The Question Changes the Work — The question changes the work

Session 04 · Extended talk track · Slide 1

The question changes the work

Expanded talk track

This is the advanced end of the series, but the core move is ordinary: name the job, name the boundary, and decide what would prove the result worked. The examples get bigger in this session. Some involve several AI helpers, a working artifact, or a future scenario. The human discipline stays the same. A vague request gives the system room to solve the wrong problem. A precise question can reveal the real constraint, remove an assumption, or create a test that the result must pass. You do not need to know every technical implementation before you begin. You do need to describe what must change, what the system may not do, and what evidence you expect at the end. As you move through these examples, listen for the question that changed the path. Then think about one question that could change a real task in your own work.

Participant recap, exercise, and public sources

Recap

Advanced AI work starts with a clear job, a boundary, and proof that the result worked.

TRY THIS

Pick one real task. Write the result you want, one rule the AI must respect, and one check you can perform.

1. The question changes the work

The Question Changes the Work — The question changes the work

Expanded talk track

This is the advanced end of the series, but the core move is ordinary: name the job, name the boundary, and decide what would prove the result worked. The examples get bigger in this session. Some involve several AI helpers, a working artifact, or a future scenario. The human discipline stays the same. A vague request gives the system room to solve the wrong problem. A precise question can reveal the real constraint, remove an assumption, or create a test that the result must pass. You do not need to know every technical implementation before you begin. You do need to describe what must change, what the system may not do, and what evidence you expect at the end. As you move through these examples, listen for the question that changed the path. Then think about one question that could change a real task in your own work.

Participant recap, exercise, and public sources

Recap

Advanced AI work starts with a clear job, a boundary, and proof that the result worked.

TRY THIS

Pick one real task. Write the result you want, one rule the AI must respect, and one check you can perform.

2. The first three sessions built a ladder

The Question Changes the Work — The first three sessions built a ladder

Expanded talk track

The four sessions form a ladder. Session 1 focused on the interaction itself: ask clearly, protect private information, check the answer, and keep important decisions with a person. Session 2 added business context, including audience, voice, approved facts, and examples. Session 3 turned that context into a repeatable process that another person could test. This session applies the same discipline to larger work. The system may divide a job, use tools, create artifacts, or coordinate several helpers. Bigger capability does not erase the earlier rules. It makes them more important. If a result disappoints you, ask which rung is missing before you change products. Was the request unclear? Was the context weak? Was the process never tested? Was the approval boundary missing? The point is not to chase the most advanced tool. The point is to carry the operating discipline upward as the work becomes more capable.

Participant recap, exercise, and public sources

Recap

The four-session progression is cumulative: ask and verify, add context, build a repeatable process, then coordinate larger work.

TRY THIS

When a result disappoints you, identify which layer is missing before changing tools.

PUBLIC SOURCES

https://marksalisburyai.com/

3. The thing I actually know

The Question Changes the Work — The thing I actually know

Expanded talk track

I often do not know the exact technical implementation that will work when a project begins. What I can usually identify is the job, what looks wrong, and what would have to pass a test. That is an operator skill. You can say that a layout feels crowded without knowing the CSS property that will fix it. You can require a report to reconcile to a known total without knowing how the spreadsheet should be built. You can define a playable game without choosing the engine first. The transferable behavior is noticing a missing constraint, describing the desired change in observable terms, and checking the result. That does not mean every beginner can reproduce every finished system immediately. It means domain knowledge and judgment can guide technical work without pretending you already know the implementation. Capability also does not grant authority. The agent may propose or build, but consequential decisions still belong with the person or role responsible for them.

Participant recap, exercise, and public sources

Recap

You do not need to know every technical step. You do need to recognize a bad result and define what must pass a test.

TRY THIS

Describe success in observable terms such as correct names, no unsupported claims, or a file that another person can use.

PUBLIC SOURCES

https://marksalisburyai.com/

4. Grocery shopping is a real task

The Question Changes the Work — Grocery shopping is a real task

Expanded talk track

Grocery shopping sounds simple until the job includes prior orders, exact products, changing availability, substitution rules, quantities, and a checkout boundary. In the workshop example, the assistant used prior purchase patterns to help rebuild a cart, kept an explicitly rejected substitution out, omitted an unavailable exact item instead of guessing, and stopped before checkout. The private account and order history are not part of the public material. What matters here is the workflow shape. The AI can research and prepare. The live store state provides evidence about availability. The shopper defines which substitutions are acceptable and which are prohibited. The person keeps authority over the purchase. For any buying task, state the spending limit, exact-item rules, allowed alternatives, prohibited alternatives, and the approval point. A plausible substitute can still be the wrong result. In this case, the exception rule mattered more than filling every line in the cart.

Participant recap, exercise, and public sources

Recap

Even an ordinary grocery task contains exact-item rules, availability problems, and a final authority boundary.

TRY THIS

For any purchase task, state allowed substitutions, prohibited substitutions, the spending limit, and the approval point.

KEEP IN MIND
The workshop example came from a private account workflow. No private account link or order record is included.

5. The grocery question that mattered

The Question Changes the Work — The grocery question that mattered

Expanded talk track

The important move was turning a previous failure into a rule the next run could check. “Do not make that substitution again” is more useful than general frustration because it changes the observable outcome. The assistant could audit the cart for the prohibited item and report whether it appeared. It could also leave out an exact product when the requested match was unavailable rather than quietly choosing something similar. The account history and underlying conversation remain private, so the public lesson is the rule itself: describe what must never happen again, why the distinction matters, and what evidence will prove compliance. Then keep the final authority boundary explicit. Preparing a cart is not the same as placing an order. A short exclusion can become an operational control when it is specific enough to test. Think of one repeated failure in your own work. What plain rule would prevent it, and what receipt, count, comparison, or checklist would prove the rule held?

Participant recap, exercise, and public sources

Recap

A short exclusion can become an operational rule when it is specific enough to audit.

TRY THIS

Turn a past failure into a plain rule: what must never happen again, and what evidence proves compliance?

KEEP IN MIND
The underlying conversation and shopping record remain private. The deck includes only the participant-safe lesson.

6. A request is not the entire system

The Question Changes the Work — A request is not the entire system

Expanded talk track

A prompt is only the request. The real workflow includes the live system the AI must inspect, the state that can change while it works, and the recovery steps when something interrupts the task. Product names may be similar without being the same item. Availability can change between search and cart review. A browser session can lose its place. The assistant may need to resume, recheck the current state, and audit the result before stopping. That is why completion criteria matter. “I tried to add the items” is not the same as a final cart that matches the rules. Add a recovery instruction to any browser-based task: if the work is interrupted, inspect the live system again, reconcile the current state against the request, and summarize what remains unresolved. Keep private screens off public displays and keep consequential actions behind approval. A reliable result needs evidence from the actual system, not merely a confident description of what should have happened.

Participant recap, exercise, and public sources

Recap

A prompt is only the request. The real system includes changing data, interruptions, recovery, and a final audit.

TRY THIS

Add a recovery instruction: if the task is interrupted, recheck the live system and summarize the final state before stopping.

7. TeamLoop, without the computer words

The Question Changes the Work — TeamLoop, without the computer words

Expanded talk track

TeamLoop is a simple idea from the user’s point of view. You give one job to one lead. The lead decides whether another kind of help would materially improve the result. A researcher might gather evidence. A builder might create the artifact. A reviewer might look for omissions or try to break the work. The lead remains responsible for combining the pieces, resolving conflicts, and returning one result with a clear account of what was checked. If extra help would add nothing, the lead should work alone. More agents do not automatically mean better work. They can also create duplicate effort, conflicting assumptions, and more review. The value comes from assigning genuinely different jobs and keeping one accountable lead. The internal routing machinery is not needed for this lesson. The human still defines success, protects the boundary, and judges the finished result. You ask once. The team handles useful handoffs. You keep the decision.

Participant recap, exercise, and public sources

Recap

A lead AI can divide a job among specialist helpers and return one checked result. The human still judges the outcome.

TRY THIS

Use a team only when research, building, or independent review would materially improve the result.

KEEP IN MIND
Private TeamLoop operating files and internal routing details are intentionally not linked.

8. Give the team a real mission

The Question Changes the Work — Give the team a real mission

Expanded talk track

Before any team starts working, define three lines. Result: what concrete thing should exist at the end? Rule: what boundary must the work respect? Proof: what observable check will tell us whether the result passed? Use a small public or fictional job for the first test. Keep private accounts, client details, and sensitive business material out. The result should be small enough to finish and inspect in the available time. Once the brief is clear, the lead can decide whether research, building, or independent review is useful. Watch the human decisions rather than the technical plumbing: why was a helper added, what job did that helper receive, and what judgment remains with the lead? The impressive part is not a long prompt or a busy screen. The useful part is that success was defined before the team began. Without that, even an impressive artifact may solve the wrong problem.

Participant recap, exercise, and public sources

Recap

The useful brief contains a result, a boundary, and a pass or fail check before any AI begins working.

TRY THIS

Write those three lines for a small public or fictional task. Keep client data and private accounts out of the first test.

9. Open The Dream Machine

The Question Changes the Work — Open The Dream Machine

Expanded talk track

A visual status layer can make complex work easier to follow, but it should not pretend to show a model’s private thoughts. Its job is to expose observable work: which helper was assigned, what artifact or evidence appeared, what changed, and what still needs a human decision. The image on this slide is a concept view, not proof that a specific run completed. To evaluate a real run, open the actual task or receipt and compare it with the original Result, Rule, and Proof. Ask the room to find one miss. If they find one, repair the work or name the next check. That is more valuable than defending the output. A polished interface can create the feeling that something important is happening even when no useful artifact exists. Ask the blunt question: what completed result or receipt proves the job finished? Cool UI earns its place only when it helps people inspect the work.

Participant recap, exercise, and public sources

Recap

A visual status layer helps only when it exposes the real work, the evidence, and the remaining judgment call.

TRY THIS

When an AI tool looks impressive, ask what completed artifact or receipt proves that the job actually finished.

10. Play a small public build

The Question Changes the Work — Play a small public build

Expanded talk track

A screenshot can show that a game screen exists. It cannot prove the game is playable. The useful test is direct interaction: can a person start the game, control the character, see the rules respond, and recover from an ordinary mistake? That same distinction applies to many AI-built artifacts. A polished mockup is not a working form. A workbook that opens is not proof that its formulas are correct. A generated website is not complete until its links, layout, and important actions work on the intended devices. Choose one behavior that a person can test directly after the build. Keep the test small and observable. If the artifact fails, record the behavior that failed rather than describing the entire project as bad. A public build becomes convincing when someone can use it and reproduce the expected behavior. Playability is an observable criterion. A screenshot is supporting evidence, not the criterion itself.

Participant recap, exercise, and public sources

Recap

A working artifact has an observable test. For a game, playability matters more than a convincing screenshot.

TRY THIS

Choose one behavior a person can test directly after the build.

PUBLIC SOURCES

https://marksalisburyai.com/work/

11. A metaphor fixed a layout

The Question Changes the Work — A metaphor fixed a layout

Expanded talk track

Several rounds of positional feedback had not solved a layout problem. The breakthrough did not come from naming a CSS property. It came from a familiar visual model: arrange the four items like the corner dots on a die. That metaphor communicated spacing and relationship more clearly than “move this a little left” and “raise that one.” The implementation could then use a simple two-by-two structure and test whether the overlap disappeared. The next review still mattered. Fixing the collision exposed too much empty space and a footer that no longer felt connected to the composition. The first successful change was not the end of quality assurance. When visual feedback stalls, compare the desired result to a familiar physical arrangement, page, tool, or object. A plain metaphor can become a technical specification when it describes relationships the builder can reproduce and the reviewer can verify.

Participant recap, exercise, and public sources

Recap

A familiar metaphor can specify a technical layout more clearly than several rounds of positional instructions.

TRY THIS

If visual feedback stalls, compare the desired layout to a familiar physical object or arrangement.

KEEP IN MIND
The detailed project conversation and local training records remain private.

12. A price question changed the lesson

The Question Changes the Work — A price question changed the lesson

Expanded talk track

A small question about price or account access can expose a hidden dependency in an otherwise good workshop plan. If an exercise requires a feature that participants do not have, the lesson fails before the learning begins. The right response is not to force the original plan. Check the current product documentation, identify which account tiers and regions have access, and design a fallback that teaches the same principle without the unavailable feature. In this case, the lesson shifted toward prepared context and a process card rather than requiring every participant to create the same reusable package live. Product availability changes, so treat this as a planning pattern rather than a permanent statement about any vendor. Before teaching a workflow, verify the account tier, regional availability, required connectors, file permissions, and recovery path. A six-word question can prevent twenty people from discovering the same access problem in the room.

Participant recap, exercise, and public sources

Recap

A small pricing or access question can reveal that a planned workshop exercise will fail for participants.

TRY THIS

Before teaching a product workflow, confirm account tier, regional access, required connectors, and the fallback path.

KEEP IN MIND
Product availability changes. Recheck the current product documentation before using this as a live exercise.

13. Playable meant changing platforms

The Question Changes the Work — Playable meant changing platforms

Expanded talk track

The project had quietly accumulated an implementation assumption: the game had to stay in the browser. The real goal was simpler. It needed to be playable in the intended environment. Once that definition of done became explicit, the team could stop optimizing the wrong platform and choose a build path that better fit the outcome. The important move was separating the need from the initially assumed format. This does not mean platform choices are trivial. A change still requires tests for launching, input, core gameplay, packaging, and the actual device. Automated checks can cover specific behaviors, but they do not certify the entire imagined game or every physical controller. Keep public prototypes distinct from later builds and describe only what was actually tested. In your own work, ask whether the named tool, file type, or platform is a true requirement or merely the first implementation idea. A clearer definition of done can remove an entire wrong path.

Participant recap, exercise, and public sources

Recap

A clearer definition of done can remove an accidental implementation constraint and change the entire build path.

TRY THIS

Separate the outcome you need from the platform or format you initially assumed.

KEEP IN MIND
The detailed game build record remains private. The public work page shows participant-safe examples.

PUBLIC SOURCES

https://marksalisburyai.com/work/

14. Bots are already a majority in one measure

The Question Changes the Work — Bots are already a majority in one measure

Expanded talk track

This statistic needs its denominator attached. Imperva reported that bots accounted for more than 53 percent of the web traffic it analyzed in 2025. That is a vendor measure of traffic in its observed environment. It does not mean that most internet activity, bytes, purchases, or conversations came from autonomous AI agents. Cloudflare uses different categories and a different network view when it distinguishes human requests, AI bots, and other bots. User-directed agents are narrower still. The useful conclusion is that automated web traffic is already widespread and businesses should expect more machine-mediated interactions. The unsupported leap would be “most of the internet is AI agents.” When a number sounds dramatic, ask three questions: what exactly was counted, what population or network was measured, and does the category match the claim being made? Keep the source, date, and denominator beside the number.

Participant recap, exercise, and public sources

Recap

Imperva reported that bots accounted for more than 53% of the web traffic it analyzed in 2025. That measure does not mean most internet activity is autonomous AI agents.

TRY THIS

When you see a dramatic statistic, check the denominator, the measurement method, and whether the source is describing bots or user-directed agents.

PUBLIC SOURCES

https://www.imperva.com/blog/bad-bot-report-2026-bots-agentic-age/
https://radar.cloudflare.com/year-in-review/2025

15. Coding agents do more of the execution

The Question Changes the Work — Coding agents do more of the execution

Expanded talk track

Coding agents now perform more of the execution inside some products, but the evidence does not support one universal speed claim. Anthropic analyzed a large set of Claude Code sessions and reported growing agent execution among users of that product during the study period. METR ran a randomized study with experienced open-source developers using early-2025 tools and found slower completion on the selected tasks. A later METR update said a newer trial suffered enough selection bias that it could not produce a reliable current estimate. These findings can coexist because they answer different questions with different users, tools, and task samples. The practical lesson is local measurement. Pick a recurring task and record total time, review effort, defects, and rework with and without the agent. A fast draft followed by a long repair may not be a productivity gain. Agentic coding is real. Its value depends on the workflow, the task, and the quality of the review.

Participant recap, exercise, and public sources

Recap

Coding agents now execute more work in some products, but controlled studies do not support one universal speedup claim.

TRY THIS

Measure your own task time, defect rate, and review effort before deciding whether an agent improves the workflow.

PUBLIC SOURCES

https://www.anthropic.com/research/claude-code-expertise
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
https://metr.org/blog/2026-02-24-uplift-update/

16. AI helps build AI today

The Question Changes the Work — AI helps build AI today

Expanded talk track

AI already helps researchers and engineers improve AI systems. AlphaEvolve, for example, proposes candidate code and uses automated evaluators to test those candidates against goals people defined. Anthropic has also described substantial AI-authored internal code under human direction. These are meaningful examples of AI-assisted AI development. They are not proof that an AI independently chooses its own goals, builds an unrestricted successor, and recursively improves without human control. The distinction matters because “AI helps build AI” and “AI autonomously replaces itself” are very different claims. In today’s systems, people still choose the objective, build the evaluation loop, decide which result counts, and control deployment. The scoring rule deserves special attention. A system can optimize the wrong target very efficiently if the evaluator rewards the wrong behavior. Keep the goal, test, and approval boundary visible. The goals and scoring rubric still matter.

Participant recap, exercise, and public sources

Recap

AI already helps researchers and engineers improve AI systems, with people setting the goals and evaluation loop.

TRY THIS

Keep the scoring rule explicit. A system can optimize the wrong thing very efficiently.

PUBLIC SOURCES

https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/
https://www.anthropic.com/institute/recursive-self-improvement

17. One possible personal AI layer

The Question Changes the Work — One possible personal AI layer

Expanded talk track

This slide presents a possible future, not a current universal product. Imagine a personal AI layer that screens incoming requests against your preferences, verifies relevance, handles a limited set of routine matters, and brings substantive choices to you. It could also work outward: learn a need, find possible businesses or products, gather evidence, and return a short list or a ready-to-approve action. Pieces of that future already exist. Conversational product discovery is available in some systems. Agent-to-agent standards describe ways software agents can announce capabilities and exchange messages. Limited commercial experiments test what happens when an agent enters the path after a person initiates contact. Those pieces do not prove that personal and business agents routinely negotiate on everyone’s behalf today. Do not assign an adoption rate or timeline. Any acting agent still needs explicit limits for identity, spending, private information, communication, and approval. The person keeps authority.

Participant recap, exercise, and public sources

Recap

A future personal AI layer could screen requests and coordinate routine work, but identity, price, privacy, and authority still require explicit controls.

TRY THIS

For any agent that may act for you, define what it may do alone and what always requires approval.

PUBLIC SOURCES

https://openai.com/index/powering-product-discovery-in-chatgpt/
https://developers.googleblog.com/a2a-a-new-era-of-agent-interoperability/
https://openai.com/index/reimagining-advertising-with-ai/

18. Discovery without a search box

The Question Changes the Work — Discovery without a search box

Expanded talk track

Imagine describing a narrow interest and receiving a small, sourced list of clubs, hobbies, places, events, or people who may share it. That can feel more useful than a blank search box if the agent explains why each suggestion fits and shows the information it used. The privacy boundary is just as important as the recommendation. The agent should ask before sharing personal details, contacting anyone, or representing your intentions. It should not manufacture intimacy, hide why a person was suggested, or speak for you inside a relationship. This remains a scenario rather than a universal current capability. The design question is still practical: what may the agent discover on its own, what data may it use, and what action always requires permission? Discovery improves when the person controls what is shared and can inspect the reasoning behind each suggestion. Convenience should not quietly become unwanted disclosure or unauthorized contact.

Participant recap, exercise, and public sources

Recap

AI-assisted discovery becomes useful when suggestions are sourced and the person controls what information gets shared.

TRY THIS

Require the AI to explain why each suggestion fits and ask before contacting anyone or sharing personal information.

KEEP IN MIND
This slide describes a scenario, not a current universal product capability.

19. Did it pass?

The Question Changes the Work — Did it pass?

Expanded talk track

Return to the criteria defined before the work began. Read the Result, Rule, and Proof exactly as written. Then inspect the finished artifact or receipt. Did the team produce the requested result? Did it respect the boundary? Does the evidence actually demonstrate completion? Score each line before explaining or polishing anything. If the work missed, do not rescue it with a persuasive story. Name the smallest repair and the test that should run again. If a live task is incomplete, say so. An unfinished run is useful evidence about task size, timing, or dependencies. A prepared example can still illustrate the audit, but it should be labeled as an example rather than presented as the live result. This is the accumulated lesson from all four sessions: ask clearly, supply relevant context, inspect the process, and keep the decision. You do not need to understand every implementation detail. You need to know what good looks like and whether the evidence supports it.

Participant recap, exercise, and public sources

Recap

Return to the original result, rule, and proof. A failed test is useful information because it names the next repair.

TRY THIS

Score the output before improving the explanation. If it fails, identify the smallest concrete repair and test again.

20. Questions and close

The Question Changes the Work — Questions and close

Expanded talk track

The technical answer can come later. Human judgment starts the work by defining the job and closes the loop by deciding whether the result passed. Across the series, the same operating habits kept reappearing: state the outcome, provide approved context, name the authority boundary, create an observable test, and repair the source that future work will use. The public links on this slide support the research examples and their limits. Keep the dates and denominators attached to measured claims, and keep future scenarios labeled as scenarios. Your next step should be smaller than your ambition. Use the question card on one low-risk task this week. Write the result you want, one rule the AI must respect, and one proof you can inspect. Then ask yourself: what answer would change the next action in the real job? That question is often more valuable than a longer prompt.

Participant recap, exercise, and public sources

Recap

Human judgment starts the work by defining success and closes the loop by deciding whether the result passed.

TRY THIS

Use your question card on one small, low-risk task this week.

PUBLIC SOURCES

https://marksalisburyai.com/work/
https://www.imperva.com/blog/bad-bot-report-2026-bots-agentic-age/
https://radar.cloudflare.com/year-in-review/2025
https://www.anthropic.com/research/claude-code-expertise
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
https://metr.org/blog/2026-02-24-uplift-update/
https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/
https://www.anthropic.com/institute/recursive-self-improvement
https://openai.com/index/powering-product-discovery-in-chatgpt/
https://developers.googleblog.com/a2a-a-new-era-of-agent-interoperability/
https://openai.com/index/reimagining-advertising-with-ai/