What Craig Hewitt learned from 100 days of AI tools

Executive overview

Most founders are using the wrong AI tools — relying on ChatGPT out of habit rather than choosing what actually performs. Craig Hewitt ran 100 days of AI experimentation and identified a clearer stack: Manus for agentic research and task execution, Claude Code for coding and content, and purpose-built agents for specific workflows.

AI can automate heavily but still fails roughly 20% of the time, requiring human oversight. The value is amplification, not full replacement — yet.

The biggest unlock is treating AI as an agent with tools and memory, not just a chat interface.

The AI tool stack worth knowing

• Claude Code is the recommended coding tool — "bare metal" AI in the terminal with file access, MCP support, and agentic capabilities; the name undersells it, as it handles content creation too
• Manus is the single recommended all-purpose tool — an agentic interface running Claude under the hood, with browser access, code execution, and report generation
• ChatGPT's main remaining use case is shareable, trained GPTs for repeatable team or customer tasks
• Manus has three modes: agent, adaptive (model chooses), and chat — making it a full ChatGPT replacement for most workflows
• Voice-to-text (super-whisper / whisper flow) rounds out the stack for fast input

Why Manus outperforms ChatGPT for complex tasks

• ChatGPT struggles when tasks exceed a single context step — e.g., transcribing an MP3 attached mid-session
• Manus can autonomously browse, download, execute Python, and aggregate data — Craig used it to pull transcripts from 20 YouTube videos, analyze them, and produce a rubric in ~20 minutes
• The agent/automation/chatbot distinction matters: a chatbot replies; an automation chains steps; an agent combines an LLM + memory + tool access
• ChatGPT's agent mode underperforms on real-world tasks like booking travel; Manus handles them reliably

When AI agents deliver real value

• Many hyped agent demos (e.g., pre-call research briefs) aren't worth the build-and-maintain overhead
• The clearest proven use case: customer support agents trained on a full knowledge base
• Craig deployed DocSpot at Castos — cut support load by ~50% across 4,000 customers for $80/month
• An escape hatch to human support is essential; without it, chatbots frustrate users
• Monitoring chatbot logs improved documentation quality as a side effect

What 100 AI videos taught about the technology

• The "I automated my entire life" narrative is real but overstated — 80% reliability means 20% still needs a human
• AI is not eliminating jobs at the individual-tool level yet; at the macro level, it is reducing headcount at large companies
• ~70 days in, Craig hit a realization: most of what he built was impressive but useless — pivoted to building actual products with Claude Code
• The 100-day format forced consistent output; ideation (not production) was the hardest part
• Borrowed ideas from other channels, news-jacked major AI releases, and used Creator Hooks for title/thumbnail generation
• An editor handled ~85 of 100 videos, working weekends to keep pace

Building a second product from an existing SaaS base

• Castos is profitable and stable at low seven figures ARR but unlikely to become a $20M+ company through organic growth alone
• After 1.5 years of intense growth attempts with strong marketing talent, Craig concluded the core product has reached a natural ceiling
• The strategic response: layer complementary products onto a 40,000-person email list and 4,000 existing customers
• Three product criteria Craig set: minimum $100/month entry price, expansion revenue built into the model, alignment with existing audience
• Linkberry.ai — AI LinkedIn content creation ("a month of posts for under $100") — is the lead candidate; waitlist open
• A second candidate: a standalone AI SEO/content writing tool built internally in Claude Code — larger market but harder to differentiate
• Third idea: a Slack-based command hub for a fleet of marketing agents (LinkedIn, ad copy, SEO) — compelling but complex and likely requires outside funding
• Validation instinct has changed after eight years: works backwards from the "aha moment" to the product rather than building on intuition