Episode 809 | What I Learned Diving into A.I. for 100 Days (with Craig Hewitt)

Summary of Episode 809 | What I Learned Diving into A.I. for 100 Days (with Craig Hewitt)

by Rob Walling

39m•December 2, 2025
Listen to Episode

Overview of Episode 809 | What I Learned Diving into A.I. for 100 Days (with Craig Hewitt)

Rob Walling interviews Craig Hewitt (founder of Castos) about Craig’s 100 Days of AI YouTube series, the practical takeaways he learned while building agentic tools and workflows, and how those lessons inform new product ideas at Castos. The episode covers what tools Craig thinks are best-in-class (and which ones he avoids), how agents differ from simple chatbots or automations, real use cases (customer support, content creation), how Craig produced 100 daily videos, and his product strategy for a second product under the Castos brand.

Key takeaways

  • AI is powerful but still requires human-in-the-loop for reliability — many workflows work ~80% of the time and need supervision/iteration.
  • For code-heavy or "bare metal" workflows, Claude Code (and Claude-based tools) are Craig’s top pick.
  • Agentic tools (LLM + memory + tools) are the next level: they can fetch data, run scripts, and act autonomously in ways chat-only models can’t.
  • Manus (a Claude-based, agentic web UI) is Craig’s favorite single tool if he could use only one.
  • ChatGPT is useful, but Craig finds it more of a consumer product; GPTs (custom GPTs) are useful for narrow workflows.
  • Practical, high-value AI deployments today: intelligent customer support (vector-store + docs), content writing accelerators, and specialized marketing agents.
  • If you’re a founder building or validating a new product, target higher ARPU (Craig uses $100/mo minimum as a rule) and design for expansion revenue.

Recommended tools & best-in-class picks

  • Claude Code (Anthropic) — Craig’s top pick for coding and complex, context-rich workflows.
  • Manus — a Claude-based agentic product with a browser/tool access layer; recommended for agentic work (market research, scraping, automated briefs, code execution).
  • ChatGPT — still widely useful for consumer-style chat; Craig mainly uses GPTs (custom agents) rather than general chat for business tasks.
  • Whisper / voice-to-text tools — useful for transcribing audio quickly (Craig referenced “whisper” workflows).
  • Automation platforms — Make (Integromat), n8n, LinedIn/Lendi/Relevance/Google Opal (examples of automation/agent frameworks).
  • Vector-store + knowledge base tools — used to build effective support agents (example: DocSpot doing Castos support).
  • Creator Hooks, Canva — production tools Craig relied on for ideation/titles/thumbnails and simple design.

Agents vs. automations vs. chatbots (Craig’s mental model)

  • Chatbot: An LLM interaction — you ask, it answers (e.g., ChatGPT or Claude chat).
  • Automation: If-this-then-that workflows calling LLMs or APIs (Make, n8n, Zapier-style).
  • Agent: Combines three components:
    1. LLM (Claude, GPT, etc.)
    2. Memory / persistent state (vector store, session memory)
    3. Tool access (web browsing, database queries, email/calendar, APIs)
  • Agents can act autonomously and chain operations (e.g., fetch 20 transcripts, analyze, generate frameworks) — agents can be embedded into automations.

Real-world examples Craig used

  • DocSpot (support agent): Castos put their docs and site into a vector store and used it to answer support questions. Result: support load roughly halved, and docs improved because the bot surfaced unclear areas.
  • Manus example: Craig asked Manus to fetch the latest 20 videos from a creator’s YouTube channel, download transcripts, analyze patterns, and produce three intro options for a podcast — a multi-step, agentic task that took ~20 minutes.
  • Internal SEO/content tool: Craig built an advanced content-writing workflow using Claude Code (more powerful than off-the-shelf blog generators).
  • Castos production tools: Podscan, PodSqueeze, and other utilities for show notes and prep.

How Craig produced 100 YouTube videos in 100 days — process & tooling

  • No “magic AI” for production — most of the heavy lifting was human work + systemization:
    • Notion board for ideas and planning.
    • Borrowed/iterated on ideas from other creators and used "newsjacking" (timely content) to capture views.
    • Editor on the team did ~85 of the videos and worked weekends — being a little ahead was crucial.
    • Creator Hooks (AI tool) for title and thumbnail suggestions; thumbnails made in Canva.
    • Stayed minimal & scrappy: two times were close calls (recording issues), but being able to ship consistently mattered more than perfection.
  • Result: channel grew to ~11k subscribers; a couple of listicle-style videos hit >50k views.

Castos strategy & the new-product thinking

  • Castos is mature, profitable, stable, but hitting growth limits — Craig wants to leverage their existing audience (40k email list, 4k customers) to build complementary, higher-ARPU products.
  • Product attributes Craig requires:
    • Minimum entry price: $100/month (makes go-to-market and churn economics easier).
    • Clear expansion revenue model (usage-based, ESP/CRM-style, or increasing tiers).
    • Strong fit with existing customer base (cross-sell potential).
  • Three product ideas under evaluation:
    1. LinkBerry.ai — AI-driven LinkedIn content service: generate a month of content for creators/entrepreneurs for ~$100/mo, positioned as an affordable alternative to ghostwriting agencies.
    2. SaaS wrapper for Claude Code content workflows — "AI SEO/content tool" offering strong context and quality writing (harder to differentiate vs. crowded market).
    3. Suite of marketing agents controllable from Slack — army of specialized agents: SEO data pulls, write articles, ad copy, LinkedIn posts, etc. (ambitious, likely requires more funding and engineering).
  • Emotional dynamics: excited about building again but nervous about starting from scratch and validating product-market fit after 8 years.

Actionable recommendations for founders (practical next steps)

  • Start small with high-value agent use cases: customer support (vectorize docs + deploy a support agent) has immediate ROI and reduces load.
  • If you build bots, include an escape hatch (easy route to contact a human) and monitor logs to iterate docs and responses.
  • Consider one AI automation/agent per month to compound improvements over a year.
  • When exploring new products:
    • Aim for meaningful ARPU ($100+/mo) and expansion revenue.
    • Validate tightly with your existing audience first (waitlists, early access).
    • Focus on an “aha moment” — product should deliver clear, immediate value.
  • Tool selection:
    • Use agentic UIs (Manus, Claude Code) for multi-step, data-driven tasks.
    • Use ChatGPT/GPTs for narrow, reusable workflows where a simple chat interface suffices.

Notable quotes

  • “It’s incredible and not there all at the same time.” — on current AI capabilities.
  • “An agent has three components: an LLM, memory, and access to tools.” — Craig’s concise agent definition.
  • “If you just did one of these [agents/automations] a month, where would you be the next year?” — on compounding automation investments.

Where to follow Craig / links (as mentioned)

  • X / Twitter: @TheCraigHewitt
  • YouTube: youtube.com/TheCraigHewitt
  • Newsletter / site: craighewitt.com
  • LinkBerry waitlist: linkberry.ai

Short summary: Craig’s 100-day AI deep dive surfaced that agentic, tool-enabled LLMs (Claude Code / Manus) are where the biggest practical gains are today. ChatGPT remains useful but more consumer-focused; deploy agents for support and specialized workflows, monitor and refine, and when building new products aim for clear value, $100+/mo pricing, and expansion revenue.