A seven-step process for making AI write in any human style

Dylan Davis July 15, 2025

• Build two AIs: one extracts style, one mimics it
• Mixing multiple writing styles confuses the model — pick one
• Same-model aggregation is biased; use a third-party model to judge

Using AI as a custom consultant to find tax savings

Dylan Davis July 10, 2025

• Interview AI first so it knows your full context before advising
• Build a personalised system prompt from the synthesised conversation
• Always validate AI tax advice with a human before acting on it

Seven lessons from building 50+ AI automations

Dylan Davis July 8, 2025

• Automate the highest-impact bottleneck first, not the most interesting one.
• Prove feasibility with a top model before optimising for cost or speed.
• One model and one hour beat three weeks of regex pipelines.

Seven habits separating top 1% AI users from everyone else

Dylan Davis July 3, 2025

• Dictating instead of typing unlocks dramatically faster AI throughput.
• Projects and context-loading separate casual users from power users.
• Model loyalty is a trap — top performers switch models by task.

Three AI prompt mistakes killing your output quality

Dylan Davis July 1, 2025

• Why prompting once and hoping is guaranteed to fail
• Context dominates prompt quality — here's how to engineer it
• Iterating across multiple models beats any single perfect prompt

The Hidden Power Behind Every ‘Smart’ AI Tool

Dylan Davis June 26, 2025

Matching AI agent architecture to the right problem

Dylan Davis June 17, 2025

• Multi-agent systems beat single agents by 80% — but only for research
• Coding agents fail when sub-tasks share dependencies; use single agents
• A three-question framework tells you which architecture fits your problem

How AI voice agents work and their key constraints in 2025

Dylan Davis June 13, 2025

• 500ms is the human threshold — beyond it, voice agents feel robotic
• Conversation cost grows exponentially: 30 mins can cost 100x more than 3
• Open-ended voice agents fail; constrained, single-task agents succeed

MCP in production: why it fails and how to use it safely

Dylan Davis June 10, 2025

• MCP agents fail 20–50% of the time with today's best models.
• Four rules cut failure rate: fewer tools, human language, chaos limits, locked dangerous actions.
• Only coding, research, and triage have proven production-scale success today.

Mary Meeker's 340-slide AI report: key findings distilled

Dylan Davis June 5, 2025

• Inference costs fell 105,000x — smarter AI is now nearly free to serve.
• Only 37% of US adults have ever used ChatGPT.
• Nvidia captures ~25% of all global data center capex spend.

Seven prompt engineering lessons from Claude 4's leaked system prompt

Dylan Davis June 3, 2025

• Match tone and length to conversation type using style rules in your system prompt.
• Use a no/maybe/now decision tree to avoid unnecessary tool calls and latency.
• Reserve hard 'never' rules for legal or safety limits; default to positive instructions.

Build an MCP server in 20 minutes using only AI

Dylan Davis May 29, 2025

• A reverse AI interview generates the system prompt that drives everything.
• Gemini 2.5 Pro's large output window produces the best standalone blueprints.
• Claude Sonnet 4 beats GPT-4.1 for error-free prompt-by-prompt implementation.

Claude 4 prompting techniques for advanced reasoning models

Dylan Davis May 27, 2025

• Your prompt's style directly shapes how Claude 4 responds.
• State what you want — negation prompts are now obsolete.
• Parallel tool calling lets Claude synthesise multiple results at once.

OpenAI Codex: how autonomous agents will reshape software teams

Dylan Davis May 18, 2025

• Codex runs parallel agents autonomously — no back-and-forth prompting needed
• Agents absorb tech debt so engineers focus only on revenue features
• Enterprise teams over 200 people are the primary target, not solo devs

Cutting AI noise: 13 high-signal sources and how to consume them

Dylan Davis May 15, 2025

• Filter by utility first — most AI content fails that test immediately
• Sampling long podcasts with AI saves hours without missing key insights
• The Latent Space Discord replaces X as a high-signal AI news feed