Show notes: The fastest feedback loop wins in the AI era
Shoreline Episodes Ep. 4
EP 004 Season 01 · Expertise & mastery

The fastest feedback loop wins in the AI era.

A practical guide to building software from your phone, pairing expertise with AI, turning repeated prompts into reliable systems, choosing models by task, and staying open to new ways of working.

June 7, 2026 1:02:09 25 chapters Transcript-derived Expertise & mastery
The gist

Episode 4 asks what expertise is worth when capable AI can be directed from a phone. Enzo opens with a concrete example: while using his workout app at the gym, he speaks the problems he notices to Codex. The agent prepares changes while the experience is fresh, and he reviews them before they reach users.

The discussion moves from mobile building to the context AI needs, the changing value of specialized knowledge, repeatable systems, model costs, local AI, and Apple's on-device approach. It ends with training, the anterior mid-cingulate cortex (ACC), and the value of a beginner's mind. These notes turn the conversation into practical steps: shorten feedback loops, map missing context, test AI in a field you know well, choose models by task, and keep human review in the process.

Key takeaways

Eight practical ideas to put to work.

01

Your phone can shorten the feedback loop.

A voice-enabled coding agent can capture problems while you use a product, prepare a change, and leave the final review for you.

02

Missing context limits useful work.

An agent needs relevant goals, history, constraints, and customer information to handle difficult tasks. More data is useful only when it is current, permitted, and well organized.

03

AI changes the value of expertise.

A skilled AI user can now approach some specialist tasks much faster. Deep experience still matters, but experts also need the willingness to test new tools.

04

Expertise and AI are strongest together.

Experienced people bring judgment, relationships, and private context that a general model does not have. AI can help them apply those advantages more quickly.

05

Turn repeated prompts into a system.

A reliable workflow connects sources, instructions, checks, and approval steps. Build it once, improve it with use, and stop recreating the process for every task.

06

Choose model spend by value and visibility.

People close to the work can see when a more capable model creates value. Leaders farther away need routing, cost controls, and outcome measures—not raw token counts.

07

Expert judgment remains a quality filter.

An expert can compare many AI outputs and recognize which ones are accurate, useful, or original. AI increases the options; experience helps select the best.

08

Keep a beginner's mind.

New tools often make experienced people feel unskilled again. Curiosity and a willingness to start over make it easier to learn what the technology can actually do.

Tactical guidesWhat to try next
Guide 01

Capture product feedback from your phone.

When the user and builder are the same person, a mobile coding agent can capture feedback at the moment it occurs.

  1. Put your coding agent into voice mode so it keeps listening after you switch apps.
  2. Open a safe test build and use the product the way a real user would.
  3. Describe each problem as you experience it: slow notifications, awkward steps, or small usability issues.
  4. Ask the agent to prepare the change while you return to what you were doing.
  5. Review the code changes and test the result before anything reaches real users.
Guide 02

Map the context your agent is missing.

Give an agent the information needed for one decision instead of trying to capture everything at once.

  1. List what an agent would need to know to make your decisions: roadmap, customers, constraints, and history.
  2. Include useful details from calls, reading, and in-person conversations that were never documented.
  3. Choose which approved devices, notes, and systems may provide that context, and set a refresh schedule.
  4. Separate sensitive information that must stay local from data that may be sent to a cloud model.
  5. Test one task, identify the missing context that caused the failure, and close that gap first.
Guide 03

Test AI in a field you know well.

Use your own expertise to measure where a model is helpful, shallow, or confidently wrong.

  1. Choose a domain you know deeply, such as your sport, craft, or professional function.
  2. Ask detailed questions and separate correct, useful answers from confident mistakes.
  3. Present a poor plan and see whether the model challenges it or simply agrees.
  4. Use the results to set appropriate review requirements in areas you cannot verify yourself.
  5. Repeat the test after major model updates because capabilities and weaknesses change.
Guide 04

Turn a prompt into a system.

Will's client proposal shows the value of a repeatable process with reliable sources, checks, and review.

  1. Find a specialized task you repeat by hand, such as proposals, reports, or research.
  2. Write down where the information lives, what counts as accurate, and how you check it.
  3. Connect those steps so the AI follows the same process from source material to finished draft.
  4. Plan for more work at the beginning and improve the system through repeated tests.
  5. Keep proofreading and approval steps. Trust in the process should never replace human review.
Guide 05

Match model spend to the task.

People closest to the work can judge when a more capable model is worth the cost. Others need clearer controls and evidence.

  1. If you work directly with the AI, record where a stronger model materially improves the result.
  2. If you manage the budget from a distance, use routing rules and review outcome data.
  3. Send extraction, drafts, and simple questions to lower-cost or open-source models.
  4. Reserve the most expensive models for difficult or high-stakes problems, not routine interface or copy changes.
  5. Revisit the episode's “blowtorch” test: using the strongest model for every task is often more power than the job needs.
Guide 06

Stand up a local model fallback.

A model that runs on your own computer can support private, offline, or high-volume work without relying on a cloud service.

  1. Install a local runtime such as Ollama and download a capable open-source model.
  2. Identify work that should not leave the device, including private notes, sensitive data, and offline tasks.
  3. Compare it with a recent leading cloud model so you know where the local option falls short.
  4. Use it for lower-cost, high-volume, or no-connectivity tasks when the quality is sufficient.
  5. Test again each quarter because open-source model performance changes quickly.
Guide 07

Create an expert review process.

Experience helps people distinguish accurate, useful output from generic filler.

  1. Generate many candidates instead of one, then judge them like a craftsperson.
  2. Keep only the strongest two or three and remove the rest.
  3. Explain why the selected outputs work so the standard becomes repeatable.
  4. Add those examples and rules to the system's instructions.
  5. Ask whether a real expert would feel the difference after five seconds, five minutes, and five uses.
Guide 08

Practice being a beginner.

The closing discussion connects difficult physical practice with the patience needed to learn unfamiliar AI tools.

  1. Choose a difficult practice with clear feedback: the bar clears or it does not.
  2. Experience the discomfort of being a beginner in a low-stakes setting.
  3. Use that patience when a new tool feels confusing at first.
  4. Treat each new model or workflow as practice, not a threat to your expertise.
  5. Train deliberately rather than endlessly: choose the right repetitions and allow time for recovery.
ChaptersTimestamp map
00:00

Testing the workout app with friends

Enzo opens from Pacifica and explains how he asks friends to use his workout app and share direct feedback.

↗ 00:40

Building with Codex Mobile and voice mode

Voice mode continues listening after Enzo switches apps, letting him describe product changes as he notices them.

↗ 01:41

Preparing app changes between gym sets

Enzo describes a problem at the gym, asks the agent to prepare a fix, and returns to his workout.

↗ 02:45

Speculating about an OpenAI device

The hosts imagine an always-available device that can capture context and interact through voice and simple visual cues.

↗ 04:27

Can AI build the product without enough context?

A head of engineering says AI cannot build the product today, then considers which missing context would change the answer.

↗ 06:31

Rethinking the product-to-engineering process

The hosts discuss how AI is changing the handoff between product, design, and engineering teams.

↗ 08:24

Specialist expertise versus AI coordination

The hosts ask whether the stronger advantage is deep subject expertise or the ability to direct AI across several fields.

↗ 10:15

How AI changes the expert advantage

The discussion considers how capable AI users can approach specialist tasks faster—and why experts still need to test the tools themselves.

↗ 13:21

Yann LeCun and the 1905 relativity example

The hosts use a model trained only on knowledge available before 1905 to discuss whether language models can produce discoveries beyond their training data.

↗ 14:24

Expert judgment as a quality filter

An expert can compare many AI outputs and quickly identify the few that are accurate, compelling, and useful.

↗ 15:45

Combining expertise with AI

The hosts argue that experience paired with AI is stronger than either alone, then discuss Elon's prediction about the scale of machine intelligence.

↗ 16:40

Networks and private data as lasting advantages

The hosts identify professional relationships and private data as forms of expertise that a general model cannot easily reproduce.

↗ 19:56

Building a reliable proposal system

Will explains how a structured process helped him produce a complex client proposal, while the client response provided an important quality check.

↗ 23:22

Measuring outcomes instead of token use

The hosts support generous experimentation when it serves the work, but recommend measuring users, product quality, and business value.

↗ 28:37

Open-source versus frontier models

The hosts compare lower-cost open models such as DeepSeek and MiniMax with the latest paid models.

↗ 29:53

Model routing and who should optimize

People close to the work can judge when a leading model adds value. Leaders farther away need clearer routing and cost controls.

↗ 35:24

Local models, Ollama, and offline AI

Running a model on a laptop makes useful AI available for private work and places without an internet connection.

↗ 38:01

The economics of frontier models

The hosts discuss falling model prices and the high cost, competition, and pressure facing companies that develop frontier models.

↗ 39:48

Apple, Siri, and on-device AI

Ahead of WWDC, the hosts consider how a local model could handle calendar and message tasks while using the cloud only when needed.

↗ 42:01

Using model routing to control costs

The hosts discuss how routing simple tasks to lower-cost models can reduce spending without lowering the quality of the result.

↗ 43:37

How downstream products reveal model strengths

The hosts argue that tool builders and customers often discover valuable uses after a frontier lab releases a new model.

↗ 44:50

Alex Karp, Palantir, and the dead hang

The conversation moves from Palantir's sales strategy to Alex Karp's reported five-and-a-half-minute dead hang.

↗ 47:39

Training hard versus training well

A longevity story prompts the hosts to question whether daily dead hangs without recovery were helping their actual goal.

↗ 51:04

The ACC, effort, and resilience

The hosts discuss claims about the anterior mid-cingulate cortex (ACC), pain tolerance, and the capacity to persist through difficult work.

↗ 53:03

The Monet test and the beginner's mind

A real Monet is criticized as AI-generated work, leading to a discussion about fear, judgment, and the willingness to begin again.

↗
The stackTools & concepts

Codex Mobile + voice-mode building

Build from your phone · real-time iteration

The opening workflow turns real product use into immediate feedback: voice mode keeps listening while you switch apps, describe a problem, and ask the agent to prepare a change for later review.

Codex docs Claude Code Ollama
Coding agent

Codex

OpenAI's coding agent for working from mobile or the command line, including the voice-enabled product-feedback loop discussed in the episode.

Open Codex docs
Coding agent

Claude Code

Anthropic's coding agent for exploring codebases, editing files, running commands, and completing longer development tasks.

Open Claude Code docs
AI editor

Cursor

An AI-powered code editor that supports several models and gives teams more control over model choice and cost.

Open Cursor docs
Local models

Ollama

A tool for running open-source models on your own computer for offline, sensitive, or high-volume work.

Open Ollama
Model routing

OpenRouter

A service for sending each task to an appropriate model based on quality, cost, speed, and risk.

Open OpenRouter docs
Model routing

Factory (Droid)

The engineering platform referenced in a discussion about using several models while controlling costs.

Open Factory
Open source

DeepSeek

An open-source model provider discussed as a lower-cost alternative to the latest paid models.

Open DeepSeek
Company brain

Notion

A home for notes and shared knowledge that can also choose among models behind the scenes.

Open Notion
AI devices

Thinking Machines Lab

Mira Murati's lab, mentioned in relation to real-time voice and vision demos for context-aware AI.

Open Thinking Machines
Mastery

Mastery

Robert Greene's book supports the episode's discussion of deliberate practice, expertise, and mastery.

Open publisher page
Perspective

Yann LeCun

His argument about the limits of language models provides context for the episode's debate about knowledge and original discovery.

Find the talks
Training science

Huberman Lab × Goggins

The source the hosts cite for their discussion of the ACC, difficult effort, and high-intensity training.

Open Huberman Lab
Links & references
EpisodeWatch the full episode — YouTube↗ GuideCapture product feedback from your phone — a voice-enabled workflow↗ GuideMap the context your agent is missing — close one gap at a time↗ GuideTest AI in a field you know well — set the right review level↗ GuideTurn a prompt into a system — connect sources, checks, and review↗ GuideMatch model spend to the task — leading cloud models versus open source↗ GuidePractice being a beginner — build patience for unfamiliar tools↗ IndexReturn to the Shoreline resource library — homepage↗
“

AI changes the value of knowledge, but it does not replace context, relationships, reliable systems, or expert judgment. The opportunity is to combine those strengths with faster tools while staying willing to learn.

Shoreline Ep. 4 · distilled operating principle
Listener checklist
Day 1

Capture feedback from your phone.

Use voice mode to describe one real product problem, then review and test the agent's proposed change.

Day 2

Map missing context.

Write down what an agent needs to make one real decision, then close the biggest information gap.

Day 3

Test AI against your expertise.

Ask detailed questions about a field you know well and use the results to set an appropriate review level.

Day 4

Systematize one task.

Turn one repeated job into a connected process with reliable sources, checks, and an approval step.

Day 5

Route your model spend.

Decide which tasks need the most capable model, which can use open source, and which should run locally.

Day 6-7

Practice beginner's mind.

Choose one difficult activity with clear feedback, then bring that patience to the next unfamiliar tool.