Netta & Char
@letstalkaitools.bsky.social
👩💻 Just two product gals (and their chatbots) exploring Gen AI tools. Creativity, coding, & weekly AI adventures. We walk you through case studies & our wins/fails.
11/11 - We're witnessing the shift from manual coding, or single-action tools, to multi-agent-orchestrated development. The results speak for themselves: production tools that actually solve real business problems.
Full technical breakdown: medium.com/p/142f8da60e...
Full technical breakdown: medium.com/p/142f8da60e...
Weeks 20–21: We Build Agents (and other efficiencies)
The following blog post was written by Claude, and then edited by a human. The draft was based on a recording of our weekly meeting…
medium.com
September 4, 2025 at 8:23 AM
11/11 - We're witnessing the shift from manual coding, or single-action tools, to multi-agent-orchestrated development. The results speak for themselves: production tools that actually solve real business problems.
Full technical breakdown: medium.com/p/142f8da60e...
Full technical breakdown: medium.com/p/142f8da60e...
10/11 - NotebookLM turned a 6-hour transcript into a professional illustration-style video in 5 minutes. We could imagine educators turning textbook chapters into quick-reference videos. Multi-medium learning for the win.
September 4, 2025 at 8:23 AM
10/11 - NotebookLM turned a 6-hour transcript into a professional illustration-style video in 5 minutes. We could imagine educators turning textbook chapters into quick-reference videos. Multi-medium learning for the win.
9/11 - Gemini's storybook creation, which generated text & images based on a prompt, for a whole storyline. It maintained perfect character consistency throughout our sample children's book (and had pretty nice visuals to boot!)
September 4, 2025 at 8:23 AM
9/11 - Gemini's storybook creation, which generated text & images based on a prompt, for a whole storyline. It maintained perfect character consistency throughout our sample children's book (and had pretty nice visuals to boot!)
8/11 - Lindy AI's agent builder, a real step above the competition (Make, n8n). You describe your project in natural language, and it builds the entire workflow with pre-filled prompts. No more configuring technical nodes on your own.
September 4, 2025 at 8:23 AM
8/11 - Lindy AI's agent builder, a real step above the competition (Make, n8n). You describe your project in natural language, and it builds the entire workflow with pre-filled prompts. No more configuring technical nodes on your own.
7/11 - We also tested some other tools. Our three faves?
September 4, 2025 at 8:23 AM
7/11 - We also tested some other tools. Our three faves?
6/11 - Project 3: Newsletter automation that receives emails with upcoming events and links to event posters, extracts times/locations from the PDFs using Google Cloud Vision OCR, then creates calendar events automatically. Runs 3x/week with a shareable calendar.
September 4, 2025 at 8:23 AM
6/11 - Project 3: Newsletter automation that receives emails with upcoming events and links to event posters, extracts times/locations from the PDFs using Google Cloud Vision OCR, then creates calendar events automatically. Runs 3x/week with a shareable calendar.
5/11 - The result is an interactive financial tool that integrates live APIs for interest rates, mortgage data, and tax rates by country. Put in key inputs (like price and sq meterage), and get sensitivity analyses and cash flow projections, all exportable to Excel.
September 4, 2025 at 8:23 AM
5/11 - The result is an interactive financial tool that integrates live APIs for interest rates, mortgage data, and tax rates by country. Put in key inputs (like price and sq meterage), and get sensitivity analyses and cash flow projections, all exportable to Excel.
4/11 - Then Claude Code set up the dev environment: 5 coding sub-agents working in parallel on isolated Git worktrees that could merge without conflicts, plus a dedicated code review sub-agent. It suggested and configured everything on its own.
September 4, 2025 at 8:23 AM
4/11 - Then Claude Code set up the dev environment: 5 coding sub-agents working in parallel on isolated Git worktrees that could merge without conflicts, plus a dedicated code review sub-agent. It suggested and configured everything on its own.
3/11 - Project 2: Built a warehouse buy/rent decision tool from a meeting transcript. Claude Code created specialized business analyst & technical architect sub-agents for fleshing out the idea, then a compiler agent to turn all that into a (28-page!) PRD
September 4, 2025 at 8:23 AM
3/11 - Project 2: Built a warehouse buy/rent decision tool from a meeting transcript. Claude Code created specialized business analyst & technical architect sub-agents for fleshing out the idea, then a compiler agent to turn all that into a (28-page!) PRD
2/11 - Project 1: Automated monthly driving summaries for 140+ employees. What used to take managers DAYS of manual spreadsheet work now happens with one button press. Complete with email previews, error handling, and historical data backfill.
September 4, 2025 at 8:23 AM
2/11 - Project 1: Automated monthly driving summaries for 140+ employees. What used to take managers DAYS of manual spreadsheet work now happens with one button press. Complete with email previews, error handling, and historical data backfill.
8/8 - Specialized AI tools are outperforming general-purpose ones for professional work.
This week's winners: ✅ Midjourney (brand/video work) ✅ Elicit (academic research)
Full breakdown: medium.com/p/729cccc854...
This week's winners: ✅ Midjourney (brand/video work) ✅ Elicit (academic research)
Full breakdown: medium.com/p/729cccc854...
September 3, 2025 at 10:04 AM
8/8 - Specialized AI tools are outperforming general-purpose ones for professional work.
This week's winners: ✅ Midjourney (brand/video work) ✅ Elicit (academic research)
Full breakdown: medium.com/p/729cccc854...
This week's winners: ✅ Midjourney (brand/video work) ✅ Elicit (academic research)
Full breakdown: medium.com/p/729cccc854...
7/8 - Our debugging tip: When stuck in AI loops, don't keep hitting the same wall.
Switch tools. Have one AI summarize what’s been tried, and another analyze what another got wrong.
We got out of endless debugging loops using this approach.
Switch tools. Have one AI summarize what’s been tried, and another analyze what another got wrong.
We got out of endless debugging loops using this approach.
September 3, 2025 at 10:04 AM
7/8 - Our debugging tip: When stuck in AI loops, don't keep hitting the same wall.
Switch tools. Have one AI summarize what’s been tried, and another analyze what another got wrong.
We got out of endless debugging loops using this approach.
Switch tools. Have one AI summarize what’s been tried, and another analyze what another got wrong.
We got out of endless debugging loops using this approach.
6/8 - Component libraries from screenshots: Great concept, poor execution.
v0 extracted a brand’s design elements perfectly (for brand-aligned prototypes)... but couldn't build decent websites with them.
Sometimes the technical capability is there, but the design judgment isn't.
v0 extracted a brand’s design elements perfectly (for brand-aligned prototypes)... but couldn't build decent websites with them.
Sometimes the technical capability is there, but the design judgment isn't.
September 3, 2025 at 10:04 AM
6/8 - Component libraries from screenshots: Great concept, poor execution.
v0 extracted a brand’s design elements perfectly (for brand-aligned prototypes)... but couldn't build decent websites with them.
Sometimes the technical capability is there, but the design judgment isn't.
v0 extracted a brand’s design elements perfectly (for brand-aligned prototypes)... but couldn't build decent websites with them.
Sometimes the technical capability is there, but the design judgment isn't.
5/8 - We built a script to connect to our email, find a newsletter with upcoming events listings, and turn that info (incl "clicking" to the PDF event invites & OCRing the key details) into Google calendar events. AI’s amazing for building these micro-workflows
September 3, 2025 at 10:04 AM
5/8 - We built a script to connect to our email, find a newsletter with upcoming events listings, and turn that info (incl "clicking" to the PDF event invites & OCRing the key details) into Google calendar events. AI’s amazing for building these micro-workflows
4/8 - Academic research just got a massive upgrade.
We tested Elicit, Claude Deep Research, ChatGPT, Perplexity, Consensus, and Scite for lit reviews.
Result: 2+ weeks of research work done in 2 hrs.
Elicit wins by a landslide - proper academic formatting, methodology, citations
We tested Elicit, Claude Deep Research, ChatGPT, Perplexity, Consensus, and Scite for lit reviews.
Result: 2+ weeks of research work done in 2 hrs.
Elicit wins by a landslide - proper academic formatting, methodology, citations
September 3, 2025 at 10:04 AM
4/8 - Academic research just got a massive upgrade.
We tested Elicit, Claude Deep Research, ChatGPT, Perplexity, Consensus, and Scite for lit reviews.
Result: 2+ weeks of research work done in 2 hrs.
Elicit wins by a landslide - proper academic formatting, methodology, citations
We tested Elicit, Claude Deep Research, ChatGPT, Perplexity, Consensus, and Scite for lit reviews.
Result: 2+ weeks of research work done in 2 hrs.
Elicit wins by a landslide - proper academic formatting, methodology, citations
3/8 - BUT: Sometimes the most confident AI is the least trustworthy AI.
Browserbase MCP testing gave us a "95% confidence" our website passed all QA... when functionality was clearly broken.
Browserbase MCP testing gave us a "95% confidence" our website passed all QA... when functionality was clearly broken.
September 3, 2025 at 10:04 AM
3/8 - BUT: Sometimes the most confident AI is the least trustworthy AI.
Browserbase MCP testing gave us a "95% confidence" our website passed all QA... when functionality was clearly broken.
Browserbase MCP testing gave us a "95% confidence" our website passed all QA... when functionality was clearly broken.
2/8 - "At $9/month, Midjourney delivers design work that would cost hundreds to outsource"
This week we created: Holiday logo variations, Brand mascots in multiple styles, Photorealistic videos. The temperature controls for character consistency blew our minds 🤯
This week we created: Holiday logo variations, Brand mascots in multiple styles, Photorealistic videos. The temperature controls for character consistency blew our minds 🤯
September 3, 2025 at 10:04 AM
2/8 - "At $9/month, Midjourney delivers design work that would cost hundreds to outsource"
This week we created: Holiday logo variations, Brand mascots in multiple styles, Photorealistic videos. The temperature controls for character consistency blew our minds 🤯
This week we created: Holiday logo variations, Brand mascots in multiple styles, Photorealistic videos. The temperature controls for character consistency blew our minds 🤯
9/9
The AI world is weird: Premium pricing ≠ premium performance. Cultural/linguistic AI gaps are real. And sometimes the best solution is knowing when NOT to use AI.
Full deep-dive: medium.com/p/e516b7ce51...
The AI world is weird: Premium pricing ≠ premium performance. Cultural/linguistic AI gaps are real. And sometimes the best solution is knowing when NOT to use AI.
Full deep-dive: medium.com/p/e516b7ce51...
Week 17: The Great Model Divide: When AI Models Face Their Limits
Picture this: You run the same AI task through six different models and five of them completely fail while one nails it perfectly.
medium.com
August 11, 2025 at 6:48 AM
9/9
The AI world is weird: Premium pricing ≠ premium performance. Cultural/linguistic AI gaps are real. And sometimes the best solution is knowing when NOT to use AI.
Full deep-dive: medium.com/p/e516b7ce51...
The AI world is weird: Premium pricing ≠ premium performance. Cultural/linguistic AI gaps are real. And sometimes the best solution is knowing when NOT to use AI.
Full deep-dive: medium.com/p/e516b7ce51...
8/9
Key lesson: Model selection is becoming as critical as prompt engineering. Performance gaps aren't 10% better—they're "works perfectly" vs "doesn't work at all."
Key lesson: Model selection is becoming as critical as prompt engineering. Performance gaps aren't 10% better—they're "works perfectly" vs "doesn't work at all."
August 11, 2025 at 6:48 AM
8/9
Key lesson: Model selection is becoming as critical as prompt engineering. Performance gaps aren't 10% better—they're "works perfectly" vs "doesn't work at all."
Key lesson: Model selection is becoming as critical as prompt engineering. Performance gaps aren't 10% better—they're "works perfectly" vs "doesn't work at all."
7/9
Temperature settings matter WAY more than we realized. Most models default to "creative mode" (temp 1.0) when data tasks need "deterministic mode" (temp 0.2).
This one setting difference meant success vs total failure.
Temperature settings matter WAY more than we realized. Most models default to "creative mode" (temp 1.0) when data tasks need "deterministic mode" (temp 0.2).
This one setting difference meant success vs total failure.
August 11, 2025 at 6:48 AM
7/9
Temperature settings matter WAY more than we realized. Most models default to "creative mode" (temp 1.0) when data tasks need "deterministic mode" (temp 0.2).
This one setting difference meant success vs total failure.
Temperature settings matter WAY more than we realized. Most models default to "creative mode" (temp 1.0) when data tasks need "deterministic mode" (temp 0.2).
This one setting difference meant success vs total failure.
6/9
The breakthrough for both projects? Stop expecting AI to do everything.
Comic book: AI for visuals + story, humans for Hebrew text
Spreadsheets: AI for data logic, Python for edge cases
Hybrid workflows > pure AI
The breakthrough for both projects? Stop expecting AI to do everything.
Comic book: AI for visuals + story, humans for Hebrew text
Spreadsheets: AI for data logic, Python for edge cases
Hybrid workflows > pure AI
August 11, 2025 at 6:48 AM
6/9
The breakthrough for both projects? Stop expecting AI to do everything.
Comic book: AI for visuals + story, humans for Hebrew text
Spreadsheets: AI for data logic, Python for edge cases
Hybrid workflows > pure AI
The breakthrough for both projects? Stop expecting AI to do everything.
Comic book: AI for visuals + story, humans for Hebrew text
Spreadsheets: AI for data logic, Python for edge cases
Hybrid workflows > pure AI
5/9
"Sometimes the grandfather character had a beard, sometimes he didn't. In one memorable panel, he was somehow playing violin during a war scene."
AI consistency is... a work in progress 😅
"Sometimes the grandfather character had a beard, sometimes he didn't. In one memorable panel, he was somehow playing violin during a war scene."
AI consistency is... a work in progress 😅
August 11, 2025 at 6:48 AM
5/9
"Sometimes the grandfather character had a beard, sometimes he didn't. In one memorable panel, he was somehow playing violin during a war scene."
AI consistency is... a work in progress 😅
"Sometimes the grandfather character had a beard, sometimes he didn't. In one memorable panel, he was somehow playing violin during a war scene."
AI consistency is... a work in progress 😅
4/9
Meanwhile, Netta tried creating a Hebrew comic book from her grandfather's war stories.
AI nailed the story structure and panel composition. But Hebrew text? Complete gibberish. Every. Single. Time.
Meanwhile, Netta tried creating a Hebrew comic book from her grandfather's war stories.
AI nailed the story structure and panel composition. But Hebrew text? Complete gibberish. Every. Single. Time.
August 11, 2025 at 6:48 AM
4/9
Meanwhile, Netta tried creating a Hebrew comic book from her grandfather's war stories.
AI nailed the story structure and panel composition. But Hebrew text? Complete gibberish. Every. Single. Time.
Meanwhile, Netta tried creating a Hebrew comic book from her grandfather's war stories.
AI nailed the story structure and panel composition. But Hebrew text? Complete gibberish. Every. Single. Time.