but you’re now trying to wire LLMs into a product, an internal workflow, or some form of automation - You keep hearing prompt engineering, RAG, tool calling, agents, and it

Category Design for Founders Part 2 － How Founders Actually Do Category Design: Problem Framing, POV, Languaging, Fundraising, and Go-to-Market

Part 1 dealt with one thing first: category design is not a naming game, and it is not a grand branding wager every startup should force. At its core, it is about correcting a market that is using the

After the Pause | On Starting, Pivoting, and Pausing a Venture Over Two Years — Part 04: I Thought Product Would Be the Hard Part

At the beginning of my second venture, I placed most of my attention where many founders do.

After the Pause | On Starting, Pivoting, and Pausing a Venture Over Two Years — Part 03: It Was Not a Lack of Direction. Reality Kept Changing the Question

There was a stretch when my pitch deck changed every few weeks.

PM User Research and Fieldwork 04 － Five Participants Is Not a Recruitment Strategy: How PMs Find the Right Users, Screen Them Properly, and Avoid the Wrong Ones

Recruiting five participants is not a strategy. This article explains how research goals, segment variation, risk, and saturation should shape participant selection.

PM User Research and Fieldwork 09 － A Complaint Is Not Yet a Job: How PMs Confirm JTBD, Switching Moments, and Pain Intensity Without Jumping to Solutions

A complaint is raw input, not yet a problem definition. This article shows how to distinguish symptoms, needs, jobs, constraints, and actionable opportunities.

After the Pause | On Starting, Pivoting, and Pausing a Venture Over Two Years — Part 01: I Did Not Suddenly Decide to Start a Company

One of the most common reactions I get, once people learn I went off to build a start-up, goes something like this.

PM User Research and Fieldwork 07 － Interviews and Fieldwork Are Not Casual Conversations: How PMs Facilitate, Observe, Take Notes, and Avoid Distorting the Session

Good interviews and fieldwork require disciplined facilitation. This article covers rapport, neutrality, probing, observation, and session control.

PM User Research and Fieldwork 06 － An Interview Guide Is Not a List of Questions: How PMs Design Non-Leading Prompts That Surface Real Evidence

An interview guide is a decision tool, not a list of questions. This article shows how to structure prompts, probes, and follow-ups without leading participants.

PM User Research and Fieldwork 03 － Not Every Research Session Is an Interview: User Interviews, Usability Tests, Field Studies, and Diary Studies Serve Different Jobs

Interviews are only one research method. This article compares interviews, observation, diary studies, usability tests, and other approaches by the evidence they produce.

PM User Research and Fieldwork 02 － Qualitative, Quantitative, and Mixed Methods: PMs Do Not Need a Side, They Need the Right Question

A practical guide to choosing qualitative, quantitative, or mixed methods based on the decision you need to make, not the method you prefer.

PM Growth Levers and Monetisation 06 － Growth OS: Mature PMs Do Not Just Generate Ideas—They Turn Metrics, Experiments, Decisions, and Scale into a Repeatable Operating System

A surprising number of teams are not short of ideas. What they are short of is rhythm.

PM Growth Levers and Monetisation 05 － Acquisition Is Not Just Conversion: PMs Should Understand SEO, Paid, Partnerships, Attribution, and Incrementality

When PMs talk about acquisition, the conversation often collapses far too quickly into a few familiar questions: what is CAC, what is ROAS, which channel converts better?

PM Growth Levers and Monetisation 03 － Lifecycle Is Not About Sending More Nudges: Retention Is About Designing a Reason to Return

When teams say they want to improve lifecycle, what they often mean is “we should probably send more messages”.

PM Growth Levers and Monetisation 02 － Value Proposition, Message Tests, and Landing Pages: Many Growth Experiments Fail Because the Promise Is Vague

A surprising number of growth problems are really translation problems.

PM Product Data and Experimentation 08 － Data Rarely Breaks at the SQL Layer: PMs More Often Fight Identity Issues, Late Events, Bot Traffic, and Failed Rollouts

When PMs see strange numbers, the first instinct is often something like this:

PM Product Data and Experimentation 07 － A/B Testing Is Not Just Hypotheses and P-Values: PMs Need Exposure Rules, SRM, Guardrails, and Validity

A surprising number of experiments do not fail because the idea was weak. They fail for a far more awkward reason.

PM Product Data and Experimentation 05 － Window Functions, Sessionisation, and Funnel SQL: Where SQL Starts Becoming a PM Superpower

A lot of PMs learn SQL by learning how to count.

PM Product Data and Experimentation 04 － SQL for PMs: Learn to Pull Evidence Before You Claim to Be Data-Driven

When PMs say they want to become more data-driven, the first instinct is often to ask for more dashboards.

PM Product Data and Experimentation 02 － Metric Trees Are Not Enough: Why PMs Should Understand Metric Dictionaries, Governance, and the Semantic Layer

When PMs start building real data fluency, one of the first “advanced” tools they tend to learn is the metric tree.

PM Product Data and Experimentation 01 － North Star Is Not Enough: How PMs Use HEART, Guardrails, and Counter-Metrics to Define Success

When PMs first get introduced to product metrics, they usually hear a powerful idea very early on: find your North Star Metric, and the whole team will start pulling in the same direction.

RAG Engineering in Practice 07 – How My Job Agent Turns JDs, CVs, and Rubrics into a Working Evidence Pipeline

If the first six articles in this series were mostly about sharpening concepts, setting boundaries, and clearing technical debris off the road, this one is about putting those ideas back into a real workflow and seeing how they actually hold together.

RAG Engineering in Practice 06 – Taking RAG into Production: Retrieval, Citation, Evaluation, and Observability Are Not Optional

A lot of RAG systems are very good at performing in demo mode.

RAG Engineering in Practice 04 – Why Qdrant Throws `Index required`: Payload Indexes, Schema, and Filter Design Explained

The first time you add a filter in Qdrant, there is a decent chance you will run into an error that feels oddly rude:

RAG Engineering in Practice 03 – The Engineering Philosophy of Chunking: You’re Designing Evidence Units, Not Splitting Text

One of the first terms people run into when building RAG is chunking. And that is often where things start to drift.

RAG Engineering in Practice 01 – RAG Is More Than a Vector Database: Start with the Full System Map

When people first get into RAG, their attention often goes straight to the vector database. That is understandable. It is usually the part that sounds the most new, the most technical, and the most likely to impress someone in a meeting.

OpenClaw Deployment and Configuration Part4

When OpenClaw breaks, how to sort the failure into the correct fault line before you waste hours guessing.

OpenClaw Deployment and Configuration Part2

After installation, which settings are worth touching first, and which ones should wait.

OpenClaw Deployment and Configuration Part1

Why I usually recommend starting OpenClaw on macOS or a Mac mini before reaching for a VPS or a tiny Linux VM.

ComfyUI Series 07 | Troubleshooting and common pitfalls: SQLAlchemy, requirements.txt, OOM, red nodes, and why safetensors keeps coming up

If you have made it through the earlier parts of this series, you are probably no longer stuck at “how do I install this”, but rather at questions like these:

ComfyUI Series 06 | How do you install models? Checkpoints, clips, loras, vae, T5XXL, GGUF, and FP8 explained

This is the point where things stop being “I vaguely understand the theory” and become “why is my workflow full of angry red boxes”.

ComfyUI Series 04 | Where do you actually get models from? What Civitai and Hugging Face do differently, and how to read a model page properly

Who is this for? Anyone who has ComfyUI running and is now ready to install models, only to discover that one page is full of checkpoints, another is full of LoRAs, and everything has version numbers, strange file names, and far too many download buttons. If you’ve just finished the install, this is the right place to go next.

ComfyUI Series 03 | Don’t just click about blindly: `127.0.0.1:8188`, your first image, tmux, background runs, and what happens after a reboot

Who is this for? Anyone who has managed to install ComfyUI, but is now staring at the interface wondering what on earth to do next. If you haven’t installed it yet, read the previous piece first. This one assumes ComfyUI already launches on your Mac.

ComfyUI Series 01 | Picking the right tool first: why I’d start with ComfyUI over A1111, Forge or InvokeAI, and whether a Mac mini M4 with 16GB is enough

Who is this for? Anyone eyeing up local image generation on a Mac mini M4 with 16GB, and trying to work out which UI is actually worth installing first. If you’re new to the Stable Diffusion world, start here, then move on to the installation guide.

OpenClaw Getting Started 05 | What OpenClaw Should Actually Do for You, and What It Should Not

A practical framework for deciding which tasks OpenClaw should handle, based on my real workflow.

OpenClaw Getting Started 03 | Discord, Skills, Memory, Remote Access, and the Safety Baseline

For readers who already have the base install working and now want a usable but controlled OpenClaw setup with Discord, skills, memory, remote access, and safer execution boundaries.

OpenClaw Getting Started 02 | A Recoverable First Install on macOS / Mac mini

A practical macOS install guide for technically comfortable readers who want a boring, recoverable OpenClaw baseline before adding channels, skills, or remote access.

AI Agentic Workflow Series 7 — How AI connects to the world: MCP will not replace workflows, it is the middle layer that lets agents use tools safely

MCP is not a workflow killer, and it is not merely a trendier wrapper around webhooks or function calling. Its real role is to expose tools, data, and prompts through a standardised interface so agents can connect to external systems more safely and consistently.

AI Agentic Workflow Series 6 — When should you use a Workflow, and when do you actually need an Agent?

This article combines my practical heuristics with mainstream definitions from Anthropic and OpenAI to decide when a workflow is enough and when an agent is justified.

Building AI Skills Series Part 2: Why ChatGPT should not connect directly to low-level tools

Why exposing low-level tools directly to ChatGPT becomes difficult to govern, and how a thin skill gateway creates safer task boundaries.

Building AI Skills Series Part 1: What is the actual difference between a Skill, Tool, MCP, Runtime, and Orchestrator?

A practical framework for distinguishing skills, tools, MCP, runtimes, and orchestrators, and understanding which responsibility belongs at each architectural layer.

MCP Engineering Deep Dive 04: Skills Are Not the Same Thing as Runtime Tool Selection

Having skills in your repo does not mean the runtime will actually route by them. In practice, the model is far more influenced by the tool surface you expose, the live inventory a session can see, and whether your descriptions, schemas, and instructions line up.

MCP Engineering Deep Dive 03: Contract Design, Schema Discipline, and Versioning Decide Whether Your Tools Can Grow Up

The hardest MCP servers to evolve are not the ones with too many tools. They are the ones with loose contracts. Once you have multiple clients, public tools, and helper tools, schema discipline becomes the real stability layer.

MCP Engineering Deep Dive 01: Transport and Remote Deployment Are Part of the Design

Part 01 of the MCP Engineering Deep Dive series.

Build Your Own MCP Server — Part 5: The Pitfalls and Operations Manual for Oracle VM and FastMCP

Part 5 of the MCP Server build series.

Build Your Own MCP Server — Part 3: How to Compare MCP Frameworks and Choose One

FastMCP, MCP Framework, xmcp, and Spring AI MCP Server are not four skins over the same idea. The real question is which one fits your language habitat, deployment model, auth pressure, and long-term operating style.

Build Your Own MCP Server — Part 2: How to Deploy a Public MCP Server on Oracle VM

Using Oracle VM, Cloudflare, nginx, and FastMCP to turn an MCP server that merely runs in a terminal into a public `/mcp` endpoint that ChatGPT can actually reach.

Refactoring a Make-first workflow into an MCP execution engine: a practical migration from toolisation to contracts

How I migrated a Make-first job agent into a clean MCP execution engine, moving from flow-driven design to contract-driven architecture.

AI Agentic Workflow Series 3 — How to Build a Context-Aware Job Agent in Make — Part 2

Part 2 of the job-agent build log: fast scoring, RAG-based deep analysis, and guardrails that turn shortlisted roles into decision-quality output.

AI Agentic Workflow Series 1 — How I Choose Between LangGraph, n8n, and Make: Why I'm Backing Make for Now

A practical comparison of LangGraph, n8n, and Make - and why, for my current workflow shape, I'm choosing Make first.