A11 / Practical guides
Keep the vision, delegate the execution
Decide what to delegate, which decisions to keep, and who can supervise AI-assisted work. Compare total team effort and use a practical handoff checklist.
- 01Prepare the brief
- 02Pair human + agent
- 03Verify the outcome
The person with a clear idea may have little time to implement it. Doing everything themselves creates a bottleneck. Handing over a vague task can create a different one: repeated explanations, reviews and corrections.
AI tools can reduce parts of the execution effort. They do not automatically transfer intent or guarantee the result. A useful handoff makes both the desired outcome and the decision boundaries explicit.
Keep intent and the acceptance decision
The requester owns what should change, why it matters and the criteria used to judge it. Those criteria can include taste and judgment, but the executor needs to know where those judgments apply.
The executor owns running the work: planning steps within scope, supervising the agent, checking intermediate results, asking necessary questions and assembling evidence. They need room to act without inheriting decisions the requester meant to keep.
| Decision | Default owner | Boundary |
|---|---|---|
| Intended outcome and priorities | Requester | Confirm before execution; revisit when intent changes |
| Implementation choices within scope | Executor | Respect the agreed constraints and project rules |
| New scope, changed acceptance or trade-offs outside the brief | Requester | Executor brings options and consequences |
| Domain-specific risk or correctness | Qualified specialist | Add review where requester or executor cannot judge it |
| Delivery readiness and evidence | Executor | Report actual checks and known limits |
| Acceptance | Authorized reviewer, often requester | Judge against the agreement; distinguish self-verification |
Write a vision that can travel
Consider a hypothetical mobile-game prototype: a mining loop with upgrades. A half-page brief could state:
- The question: does each upgrade make the mining interaction feel different enough to sustain interest?
- The minimum scope: one playable loop and three upgrades that change the interaction, not only its numbers.
- What does not matter yet: placeholder art is acceptable; menus, saving and monetization are out of scope.
- A reference: one clip or sketch that shows the intended rhythm, with a note about what to borrow from it.
- The review: a ten-minute play session on the agreed phone, followed by a decision about whether the loop is worth developing further.
The prototype is an experiment. The reviewer may decide the mechanic is not promising even when the executor delivered a good prototype. Separate those questions: did the prototype meet the brief, and what did the experiment teach us?
If the feeling is still hard to describe, agree an early checkpoint before the executor builds all three upgrades. That gives judgment a place in the process without turning every implementation choice into a new approval.
Count the whole cost
A handoff is useful when the benefits exceed the effort it creates. Count preparation, questions, execution, supervision, review and rework. Look at total team effort as well as the requester's attention.
| A handoff is more promising when | It may cost more when |
|---|---|
| The outcome and boundaries are clear | The requester is still deciding what they want |
| The executor can recognize a wrong result | Nobody available can judge the domain |
| Checking is substantially easier than doing | Review requires redoing the work |
| There is a meaningful unit of work to transfer | It is a trivial change with costly setup |
| Requirements are stable enough for the task | Direction changes before each review |
| The executor has real capacity | The work would displace higher-priority commitments |
Task size alone is not enough. A difficult task with clear checks can be a good handoff; a small but ambiguous task can consume repeated rounds of clarification.
Choose someone who can supervise
Look for domain judgment, available time, legitimate access and suitable AI tools. Ask whether the person can notice a plausible but wrong result and knows where to obtain help.
A game designer may be well placed to judge a prototype's interaction. An engineer may be needed for performance or a difficult integration. A QA teammate may be well placed to investigate a reproducible bug. The useful match depends on the checks, not a general claim that AI makes roles interchangeable.
Use authorized accounts and tools. Delegation does not require sharing logins or moving a provider subscription between people.
Avoid reserving work for someone who cannot start. If a particular model or specialist is required, explain the reason. Otherwise, give a qualified executor room to use the tools that meet the task's needs.
Send context, not your entire history
Include the outcome, relevant decisions, references, source links, constraints and delivery destination. Summarize the conclusions that affect the task and keep links to the underlying material for verification.
An agent can help draft from recorded conversations when it has authorized access. That does not remove the need to choose the relevant sources, resolve contradictions or confirm the draft. Keep credentials out of the brief and use the team's normal access mechanisms.
Make review fair and useful
Review against the criteria agreed before execution. Accept when they are met. Send back specific issues the executor can address, or reopen the task if it needs a different match.
If the brief omitted a requirement, say so. Agree the new scope rather than blaming the executor for following the instructions. If the result reveals an important correctness or safety problem, resolve it before acceptance even if it was not anticipated.
Bring in a specialist when you cannot judge an important part of the result. Keeping the final call does not mean pretending to have every kind of expertise.
Know when to change the handoff
When work repeatedly returns for rework, ask whether the brief is missing context, the executor needs different support or the task is actually exploration.
Exploration is legitimate. Give it an outcome such as options, a prototype and evidence, rather than pretending an unsettled direction is ready for implementation.
Keep work that changes direction constantly, cannot be safely delegated with appropriate access or takes as much effort to review as to do.
The handoff checklist
KEEP
[ ] Intended outcome and reason
[ ] Non-negotiables and excluded scope
[ ] Named decisions that remain with the requester
[ ] Acceptance criteria and reviewer
TRANSFER
[ ] Executor can judge this kind of work
[ ] Time, access and authorized AI tools are available
[ ] Relevant context and references are accessible
[ ] Implementation freedom and stop points are explicit
[ ] Early checkpoint for uncertain or subjective work
[ ] Delivery form and evidence are agreed
REVIEW
[ ] Checks and evidence come from this work
[ ] Known limits and omitted scope are visible
[ ] Specialist review is included where needed
[ ] New requirements are identified as changes
LEARN
[ ] Count preparation, questions, supervision and rework
[ ] Compare requester attention and total team effort
[ ] Check quality and usefulness, not only speed