Becoming an AI Power User

5.7Infrastructure and Learning Systems

They match delegation to the task and the checks available

Developing

Delegation needs checks that can detect the errors that matter

Suppose an assistant drafts a client update from the project plan. The plan says Friday, but you agreed to deliver on Thursday in a call the assistant cannot access. A careful-looking draft can still give the wrong deadline. A useful review needs the missing evidence, not merely another reader. Supply the agreement, connect the source, or ask someone with that context to check the deadline.

A second model can compare the draft with a transcript, and a simple check can flag a date that differs from the approved schedule. If the disagreement requires interpreting a promise made to the client, involve an evaluator who can understand that promise. The relevant advantage may be information, domain skill, or a reliable test. It does not follow automatically from being human or AI.

Match assurance to the consequences of error and the strength of the available checks. A low-stakes draft may need a quick review. A consequential recommendation may need source verification, specialist assessment, or a tested evaluation process. The aim is an arrangement that produces the best aligned result and makes failures discoverable.

Delegating the labor still leaves you holding the call

When you delegate a client update, you also decide how the result will be checked. A model can flag unsupported claims; software can compare dates against a schedule; someone who knows the relationship can assess an ambiguous commitment. Ask what each check can see and what it can miss. A fluent review is no more self-validating than a fluent draft.

Delegate evaluation after establishing that it works. Test a proposed checker with an update containing a known wrong deadline and another that is correct. See whether it catches the error without inventing one in the clean draft. Then add cases that reflect other important failures. Your responsibility is to choose and maintain a trustworthy arrangement, which can include checks that run without your reviewing every result.

Practice matters when you still rely on a skill

If a workflow depends on your ability to catch a subtle error, check that ability in practice. For the client update, ask an assistant to prepare a separate set of practice drafts using verifiable sources. Include valid drafts and drafts with errors in dates, evidence, or commitments, and keep the answer key separate until you have assessed them. Then compare your findings with the sources and key. This tests both missed errors and false alarms.

Skills can weaken through disuse, but delegation does not prescribe one inevitable trajectory. You may practice less of the original task while learning to design better checks, compare evidence, or diagnose failures. Decide which capabilities your goals still require. If independent writing or domain judgment is one of those goals, include practice and feedback that exercise it.

A capability you continue to rely on needs a way to stay available when it matters.

Choose a collaboration mode by capability, evidence, and consequences

Full delegation can fit work whose requirements and important errors can be checked reliably. Close collaboration can help when requirements are still being discovered or a person holds relevant information the system lacks. High stakes require stronger assurance, which may come from specialist review, independent data, mechanical checks, a validated AI evaluator, or a combination.

An experienced user may delegate one task fully and work interactively on another. A settled reporting format can run with automated source checks; a new strategy question may benefit from a conversation that explores competing aims. The choice follows the task and evidence. It need not progress in one direction as the user gains experience.

Describe both who performs the task and when its results are checked. The modes below describe approaches that also work in combination: you might work closely with an assistant to clarify a new requirement, delegate the draft, and use tested checks to supervise repeated runs. Their names help describe an arrangement; they do not establish its quality.

Tool use: you invoke the model for a discrete sub-step and do the core work yourself. As of 2026 this looks like asking for one paragraph or one calculation inside a task you are otherwise driving.
Pair work: continuous partnership where you stay in the loop on every decision, trading the work back and forth turn by turn.
Delegation: the model executes the whole task and you review the output after the fact.
Supervision: the model runs on its own within bounds you set, and you watch its behavior in aggregate rather than checking each result.

Reliable execution supports improvements to the whole workflow

Once a recurring task and its checks work, a broader question becomes useful: what would improve the whole workflow? Perhaps the assistant should collect the current sources itself, catch contradictory dates before drafting, or request a decision only when a material uncertainty remains. This is abstraction jumping: encode a practice you have established and then work on the design that coordinates it.

For the client update, start by checking whether the assistant reliably extracts commitments from the approved notes. Once that works, ask it to compare those commitments with the project schedule before writing. Later, a reusable workflow could handle that comparison on every update and bring you only unresolved conflicts. Each level depends on evidence that the lower-level work and its checks are reliable enough for the new use.

Judge an improvement by the outcomes you care about. Compare accuracy, usefulness to the client, preparation and review effort, and any capabilities you want to develop. Enjoyment matters too: experimenting with a new format or a more revealing test can be part of why you continue. These qualities can trade off. A valuable workflow need not be simultaneously faster, more relaxing, and more fun on every task.

Workflow checks can miss errors for different reasons

These examples can help you inspect a workflow for weak points. Each names a possible failure and a response to investigate. Several can occur together, and the same symptom may have more than one cause.

Risk	How it develops	How to prevent it
Skill atrophy	A skill required for independent checking receives little practice, so errors may become harder to detect.	Practice the specific skill with feedback where you still need it, or establish another competent check.
Overtrust	Plausible results are accepted without testing whether the current checks would catch an important error.	Test the checks on known flawed and valid cases, and verify consequential claims against the relevant sources.
Invisible errors	A subtle error survives because the checks lack evidence or cannot detect that type of mistake.	Give the evaluator the relevant sources and test it on known errors. Use a suitable independent check for consequential claims.
Context loss	The workflow produces a recommendation without retaining the goal, sources, or decisions needed to explain or check it.	Keep the goal, supporting sources, and confirmed decisions beside the output. Ask the assistant to link its conclusions to that record. Distinguish your stated reasons from explanations the assistant proposes, and resolve consequential gaps before relying on the result.

Doubt the evidence too, including this book's numbers

Read a finding together with its limits before applying it to your workflow. A survey of reported effort cannot show how an individual's skill changed over time, and results from industrial automation do not measure a current AI assistant's performance. The research can suggest what to investigate. Your own task still needs the relevant sources and checks, judged against what you want to accomplish.

Write down what you want to achieve and how you will know whether it happened. That context helps an assistant assess the work and can reveal where your own requirements need clarification. An evaluator can then compare output against the intended result instead of substituting whatever sounds plausible.

Keep the effort that serves your goals and delegate work when the resulting arrangement performs better. Those goals can include learning, enjoyment, good decisions, and reliable delivery. As you discover useful checks, save and test them so the system can carry more of the evaluation. That may leave more attention for questions the current setup does not yet answer.

Audit your workflows and reset the modes that have drifted

Audit your workflows for delegation risk

Claude interviews you about recurring tasks and the checks they rely on, then recommends where to delegate, strengthen evaluation, or collaborate. You assess the recommendations against the work's goals and consequences.

Copy the promptUse the copy button above the box.
Paste it in Claude or CodexMake sure your active project is enabled.
Answer the questionsGive clear answers so it understands what you want made.
Test it to see if it worksTry the output and refine until it does what you need.

I want to choose suitable delegation and evaluation arrangements for my recurring AI workflows. 1. Ask about the tasks I use AI for and how each result is currently checked. For each task, establish the consequences and reversibility of an error, the relevant sources, and the expertise or authority needed to assess it. 2. Produce an audit table with the current mode, important failure types, evidence available to each evaluator, and the performance evidence supporting the present checks. Do not treat a human reviewer or a second model as reliable by default. 3. Recommend whether to delegate, collaborate, add source access, use a mechanical check, or obtain suitable specialist review. Explain what each recommendation addresses and what remains uncertain. 4. Propose one representative test for a weak point. Include a known flawed case and a valid case so the check is tested for both missed errors and false alarms. Keep practice material separate from live work. 5. Ask whether developing or preserving a particular skill is also one of my goals. If it is, suggest practice with feedback that exercises that skill. Do not assume every delegated skill must be preserved. Scope: assess and recommend from the information I provide. Route by the task, evidence, consequences, reversibility, and demonstrated capability. Do not moralize about delegation or claim that a workflow is safe merely because its past runs looked good.

References

2 sources

1
Ironies of Automation View source
Lisanne Bainbridge · 1983 · Automatica 19(6), pp. 775–779
Bainbridge examined industrial automation that left people responsible for uncommon failures. She described the difficulty of maintaining intervention skills and of asking a person to monitor a system whose decisions exceed what that person can check.
The relevant design question is whether the assigned reviewer can detect and resolve the expected failures. A human checkpoint provides little assurance when the person lacks the necessary evidence, skill, or time.
This is a historical analysis of industrial process control and aviation, not an experiment on current AI knowledge work. Its design questions are useful; it does not establish a permanent human advantage over AI.
2
The Impact of Generative AI on Critical Thinking View source
Lee, Sarkar, Tankelevitch et al. · 2025 · CHI 2025
The survey collected responses from 319 knowledge workers. Confidence in generative AI and task-specific self-confidence were associated with self-reported critical thinking behavior and effort in the reported tasks.
The finding motivates checking how confidence relates to review behavior. It does not demonstrate that AI use causes a progressive loss of ability.
The study is a cross-sectional self-report survey, rather than a longitudinal skill test or randomized trial. It cannot establish causation, a timetable for skill loss, or the effect of a current AI system.