What changed how we build

2026 field notes Learnings

p10q grew from a simple habit: learn from people doing the work, test the ideas in real projects, and keep only what makes the next build better.

By p10q ~15 min read
Scroll to begin
Learning fieldSources into practice
95sources kept
Observe the work
Shape with judgment
Protect the boundary
Turn sources into a method.
  1. Observe
  2. Direct
  3. Verify
  4. Distill
01 / Learn

Follow the work, not the hype.

The useful sources are usually close to practice. We look for builders showing their process, failures, constraints, and finished evidence. Talks and technical explainers provide depth. A quiet Hacker News view provides breadth without turning the feed into the work.

Structured courses and books provide a slower foundation. A source earns a permanent place only when it changes how we plan, build, verify, protect data, or explain a product to a customer.

Selection ruleKeep the sources that change behavior.
02 / Direct

The bottleneck moved to judgment.

Better coding agents can inspect a repository, propose work, implement changes, run checks, and recover from errors. People still choose the direction, define what good means, and review consequential outcomes.

Intent and tradeoffs matter more than a long sequence of instructions. Tests, screenshots, budgets, and direct review give the loop something real to optimize. Durable repository instructions preserve the constraints and decisions worth carrying into the next session.

DirectionGive agents room inside a goal you still own.
03 / Loop

Make every build improve the next one.

The studio workflow is a closed loop. Explore the customer problem and current system. Frame the intended outcome and evidence. Build inside clear boundaries. Verify the real artifact. Distill the useful decisions back into the repository.

The shipped result matters, but the checks, methods, and decisions left behind are what compound. Each finished project should make the next project easier to direct and harder to misunderstand.

Compounding workLeave the next session a better system than the one you inherited.
04 / Protect

Draw the boundary before granting power.

An agent becomes risky when private data, untrusted input, and an outbound action meet without a meaningful approval point. Remove one leg, isolate the work, or require a person to approve the consequential step.

The customer should be able to see what evidence supports an action, where their data can travel, and how to recover when the system is wrong. Logs and reversibility are part of the product, not cleanup work.

Trust boundaryMake power visible before making it automatic.
05 / Ship

Turn the method into useful products.

Record Maps makes fragmented public records researchable while keeping every fact attached to its source, definition, and observation date. One product protects attention. The other protects evidence.

Product principleUse software to keep consequential work legible.
06 / Sources

Keep a useful source library.

Tools change quickly, so the shelf combines current practice with durable foundations. We return to people showing real agent workflows, designers explaining how judgment works, and courses or books that make the underlying systems easier to reason about.

The collection below keeps the full verified video watchlist and the slower references we return to. Grouping each item by the decision it helps with makes a large archive useful without turning it back into a feed.

Learning systemKeep the reason each source earned its place.
The source shelf

What we kept from the year.

Six highlighted talks are useful places to begin. The complete library below keeps all 68 verified videos plus the courses, books, sites, communities, and essays that changed how p10q builds.

Watch
Cursor

A computer for every agent

Agents become more trustworthy when they can operate the finished application and return recorded evidence of what they tested.

Watch
Latent Space

The AI Architect: Bret Taylor

Strong AI products keep customer understanding, product judgment, and engineering close together as verification becomes more valuable than code generation.

Video library

Every video has one reason it stayed on the list. Grouping by the decision it helps with makes the archive easier to revisit than a date-ordered feed.

17 videos

Directing coding agents

Planning, review, specifications, and the human judgment that keeps implementation pointed at the right outcome.

15 videos

Harnesses, context, and memory

The surrounding systems that give a model useful context, tools, durable state, and repeatable feedback.

06 videos

Autonomy, orchestration, and trust

Multi-agent coordination, visible authority, security boundaries, and recovery when automated work goes wrong.

09 videos

Design and visual craft

References, critique, taste, motion, and the feedback loops that keep generated interfaces from becoming generic.

12 videos

Products, founders, and organizations

How cheaper implementation changes company shape while product direction, customer understanding, and costs remain consequential.

09 videos

AI and software foundations

Models, systems, evaluations, architecture, benchmarks, and the technical history underneath current agent workflows.

Courses, books, and working references

These slower resources provide the foundations and practical habits that remain useful when the current agent tools change.

02 resources

General

A small daily signal for depth and breadth without letting the feed become the work.

  • YouTube Conference talks, tutorials, and technical explainers
  • hckrnews.com Curated HN front page, less noise
04 resources

Design & Motion

Courses and teachers that build visual judgment before adding more software.

  • Flux Academy Web design & freelancing
  • School of Motion Motion design courses covering After Effects, Cinema 4D, and animation principles
  • Learn UI Design Erik Kennedy's practical, opinionated guide to UI and UX fundamentals
  • Josh W. Comeau Tutorials and interactive articles about CSS and React
02 resources

Web Development

Durable mental models for front-end systems that continue to help as frameworks change.

  • Joy of React Josh Comeau's React course, with an emphasis on mental models and syntax
  • Frontend Masters Expert-taught courses on JavaScript, React, TypeScript, Node, CSS, and related topics
02 resources

Architecture

Distributed-systems lessons and system-design explanations grounded in operating real services.

  • Amazon Builders' Library Distributed systems patterns documented by Amazon engineers
  • YouTube System design channels, re:Invent talks, and architecture explainers
06 resources

ML & AI

Structured ways to understand models, agents, memory, planning, and the field around them.

  • swyx/ai-notes A community-maintained knowledge base for AI papers, trends, and tools
  • DeepLearning.AI Andrew Ng's AI & machine learning courses
  • YouTube Andrej Karpathy, 3Blue1Brown, Yannic Kilcher
  • Deep Learning with Python François Chollet's hands-on guide to deep learning with Keras
  • The Welch Labs Illustrated Guide to AI A 376-page visual guide from perceptrons through attention and image and video generation, with exercises, code, and companion videos
  • An Illustrated Guide to AI Agents Maarten Grootendorst and Jay Alammar explain tools, memory, planning, reasoning models, multimodal systems, and multi-agent collaboration through visual examples
02 resources

3D

Project-based Blender training that moves from first exercises into complete visual work.

  • Polygon Runway Beginner Blender tutorials built around stylized projects
  • CG Cookie Blender training in modeling, texturing, and animation
02 resources

Management

Practical writing and reading lists for the human systems around engineering work.

07 resources

Working practices

Small tools and essays that made agent feedback, memory, community, and trust boundaries more concrete.

  • Lavish Annotate a local agent-built interface and turn visual feedback into the next implementation pass.
  • Claude-Mem Distill completed sessions into durable project memory instead of retaining entire conversations.
  • AI Builder Club A community for comparing real AI-building workflows with people actively shipping products.
  • DHH on parallel agent work A practical account of the attention limits that appear when many terminal agents run at once.
  • The lethal trifecta A clear test for the dangerous combination of private data, untrusted input, and outbound communication.
  • Tidy First example Structure the environment so the next consequential change becomes smaller and easier to review.
  • Do smart things Give capable agents room to propose useful work while keeping risky actions behind human review.