Skip to content

A9 / Practical guides

The Wagglet Way: a practical guide to turning ideas into accepted work

Turn ideas into accepted work with two aligned briefs, qualified supervision and evidence-based review. Includes a worked example and reusable handoff template.

  1. 01Prepare the brief
  2. 02Pair human + agent
  3. 03Verify the outcome

The Wagglet Way is a method for AI task handoff: capture intent, prepare two aligned briefs, confirm the scope, supervise execution and review the evidence before acceptance. The requester keeps the outcome; the executor owns supervision and delivery.

An idea can be clear in one person's head and still be impossible for someone else to execute reliably. AI tools can help with execution and drafting, but they do not remove that gap. Someone must explain the outcome, provide the relevant context and judge the result.

At Rockbite Games, we call our approach the Wagglet Way. A requester captures an idea. A person or agent prepares the task. A qualified teammate supervises execution with their own AI tools. The work returns with evidence for review.

You can use this method in a shared document and a task board. The templates below are meant to be adapted, not treated as a form every small change must fill out.

In this guide

The responsibility that moves

Delegation moves more than the steps of a task. The executor takes responsibility for starting in the right place, noticing errors, handling questions within scope and making the result reviewable.

The requester keeps the intent: what should change and why. They confirm the draft, resolve decisions that remain theirs and usually review the outcome. A specialist may need to judge parts of the result that the requester cannot evaluate.

RoleOwns
RequesterIntent, context, confirmation and changes of scope
DrafterA faithful brief, explicit unknowns and usable checks
ExecutorSupervision, delivery quality and evidence
ReviewerThe acceptance decision against agreed criteria

For independent review, choose a reviewer other than the executor. If self-verification is necessary and supported by the tool, label it honestly and do not count it as independent review.

1. Capture the idea while the context is fresh

Separate having the idea from writing the task. Start with the project, one line about what should change and the source conversation or evidence. Add why it matters if the reason is not obvious.

For example, this is an illustrative request: “New hires can't find the staging login. Put the approved access instructions where they start onboarding.” It identifies a problem and a direction without pretending the solution has already been specified.

Speaking can be easier than typing. Record or dictate a short explanation using your team's agreed tools and access rules. A drafter can use recorded context only if the relevant material is available to it.

2. Prepare the task from sources

Give the drafter the relevant notes, decisions, files, references and code. Link to the originals and summarize the decisions that matter. Identify which source is authoritative when material conflicts.

An agent needs configured access to those sources. A link in a prompt does not guarantee it can open the page, read an attachment or inspect a repository. The draft should say what was unavailable and what remains unknown.

For the onboarding example, useful context might include the onboarding page, the approved access procedure and the person responsible for granting access. An unresolved question could be: “Do contractors follow the same process?” That question should stay visible until an authorized person answers it.

Split work when its parts have different owners, evidence or dependencies. Give each part an observable result and name who integrates them. Avoid turning an uncertain project into a long task whose acceptance criteria are still guesses.

3. Write two aligned briefs

The agent brief explains the intended outcome, the current situation, sources, scope, constraints, evidence and stop conditions. Write it as if the agent has no surrounding conversation. Clear instructions help, but they do not guarantee that a model will retain or follow every detail.

The human brief describes a different job: what the executor must provide, decide or watch. Keep it short, with separate QA steps when the checks need detail.

For the illustrative onboarding task:

Agent brief

Add a clear link from the onboarding start page to the approved staging-access instructions. Use the current access procedure as the authority. Preserve the access controls; do not publish credentials or create a new login flow. Ask the executor to obtain a decision from the access owner if the contractor process is unclear. Deliver the edited page and a list of checks and unresolved questions. Do not deploy without the task's authorization.

Human brief

Confirm the access owner and the approved procedure. Obtain a decision about contractors if needed, then check that a new starter can find the instructions without seeing credentials or gaining unintended access.

QA

  • From the onboarding start page, follow the link using an appropriate test account.
  • Confirm the destination is the approved procedure and that access restrictions still apply.
  • Check the normal employee path and the agreed contractor path separately.

The example is hypothetical. Its purpose is to show how intent becomes a bounded handoff, not to report a customer outcome.

4. Confirm before execution

Read the draft for three things: does it describe the intended outcome, are its acceptance checks observable, and has it turned any unknown into an assumption?

Check that the two briefs agree. If one tells the agent to publish and the other tells the person to review before publication, fix the contradiction before assigning the work. Put any decision the agent must stop for in the agent instructions themselves.

Do not rely on a generic “use your judgment” where a named person owns a decision. Name that person or role and explain how the executor should reach the decision through the normal workflow.

5. Match capability and capacity

The executor needs more than spare time or an AI subscription.

FactorQuestion to answer
JudgmentCan this person recognize a wrong or incomplete result in this domain?
TimeCan they start and supervise the work alongside existing commitments?
AccessDo they have the necessary repository, environment, data and decision access?
AI toolsDo they have authorized tools and sufficient capacity for the task?
EscalationIs there someone available to resolve questions outside their authority?

A designer might be well placed to supervise a prototype's interaction quality. That does not automatically qualify them to approve a sensitive data migration. Match the checks and risks, not just the title.

Respect existing workload. A shared board can let eligible people claim tasks when they have capacity, but a visible task is not always ready to start: dependencies or unanswered decisions may still block it.

6. Supervise the work

The executor starts in the right environment, provides authorized access, checks intermediate results and answers questions within scope. When the intended outcome or scope must change, they return to the requester.

Agent tools may support screenshots, browser inspection and automated checks. Decide which evidence the agent can gather and which checks need a person, a real device or specialist judgment. Assign those responsibilities explicitly instead of assuming a universal capability limit.

Stop when required information is missing or the work crosses an agreed boundary. Record the question and who needs to answer it. A named decision owner and a clear follow-up path are more useful than a task that only says “waiting for the team.”

7. Deliver evidence appropriate to the result

A delivery should let the reviewer understand what happened without redoing all the work:

  • The result or a link to it.
  • What changed, or why no change was needed.
  • The checks actually performed and their results.
  • Evidence from the current work, such as tests, source comparisons, screenshots or recordings.
  • Known limits, skipped scope and unresolved questions.

A no-change finding can be useful work. It still needs evidence: the rule examined, the observed behavior, the decision and its limits. If an original proof requirement no longer fits, resolve it with the task owner. Do not substitute an unrelated artifact just to satisfy an upload requirement.

8. Review against the agreement

Accept when the result meets the agreed checks. Send it back when the same executor can address a specific issue. Reopen it when the work should become available for another claim. Explain the reason without reducing every mismatch to a failure by the executor.

If a missing requirement appears during review, acknowledge the scope change. Agree the additional work or create a follow-up instead of treating an unstated preference as a previously failed criterion. Newly discovered safety or correctness problems still need resolution before acceptance.

For code, acceptance does not prove that a branch merged or a change deployed. Track those outcomes separately.

9. Learn whether the handoff helped

Track useful accepted results, quality and the attention spent on preparation, questions, supervision, review and rework. Separate elapsed time from active effort. A task can take two weeks because it waited for a decision; timestamps alone do not show how many hours people worked.

Ask why work was sent back. Was the brief ambiguous, the context missing, the implementation wrong or the outcome itself changed? Turn recurring problems into short, concrete drafting rules.

Do not optimize only for the requester's time. A handoff that saves ten minutes for a lead but creates hours of avoidable work for teammates may be a poor trade.

When to keep or narrow the work

Keep a task when checking it costs as much as doing it, when nobody available can judge it, or when the executor cannot be given appropriate access. If the outcome is still unknown, consider a research task that ends in options and evidence rather than an implementation task that pretends the decision is settled.

A reusable handoff template

Title: [one-line outcome or problem]
Requester: [name]       Reviewer: [name]
Source: [request / discussion / evidence]

AGENT BRIEF
Intent: [outcome and reason]
Today: [current behavior]
Change: [exact scope and locations]
Context: [sources, relevant decisions, access needed]
Do not change: [constraints]
Stop and ask: [unknowns, decision owner, scope boundaries]
Deliver: [artifact / branch / report; publication authority]
Done when: [3-5 observable checks]

HUMAN BRIEF
[Purpose.] [What you provide or decide.]
[What you supervise and when you escalate.]

QA
[Checks, responsible person/tool, expected result]

DELIVERY
Result: [link or finding]
Checked: [actual checks and evidence]
Limits: [not tested, uncertain, not done]

REVIEW
[Accept / send back with reason / reopen]
[Independent review or explicitly labelled self-verification]

Running it in Wagglet

Requests capture ideas before full drafting. Drafts can be written by a teammate or delegated to an available configured agent. Human-run AI tasks have separate human and agent instructions; human-only work does not need an agent brief. Claim rules, proof requirements and review actions help make the handoff explicit.

Use the correct access boundary for the workflow. Standard Task Handoff is rooted in the active claim and authorized related context. Personal MCP connections act with their owner's permissions. Neither replaces careful source selection or requester confirmation.

Continue the handoff