Talk resources / Leading Deeper with AI

Reuse the process.
Keep your judgment.

Reusable instructions for recurring work: what your agent should read, return and leave for you to decide.

Choose an agent skill
Questions lead to a proposed patch and replay test. A human approves, revises or rejects it before shared knowledge changes.QuestionsAI patchYour review

A question can improve the next answer.

Revisit the talk

Slides & takeaways

Browse the slides and the point behind each one. Choose a view to see each reveal, or open the presentation for the animations.

Leverage and ownership

Leading Deeper with AI

AI can make technical exploration easier and strengthen your leadership. Use that access where your contribution improves the outcome, while protecting the work others should own and grow through.

Read all 27 takeaways
  1. Leading Deeper with AI

    AI can make technical exploration easier and strengthen your leadership. Use that access where your contribution improves the outcome, while protecting the work others should own and grow through.

  2. Someone else's air

    Exploring unfamiliar systems used to depend heavily on your team’s expertise and context scattered across documentation and conversations. That support made exploration possible and shaped how far you could go.

  3. Free to go deeper

    AI makes it easier to investigate unfamiliar code, test ideas and answer first questions yourself. Bring that deeper understanding back to your team as a more informed perspective.

  4. Already in the water

    The studies suggest managers are becoming more hands-on. Higher code output is an activity measure; leaders still need to choose contributions that improve outcomes for customers.

  5. Leverage Grid

    Map activities by leverage and ownership. Automate recurring preparation, make room for your own technical learning and leadership, and protect work your team can own, grow through and use to deliver customer value. Leverage depends on the person.

  6. The experiences behind the grid

    Build judgment through real practice, then carry it between your own work, your team and the wider business. These experiences can overlap and strengthen one another.

  7. My practice pool

    Choose a real problem with people who can give you useful feedback. Softball IQ gave Jasmine a place to build, test and refine AI-assisted work with her daughter’s team.

  8. Learning by building

    Build reusable context with retrieval-augmented generation (RAG) and structured instructions (context engineering). Review feedback and propose improvements through recurring checks and expert review; give agents more autonomy with goals, constraints, acceptance checks, stop points and inspectable results.

  9. Leverage Grid

    Daily Debrief brings together sourced priorities and follow-through. Choose Where Attention Belongs recommends one small automation to test; you decide what matters and where to invest the attention you recover.

  10. For my team

    A smaller team still needed to support the same systems. Deeper technical involvement became a way to rebuild shared knowledge and support the team through that change.

  11. February 2026

    After Block’s workforce reduction, the team needed emotional support and a way to cover lost specialisms. Leadership meant making room for people, then working together to stabilize systems and fill knowledge gaps.

  12. What was already written

    Start with the context that already exists: code, documentation, conversations and tickets. Structure it into a shared, navigable knowledge source with evidence and explicit gaps, then keep reviewing and improving it.

  13. Questions exposed the gaps

    Working alongside the team exposes missing context. Capture explanations once in shared knowledge so the next person, or their agent, can begin with an answer instead of reconstructing the same history.

  14. Answers where people ask

    Make each engineer’s support answer improve the next one. Perch reviews threads every two weeks, uses a fixed rubric to find knowledge gaps and proposes updates, with checks on the reliability of its assessments.

  15. A question becomes a proposed update

    Turn repeated support corrections into proposed guidance. Replay the original questions, compare the answers with engineers’ corrections, and show the evidence and checks so a person can review and merge the update.

  16. How the channels compared

    Channels with sustained Perch reviews improved most; the comparison was observational and included partial help. Separately, 175 useful first passes suggest 30–60 response hours before review and upkeep, using assumed rather than measured time savings.

  17. Where the attention goes

    Use shared knowledge and AI support to reduce recurring work. Apply your technical understanding to align people and stakeholders, while engineers own delivery and build expertise through the work.

  18. An earlier target. More shared knowledge.

    The hardware-constrained migration now targets early 2027, earlier than planned, and remains in progress. Two engineers have core platform knowledge, with performance at parity, uninterrupted migrated workflows and optimization still ahead.

  19. Beyond my team

    The skills and judgment developed with your team can help elsewhere in the business. Bring experience structuring context and reviewable work, and build alongside the people who know the domain.

  20. Less effort reconstructing work

    A parallel finance pilot drafts capitalizable engineering time from scoped activity and supporting evidence; engineers review, correct and attest, while finance retains accounting judgment. The pilot is testing whether this reduces reconstruction and improves review.

  21. Nitrogen narcosis

    Nitrogen narcosis can make a diver feel confident while impairing judgment. Softball IQ taught the same leadership lesson: back confidence with checks, guardrails and review by someone who knows the domain.

  22. Low visibility

    In low visibility, divers need help staying oriented. In finance, domain experts provided that orientation: they shaped useful evidence and kept the accounting judgment.

  23. Task loading

    Too many simultaneous tasks can reduce a diver’s awareness. When the CI migration needed coding, coordination and stakeholder management, engineers owned delivery while Jasmine focused on aligning people and timing.

  24. Equipment upkeep

    Diving equipment needs ongoing care; so do AI tools. Perch still needs review and upkeep as the tools change. Decide who will maintain what you build, and include that effort when judging its value.

  25. What to take back

    Find a real practice pool and check the results. Use the grid to protect high-leverage work, technical learning and team ownership, then carry those skills beyond your team alongside domain experts.

  26. The value of each dive is what we bring back

    Bring back skills that let you contribute in new ways, and judgment that creates room for your team’s most valuable work. That is the value of going deeper.

  27. Agent skills you can use

    Try the agent skills behind these stories. Start with Daily Debrief to see what needs attention, or Choose Where Attention Belongs to identify one repeated step worth improving.

Start here

Start with one task.

Choose a skill, add your context and check the result. Next time, supply fresh context instead of rebuilding the instructions.

Fictional example / AI draft

Your daily debrief

Tonight (0)

Nothing in the supplied sources needs action tonight.

Tomorrow (1)

  • Review Morgan’s rollout checklist

Tracking (2)

Keep visible; this does not mean done.

  • Confirm the review date
  • Track Avery’s dependency check
AI proposes priorities. You choose what to act on.

Agent skills and examples

Skills on GitHub
AI generation or assessmentHuman input or judgment
Compare human effort and quality for comparable cases. Calculations stay separate from the AI drafted assessment. A person checks assumptions and decides whether to keep, revise or stop.Human effortPreparationReview + fixesKeepReviseStopBaselineAI workflowYou decideAI drafts assessment
07 / Companion agent skill

Check whether the automation is helping

Compare preparation, review, corrections and upkeep before deciding whether to keep an automation.

Possible toolsReportsCSV

Was the result useful?

Tell us what you used, what you had to correct and what would make the next run better.

Apply the frameworkWhat each quadrant leaves behind
About these resources & agent-readable formats

Adaptable resources based on my talk, with a new measurement companion. They contain no company data. Each guide explains the result, tools, human review and value to check.

Updated November 4, 2026. The tools and my approach will keep changing.