Breaking it on purpose
How do you find the ways an AI product fails before your users do?
Timeline
Compressed: 2022–24Building at work, not hereZoomed in: Oct–Nov 2026
What AI changed for me
What I built
Tools
What I learnedCourses
TalksLearned by buildingCertifications
ChatGPT launches
Large language models become a mass-market product overnight.
source (opens in a new tab)GPT-4
Claude launched the same day.
Fed into: EXP-000 Ambient care-plan scribe
source (opens in a new tab)Long context and JSON mode
GPT-4 Turbo reads about 300 pages at once, and structured output becomes practical.
source (opens in a new tab)Models that reason first
OpenAI's o1 thinks before answering: big gains on maths and code.
source (opens in a new tab)AI that uses a computer
Claude operates a screen, keyboard and mouse: AI starts acting, not just answering.
source (opens in a new tab)Model Context Protocol
An open standard for connecting AI to tools and data.
source (opens in a new tab)Coding agents arrive
Claude Code previews: an agent that works in your codebase, not just suggests lines.
Fed into: EXP-005 Fold, EXP-006 Fold's review panel, EXP-007 This lab, rebuilt
source (opens in a new tab)NHS guidance on ambient scribes
NHS England publishes guidance on AI-enabled ambient scribing products in health and care.
source (opens in a new tab)Image models get precise
Google's "Nano Banana" focuses on consistent characters and precise edits.
Fed into: EXP-005 Fold, A recipe for honest pictures: how to avoid AI image slop without copyright trouble
source (opens in a new tab)A frontier model held back for safety
Anthropic gives Claude Mythos Preview only to cyber defenders, through Project Glasswing.
source (opens in a new tab)Mythos-class models released
Claude Fable 5 becomes available to everyone, and Claude Mythos 5 to Glasswing partners only.
source (opens in a new tab)Gemini 4, defenders first
Google also releases its new frontier model to cyber defenders first, as Anthropic did with Mythos.
source (opens in a new tab)ChatGPT appears
GPT-4o inside a product; coding with Cursor
Claude holding my context
My personal context folders for Claude, which carry my projects and preferences across tools.
An agent builds a whole site
Agent teams and reviewers
Fold's review panel, then agent teams and reviewers on later builds.
EXP-000 Ambient care-plan scribe
Before care had scribes: drafting care plans with AI in 2022
Read the logEXP-001 emilybuilds on Notion + Super
Deciding not to build: when no-code is the right call, and when it stops being one
Read the logEXP-004 emilysas.com
The bottleneck moved: where the time goes when an AI agent writes the code
Read the logEXP-006 Fold's review panel
Never mark your own homework: independent AI reviewers for a solo builder
Read the logEXP-007 This lab, rebuilt
Nothing goes live unread: building a public site with AI agents, without losing control
Read the logEXP-009 Mullo
The best page, the worse journey: when an AI tool's quality isn't worth the effort
Read the logBack end and automation: automations replaced by my own code
From wiring services together in Zapier to a back end written in code.
See the stackThinking and planning: plans now live in the repo
From a chat that wrote the instructions I pasted into the editor, to a planning session that writes specs into the repository for the builder.
See the stackResearch: research moved to its own tool
From asking the same chat to a separate research tool.
See the stackWriting the code: the agent now works on its own
From an editor whose AI wrote most of the code but which I had to stop from running scripts, to an agent trusted to build, run and test in its own session.
See the stackChecking the work: one big system became small reviewers
From one evaluation system I couldn't keep running, to small reviewers that each check one thing.
See the stackData: off a spreadsheet-style database
From a spreadsheet-style database to hosted databases and private repositories.
See the stackDesign: a design tool, plus layout tests
From layouts I had to fix for phones after launch, to a design tool and layout tests at phone and iPad sizes.
See the stackEmail and sign-ups: sign-ups go straight to the newsletter
Sign-ups go straight into the newsletter tool, with no automation in between.
See the stackModels inside my products: one model became one per job
From one model for everything to a different model for each job.
See the stackWebsites: a hosted platform became my own site
From a site I had less control over to one I own, edited in a content manager.
See the stackMaking Friends with Machine Learning
Read the learning noteMtPCon London
Read the learning noteDecision Making in the Age of AI
Read the learning noteDesigning AI drafts for professional review
Learned building EXP-000Evaluating AI output with an AI judge
Learned building EXP-002How I evaluate AI output v1
Read the guideDirecting a coding agent end to end
Learned building EXP-004Designing independent AI reviewers that fail safe
Learned building EXP-006Fact-checking AI-written content against sources
Learned building EXP-007Running a multi-agent build with human sign-off at each gate
Learned building EXP-007How I evaluate AI output v2
Read the guideMachine Learning Foundations for Product Managers
View certificate (opens in a new tab)Chat
Seeing, hearing, reasoning
Agents
Gated frontier
Chat
- Nov 2022ChatGPT launches
- Nov 2022ChatGPT appears
- Dec 2022EXP-000 Ambient care-plan scribe
- Dec 2022Designing AI drafts for professional review
- about 2023Making Friends with Machine Learning
- Mar 2023GPT-4
- Jun 2023Machine Learning Foundations for Product Managers
- Jul 2023Llama 2
- Nov 2023Long context and JSON mode
- Mar 2024Claude 3
- Apr 2024Decision Making in the Age of AI
Seeing, hearing, reasoning
- May 2024GPT-4o
- Aug 2024EU AI Act in force
- Sept 2024Models that reason first
Agents
- Oct 2024AI that uses a computer
- Nov 2024Model Context Protocol
- Jan 2025DeepSeek R1
- Feb 2025"Vibe coding" coined
- Feb 2025Coding agents arrive
- Apr 2025NHS guidance on ambient scribes
- Jun 2025EXP-002 Launchling
- about 2025Back end and automation: automations replaced by my own code
- Jun 2025Evaluating AI output with an AI judge
- Jul 2025EXP-001 emilybuilds on Notion + Super
- Jul 2025GPT-4o inside a product; coding with Cursor
- Aug 2025GPT-5
- Aug 2025Image models get precise
- Sept 2025How I evaluate AI output v1
- Building at work, not here
- Jan 2026Thinking and planning: plans now live in the repo
- Mar 2026Research: research moved to its own tool
Gated frontier
- Apr 2026A frontier model held back for safety
- Jun 2026Writing the code: the agent now works on its own
- Jun 2026Mythos-class models released
- Jun 2026MtPCon London
- Jul 2026Claude holding my context
- Sept 2026Gemini 4, defenders first
- Oct 2026Analytics: cookie-free analytics
- Oct 2026Checking the work: one big system became small reviewers
- Oct 2026Data: off a spreadsheet-style database
- Oct 2026Design: a design tool, plus layout tests
- Oct 2026Email and sign-ups: sign-ups go straight to the newsletter
- Oct 2026Hosting: sites moved; apps stayed
- Oct 2026Models inside my products: one model became one per job
- Oct 2026An agent builds a whole site
- Oct 2026EXP-004 emilysas.com
- Oct 2026Directing a coding agent end to end
- Oct 2026EXP-005 Fold
- Oct 2026Agent teams and reviewers
- Oct 2026EXP-006 Fold's review panel
- Oct 2026Designing independent AI reviewers that fail safe
- Oct 2026EXP-007 This lab, rebuilt
- Oct 2026Fact-checking AI-written content against sources
- Oct 2026Running a multi-agent build with human sign-off at each gate
- Oct 2026EXP-009 Mullo
- Oct 2026How I evaluate AI output v2
- Nov 2026Websites: a hosted platform became my own site
The same timeline, as a list
- ChatGPT launches
- ChatGPT appears
- EXP-000 Ambient care-plan scribe
- Designing AI drafts for professional review
- Making Friends with Machine Learning
- GPT-4
- Machine Learning Foundations for Product Managers
- Llama 2
- Long context and JSON mode
- Claude 3
- Decision Making in the Age of AI
- GPT-4o
- EU AI Act in force
- Models that reason first
- AI that uses a computer
- Model Context Protocol
- DeepSeek R1
- "Vibe coding" coined
- Coding agents arrive
- NHS guidance on ambient scribes
- EXP-002 Launchling
- Back end and automation: automations replaced by my own code
- Evaluating AI output with an AI judge
- EXP-001 emilybuilds on Notion + Super
- GPT-4o inside a product; coding with Cursor
- GPT-5
- Image models get precise
- How I evaluate AI output v1
- Thinking and planning: plans now live in the repo
- Research: research moved to its own tool
- A frontier model held back for safety
- Writing the code: the agent now works on its own
- Mythos-class models released
- MtPCon London
- Claude holding my context
- Gemini 4, defenders first
- Analytics: cookie-free analytics
- Checking the work: one big system became small reviewers
- Data: off a spreadsheet-style database
- Design: a design tool, plus layout tests
- Email and sign-ups: sign-ups go straight to the newsletter
- Hosting: sites moved; apps stayed
- Models inside my products: one model became one per job
- An agent builds a whole site
- EXP-004 emilysas.com
- Directing a coding agent end to end
- EXP-005 Fold
- Agent teams and reviewers
- EXP-006 Fold's review panel
- Designing independent AI reviewers that fail safe
- EXP-007 This lab, rebuilt
- Fact-checking AI-written content against sources
- Running a multi-agent build with human sign-off at each gate
- EXP-009 Mullo
- How I evaluate AI output v2
- Websites: a hosted platform became my own site
Threads
Breaking it on purpose
Knowing if it's any good
- EXP-002 LaunchlingStart here
- From Side Project to Stress Test: Using AI to Evaluate TinyLaunch
- How I Used AI Pair Programming to Build a Robust Evaluation System (Without Vibe Coding)
- EXP-009 Mullo
- How I evaluate AI output
- Diverse by prompt, biased by role: how AI illustration passes society's stereotypes on
Checking the checkers
- Never let the builder mark its own homework Each reviewer starts with a fresh context, so it hasn't absorbed the reasoning that produced the work and doesn't share its blind spots.
- Decide which way every check fails A check that crashes and reports "no errors" is worse than no check; when a check can't run, it should say "couldn't be checked", not pass.
- Make every check say what it didn't check A reviewer rewarded only for finding problems will find problems, so "nothing wrong" counts as a proper answer, and every report lists what it found sound and what it didn't look at.
Building with agents
- Trying (and Failing) to Send Emails from My Domain
- TinyLaunch Update: From Framer MVP to Full React Build — Lessons, Challenges & Next Steps
- TinyLaunch Is ‘Launched’: From Framer to Fully-Coded Product
- How I Used AI Pair Programming to Build a Robust Evaluation System (Without Vibe Coding)
- EXP-004 emilysas.comStart here
- EXP-007 This lab, rebuilt
- A recipe for honest pictures: how to avoid AI image slop without copyright trouble
What it knows and can reach
Seeing the cost
Designing for trust
- EXP-000 Ambient care-plan scribeStart here
- Satirical Startups & French Pirate Founders: How Launchling Handles Absurdity by Design
- EXP-005 Fold
- EXP-006 Fold's review panel
- A recipe for honest pictures: how to avoid AI image slop without copyright trouble
- EXP-009 Mullo
- Diverse by prompt, biased by role: how AI illustration passes society's stereotypes on
Is it worth building
- EXP-000 Ambient care-plan scribe
- EXP-002 Launchling
- EXP-001 emilybuilds on Notion + Super
- Building My AI Product Portfolio Site with Notion + Super
- TinyLaunch Is ‘Launched’: From Framer to Fully-Coded Product
- Why I Built TinyLaunch – Helping Tiny Ideas Take Flight
- From TinyLaunch to Launchling
- Post-Launch Polish: What I’ve Been Improving on TinyLaunch
- From MVP to Momentum: A Week of Big Leaps for Launchling
- EXP-005 FoldStart here
Knowing if it's any good
What does “good” mean for output nobody can mark right or wrong?
Checking the checkers
When AI checks AI, what checks the checker?
Building with agents
Where do a builder's time and judgement go when agents do the typing?
What it knows and can reach
What should an AI know and be able to reach, and who decides?
Seeing the cost
When is a cheaper model the better product decision?
Designing for trust
How do you design AI that people trust exactly as much as they should?
Is it worth building
When building is this cheap, how do you decide what's worth building?
The thread map as a list
Breaking it on purpose
Knowing if it's any good
Checking the checkers
Building with agents
- Trying (and Failing) to Send Emails from My Domain
- TinyLaunch Update: From Framer MVP to Full React Build — Lessons, Challenges & Next Steps
- TinyLaunch Is ‘Launched’: From Framer to Fully-Coded Product
- How I Used AI Pair Programming to Build a Robust Evaluation System (Without Vibe Coding)
- EXP-004 emilysas.com
- EXP-007 This lab, rebuilt
- A recipe for honest pictures: how to avoid AI image slop without copyright trouble
What it knows and can reach
Seeing the cost
Designing for trust
- EXP-000 Ambient care-plan scribe
- Satirical Startups & French Pirate Founders: How Launchling Handles Absurdity by Design
- EXP-005 Fold
- EXP-006 Fold's review panel
- A recipe for honest pictures: how to avoid AI image slop without copyright trouble
- EXP-009 Mullo
- Diverse by prompt, biased by role: how AI illustration passes society's stereotypes on
Is it worth building
- EXP-000 Ambient care-plan scribe
- EXP-002 Launchling
- EXP-001 emilybuilds on Notion + Super
- Building My AI Product Portfolio Site with Notion + Super
- TinyLaunch Is ‘Launched’: From Framer to Fully-Coded Product
- Why I Built TinyLaunch – Helping Tiny Ideas Take Flight
- From TinyLaunch to Launchling
- Post-Launch Polish: What I’ve Been Improving on TinyLaunch
- From MVP to Momentum: A Week of Big Leaps for Launchling
- EXP-005 Fold
How this page is made
Built automatically from each build's start date, each learning note, skill, certification and guide version, the dates of the tools I switched to, the events in "What AI changed for me" and the AI milestones. Adding an item updates the timeline the next time the site is published.