AI workflows

Every answer 29 answers

• Can AI fix software bugs automatically?

  Partly, yes. My setup sends error-tracker alerts to Slack, and an agent on a separate computer picks them up and patches the bug, often without me knowing. Teammates can also drop a screen-recording link in a bugs channel. It isn't perfect yet, and it only works if errors are visible in the first place.

• Can AI pick the best stories for my newsletter?

  Yes, and selection is a better use of AI than writing. Export your click report with the link for each story, give it to ChatGPT or Claude with your candidate stories, and ask which ones to include based on what readers actually clicked. A click means little until the model can see the content behind it.

• Can AI write and run a local newsletter?

  Mostly, yes. I automated a local newsletter to roughly 97% quality in about an hour with Claude's multi-agent workflows, and Claude Code has built a full issue inside Beehiiv with nobody drafting. The hard part is sourcing, not writing. Local judgment and a human read before send are still people jobs.

• Can AI write local news from city council meetings and permits?

  Yes, and it's ready now. larkfeed's article engine watches public data like council meetings and permits, researches, and drafts the story on its own. It has adapters for the seven main platforms that host council content in the US. The person becomes the editor: review, edit or hit publish.

• Can Claude use my software for me, like Beehiiv, Canva or ClickUp?

  Yes, through connectors. Connect your tools to Claude and give instructions in plain speech, like design four Instagram posts in Canva from this week's stories. When I checked in March 2026, Beehiiv's own connector (MCP server) was read-only: data, questions and reports, but no changes. Assume read-only until a write path is proven.

• Claude vs ChatGPT vs Codex: which AI should I use?

  Use more than one, but pick one per job and stick with it. In mid-2026 I told companies ChatGPT was ahead for general use, yet consistency beats chasing the leader. For coding I lean on Codex for parallel work, and Claude for design-heavy builds and Notion. One AI coding tool per project.

• How can a non-technical founder build software with AI?

  Talk to the coding agent in plain English, one chat per codebase, and ignore its time estimates. I never touched code until late 2025. Almost everything that got my form product ready to launch went through Codex by voice or text, not by me editing code. Demand proof it works live, not just locally.

• How do I automate the whole sales process after a call with AI?

  Let an agent work from the call transcript. After a sales call, mine reads the recording's transcript, builds a client-specific proposal website, creates a Stripe payment link at the price we discussed, and once paid sends onboarding and an intake form. The intake then produces a company overview and an AI context file for the client.

• How do I cut AI costs when using Claude or Codex all day?

  Use the top model for planning and a cheaper model for doing, avoid burning tokens on images and video, and default to the smaller model unless the task needs depth. My standing rule: Sonnet by default, Opus for research. And remember the $200 plans hide thousands of dollars of API-equivalent usage.

• How do I give AI context about my business without re-explaining it every time?

  Have the AI interview you once. Ask Claude or ChatGPT to ask you about 40 questions about you, your business, your processes and workflows, then turn the answers into a context document. Put that file in your project folders or upload it to any chat, and you never have to give it that context again.

• How do I know an AI coding agent really fixed something?

  Make it prove it. I don't accept done, fixed or live without evidence: screenshots of the agent using the feature like a human, the live link, or test results. AI agents regularly say they tested something when they didn't, and fix things locally without saying whether they're deployed. A merged badge doesn't prove production updated.

• How do you get AI agents to do useful work without being prompted?

  Set a bar and give them the access to meet it. My bar is one valuable thing a day per agent without me asking: build something, redo something, or hand me a strategy I can drop into Claude or Codex. A report alone doesn't count. The biggest blocker is logins and permissions.

• How do you run AI agents inside Slack with a team?

  Treat Slack as the agents' office. Teammates talk to agents in shared channels, agents act on what they hear and report back in the thread when done. My rules: replies inside threads, three task roundups a day instead of constant pings, no agent-to-agent DMs, and plain-English recaps. Expect setup pain.

• How do you stop AI from producing generic slop?

  Don't let it freestyle. Give it an approved system to follow: a locked brand and design system for visuals, a banned-words list and real local knowledge for writing, and a reviewer that checks before you see it. Generic model taste goes last. AI defaults are where slop comes from.

• How should I set up an AI second brain for my business?

  Keep one source of truth, split only into business and personal. I built a private second brain five times before realizing 98% of it was company knowledge. Now everything business-related lives in one shared brain every teammate and agent can read, and personal material stays completely separate.

• Should an AI agent start with read-only access?

  Yes, start read-only and expand on purpose. My chief-of-staff agent began with read access to code and a read-only work inbox, with broad write access deferred. The reason: an AI-assisted database job once wrote 725,000 rows into production because production and staging weren't separated. Access is still the biggest blocker.

• Should an AI reviewer check work before a human sees it?

  Yes. Nothing from my agents should reach me without an automatic reviewer checking it first for bad design, bad language, spacing and banned words. Cap it at two or three passes. If it still misses the bar, it comes to me with the reason instead of looping forever.

• Should you plan with one AI and build with another?

  Yes. I use a planning chat as my product manager and a coding agent as the developer. For big builds, one model writes the plan, ChatGPT's top model on high effort redlines it, and cheaper models do the searching and building. Chats forget; coding tools keep files, so the plan has to land there.

• Should you send an AI-written newsletter without reading it?

  No, not until it consistently earns it. My rule today is a human read before every send: check it fits your content, get the details right and double-check the links. The only way I'd skip the person is an automatic reviewer scoring issues 9 out of 10, and early tests scored 5 and 6.

• Should you talk or type to AI tools like ChatGPT and Claude?

  Talk. Speaking is at least three times faster than typing by my estimate, and the bigger win is that you stop self-censoring, so the model gets more context. I use Wispr Flow for dictation, don't type more than a sentence anymore, and even code by voice. Type only exact strings like URLs, numbers and code.

• What AI tools should a non-technical business owner pay for first?

  Three paid tools and one free account: Claude at $20 a month, ChatGPT at $20 a month, a voice dictation tool at about $15 a month, and a free GitHub account. If you're not paying, you don't see what's possible, because the free tiers don't give you access to it.

• What are multi-agent AI workflows, and are they better than one agent?

  A multi-agent workflow has a coordinator agent hand pieces of a job to sub-agents running at the same time, like one researching, one writing and one assembling. In my tests it was faster and less error-prone than one agent doing everything. I got a local newsletter to roughly 97% quality in about an hour this way.

• What is meta-prompting, and why does it improve AI output?

  Meta-prompting means you never prompt AI directly for anything serious. You ask the AI to write the prompt first, then you run that prompt. I use it for deep research, big builds and anything that matters. Paired with voice dictation, it turns a rambling three-sentence ask into a far richer set of instructions.

• What research tools give AI agents better results than plain Google?

  Give agents AI-native search and scraping tools, and keep them on free tiers until they prove out. I gave my growth agent Exa for AI search and Firecrawl for scraping, both on free tiers, plus a prepaid $50 keyword-data API. A peer swears by the Perplexity API for fresh, source-level results.

• What should go in AI agent instruction files like CLAUDE.md and AGENTS.md?

  Short, clear rules that stop the mistakes you keep seeing, not a growing pile of restrictions. I spent about 40 hours over four months tuning my CLAUDE.md, AGENTS.md and skill files, mostly to stop AI tools inventing problems. My rule of thumb: fewer rules, no conflicts, and delete before you add.

• What should you automate with AI first in a newsletter business?

  The back end first: the grunt work and annoying admin tasks, not the writing. My test is simple. If a reader or advertiser will experience the output, keep a human in the loop. If it's internal, default to automating it. Start with recurring per-newsletter steps that add up across every brand.

• Which AI app builder should a beginner start with: Replit, Lovable, v0 or Claude Code?

  Start with a hosted builder like Replit, Lovable or v0, then graduate to Claude Code or Codex. Claude Code is level three, four or five. I couldn't have gone from nothing straight to it. For front-end designs, v0 won about 90% of my side-by-side tests.

• Why do AI coding agents write too much code, and how do you fix it?

  Because agents add rules, restrictions and code far more readily than they delete. My team's custom agent system grew to 70,000 lines duplicating an open-source tool that already worked, which explained the random failures. Told to start over from zero, the coding agent cut less than 1% in two days. You have to insist on deletion.

• Why does ChatGPT or Claude forget my work, and how do I stop it?

  Regular chats don't reliably save things, and long chats run out of context. I lost a four-hour website build that way. For anything you're building, work in tools that read and write folders on your computer, like Claude Cowork, Claude Code or Codex, so the work lives in files instead of chat memory.