Divya van Mahajan

15 Ways I Use AI (And the Models I Use for Each)

ai-daily-brief-podcast

Overview

The speaker — host of the AI Daily Brief podcast, affiliated with a company called Super Intelligent — walks through approximately 15 real-world AI workflows he uses personally and professionally. The talk is framed as a practical, honest inventory: which tasks he delegates to AI, which specific models or tools he reaches for in each case, and where he acknowledges gaps in his own adoption. The premise is that concrete personal examples are among the most useful things practitioners can share with others on their AI journey.

Source video: No URL was provided in the submission.

Prerequisites

  • Basic familiarity with large language models (LLMs) and the major providers: OpenAI (ChatGPT, O3, GPT-4o), Anthropic (Claude), Google (Gemini), and xAI (Grok).
  • General awareness of AI-assisted content creation tools (Descript, Midjourney, etc.).
  • Understanding of what “deep research,” “reasoning models,” and “vibe coding” mean in the context of current AI tooling.
  • Familiarity with podcast production workflows and content marketing concepts (show notes, short-form clips, newsletters).

Main Points

1. Voice Agents for Enterprise Consulting

  • Super Intelligent uses a custom voice agent built on the OpenAI API to conduct agent-readiness audits for enterprise clients.
  • The agent helps companies identify their most relevant AI agent use cases by understanding how they currently work.
  • Voice agents are described as being at the very beginning of mainstream adoption.

2. Podcast Editing with Descript and VEED

  • The podcast is edited entirely by the speaker’s team using Descript, an AI-powered, collaborative video/audio editing platform.
  • Descript’s Underlord AI suite handles filler word removal, retake detection, and an “eye contact” feature that corrects gaze when reading from a teleprompter.
  • VEED is used specifically for its ability to review and edit footage at 2× or 3× playback speed, a feature unavailable in tools like Adobe Audition.
  • Automatic clipping tools (e.g., Opus) for generating short-form social content are acknowledged as a gap — not yet in use but planned.

3. Podcast Descriptions and Show Notes

  • Full episode transcripts (auto-generated by Descript) are dropped into ChatGPT (GPT-4o or GPT-4.5 research preview) to generate show notes descriptions.
  • This is treated as a purely functional writing task, not an artistic one.
  • AI-suggested titles are generally rejected; the speaker finds ChatGPT over-relies on colons/dashes and multi-part title formats that underperform.

4. Long-Form Writing with Claude

  • For most longer-form writing tasks, the speaker defaults to Claude (currently Opus 4) rather than ChatGPT.
  • Use cases include an enterprise AI weekly newsletter, bonus podcast episode scripts, and Patreon posts for a true-crime podcast.
  • A lightweight prompt example: “Craft a tight email summary of key news and stories in AI last week based on the attached transcripts. Focus on an enterprise audience.”
  • Key reason Claude excels here: custom writing styles (comparable to Midjourney style templates) that can be saved and reused — e.g., “Tech Translator,” “Cheeky Crime Companion,” “Storyteller’s Lens.”
  • Heavy context input (hundreds of minutes of transcripts) causes Claude to naturally mirror the speaker’s voice and thinking.

5. Using O3 as Writer After Deep Research

  • There is one specific situation where O3 outperforms Claude or GPT-4o for writing: when O3 has just completed a deep research dossier and then converts it into a script or document.
  • Example: a detailed background dossier on a major British theft was produced via Deep Research, then O3 turned it into a podcast script with strong results.

6. Deep Research with O3 and Perplexity

  • ChatGPT’s O3-powered Deep Research is the primary tool for substantive research tasks: market sizing, pitch framing, background dossiers, and competitive landscape analysis.
  • Example prompts include: researching projected enterprise spend on building AI agents over five years; generating a list of boutique strategy consulting firms benchmarked against a named example.
  • For quick, everyday research (e.g., P/E ratio lookups), Perplexity is preferred over O3.

7. O3 as Strategic Collaborator

  • The speaker treats O3 as a genuine strategic collaborator, not just a tool — engaging it on almost every significant strategic idea.
  • Use cases range from core business strategy (not shown) to exploratory “side quest” ideation (e.g., hypothetical podcast-startup interaction models).
  • A key benefit: externalizing half-formed ideas into O3 makes it easier to leave them behind and refocus, while also providing a way to test whether an idea has real merit.
  • Grok 4 is being trialed as a parallel competitor to O3 for strategic reasoning tasks following its release.

8. Pitch and Information Architecture

  • Reasoning models (primarily O3) are used to tighten pitches — moving from rough outlines to polished memos and information architectures.
  • Pitch decks are harder for models than memos because training data skews toward formulaic deck structures.
  • Critical workflow note: First responses tend to be sycophantic (expanding the speaker’s existing structure rather than challenging it). The fix is an explicit follow-up prompt: “Use your own prerogative, don’t assume my architecture is correct, tell me how you would change it.”

9. Voice Input with WisperFlow

  • WisperFlow (whisperflow.ai) is used as a system-wide voice recognition tool across iPhone apps, replacing Apple’s native (and poorly rated) dictation.
  • Described as a “total game changer” for turning speech into text across all contexts.

10. Image Generation — Ideogram for Assets

  • Ideogram is the tool of choice for functional collateral: podcast episode covers, title cards, and assets requiring accurate text rendering.
  • Key strengths: high prompt fidelity, excellent text-in-image generation, and automatic prompt expansion with editable output.
  • The tool exposes its expanded prompt so users can iterate on specific elements.

11. Image Generation — Midjourney for Creativity

  • Midjourney is retained for tasks requiring deep aesthetic creativity: artistic experiments, abstract presentation backgrounds, visually ambitious concepts.
  • Example: combining Renaissance Vitruvian Man / da Vinci notebook aesthetics with modern technology imagery.
  • Acknowledged trade-off: Midjourney produces the most visually arresting results but is less reliable for precise, functional requirements.

12. Vibe Coding for Feature Exploration

  • The team at Super Intelligent has a soft ban on verbally describing new feature ideas; they must be prototyped via vibe coding first.
  • Benefits: (1) the originator refines and validates the idea before socializing it; (2) it is far easier to show than tell when presenting to others.

13. Automations — An Acknowledged Gap

  • The speaker does not currently have automated/agentic content pipelines in place, despite having experimented with Lindy, n8n, and Zapier.
  • The identified gap: taking published content → auto-generating short-form videos → distributing across social channels without manual intervention.
  • This is framed as a known failure, attributed to inertia, time pressure, and the “new systems thinking” required to deploy automations.

14. AI Is Bad at Naming

  • A notable negative finding: LLMs are consistently poor at naming — features, projects, products, companies.
  • Output is typically described as too cringy, too long, or both.
  • The speaker has never accepted a model-generated name as usable and continues testing only to track whether this improves over time.

15. Gemini — Growing Experimentation

  • The speaker is increasingly experimenting with Gemini due to Google Workspace integration (Docs, Sheets).
  • Describes Gemini’s models, interfaces, and tools (e.g., AI Studio) as improving steadily.
  • For video generation specifically, Veo 3 (VO3) is the current model of choice when needed.
  • Predicts meaningfully higher personal Gemini usage within 6–12 months.

Key Concepts

  • Descript / Underlord: A collaborative podcast and video editing platform with an AI suite that handles filler word removal, retake detection, and eye contact correction.
  • VEED: A video editing tool notable for supporting 2×/3× playback speed review.
  • WisperFlow: A third-party voice recognition app for iPhone that uses Whisper-class transcription system-wide across apps.
  • Deep Research (O3): OpenAI’s O3-powered research mode that performs multi-step web search, synthesis, and reasoning to produce long-form research dossiers.
  • Perplexity: An AI-native search engine used for quick, lightweight research queries.
  • Ideogram: An image generation tool optimised for text rendering and prompt fidelity, used for functional content assets.
  • Midjourney: An image generation tool favoured for high aesthetic creativity and visually striking outputs.
  • Claude Custom Writing Styles: Saved style profiles in Claude (e.g., “Tech Translator,” “Storyteller’s Lens”) that function like reusable tone-and-voice templates.
  • Vibe coding: Prototyping a software feature or product idea using AI-assisted code generation, without necessarily writing production-ready code, as a way to explore and validate concepts.
  • Sycophancy (in LLMs): The tendency of models to agree with or expand the user’s existing framing rather than challenging it; requires explicit prompting to overcome.
  • Grok 4: xAI’s reasoning model being evaluated as a competitor to O3 for strategic analysis tasks.
  • Veo 3 (VO3): Google DeepMind’s video generation model, used when AI video generation is required.
  • Agent Readiness Audit: Super Intelligent’s consulting product that uses a voice agent to help enterprises identify their highest-value AI agent use cases.

Summary

The speaker offers a candid, tool-specific inventory of roughly 15 AI workflows spanning podcast production, long-form writing, research, strategic planning, image generation, voice input, and software prototyping. The central argument is that model selection is use-case-specific: ChatGPT (GPT-4o/4.5) handles functional copywriting; Claude Opus 4 excels at longer-form writing, especially with custom style profiles and rich context; O3 is the preferred engine for deep research, strategic collaboration, and post-research writing; Perplexity serves quick lookups; Ideogram handles text-heavy assets while Midjourney handles creative imagery; and WisperFlow replaces native mobile dictation. The speaker is equally forthcoming about failures — LLMs are consistently bad at naming, sycophancy must be actively countered in pitch work, and his own automation pipeline remains unbuilt despite recognising it as an obvious priority. The talk closes with a forward-looking note: Gemini’s improving integration with Google Workspace and Grok 4’s early promise suggest the current OpenAI-heavy balance of usage may shift meaningfully within the year.