Becoming an AI Power User

3.2Decomposition and Parallel Work

They give conversations clear responsibilities and check how results fit together

Developing

Each conversation needs a defined responsibility and check

Give a separate conversation a clear job when the separation helps you check the work. A research assignment can return source-backed findings without drafting a persuasive proposal. A reviewer can test a completed draft against the brief. One assistant may handle both roles well; separate contexts are a design choice to test when instructions conflict, history becomes confusing, or a fresh review adds useful coverage.

A hub can coordinate focused assignments. It holds the goal, shared facts, and current decisions, then combines the results. Some assignments can run at the same time, such as reviewing separate source documents. Others must wait: a critic cannot review a draft that has not been written, and analysis cannot use evidence that has not arrived.

Separate conversations require clear handoffs and a place to reconcile their answers. A simple task may be easier in one session. For a complex project, assign coordination, specialist work, and review where those responsibilities need separate contexts. Run the critic alongside the specialist only if it can assess the brief independently; otherwise give it the specialist's completed result.

An architectural cutaway places shared goals and decisions in a central hub, with focused workrooms and a draft-to-review handoff. — A coordinator chooses which assignments stay together, run in sequence, or run in parallel, then checks how their results fit.

Split when responsibilities or old instructions become hard to track

Picture preparing a client proposal. You open one conversation and put everything into it: research the client's industry, find their pain points, draft the proposal structure, estimate the timeline, flag the risks, match the company's tone, and write talking points for the meeting. The conversation gets long. The model starts blending research with its own assumptions. You lose a strategic decision between two formatting tweaks. You end up with something polished and hard to trust, because you cannot tell which parts were reasoned and which were the model filling gaps to keep the text moving.

Look for a specific failure before adding conversations. A rejected recommendation may return, a source finding may become an unsupported assumption, or an old formatting instruction may still affect the draft. A clearer brief may resolve the problem. A separate assignment can help when the responsibilities need distinct context or checks.

Fluent prose does not establish that the requirements were followed. Compare claims with sources and the draft with the current decisions. A separate reviewer can help make that comparison, but opening another conversation does not itself establish that a check was performed or that its verdict is correct.

A project coordinator can give several specialists bounded assignments, or ask one capable collaborator to complete a sequence with checkpoints. The useful principle is explicit responsibility: who gathers the facts, who checks them, and what the writer may infer from them.

Separate responsibilities while preserving their dependencies

In On the Role of Scientific Thought (opens in a new tab), Edsger Dijkstra describes as a way to order thinking. His 1974 discussion considers different aspects of a problem separately while still recognizing their connections. That distinction helps organize AI assignments; it is not evidence that every extra task divides a model's attention by a fixed amount.

The same distinction appears in familiar work. A film editor needs the footage before completing the edit, while separate camera teams may shoot independent scenes at the same time. Distinct responsibilities and simultaneous execution are different decisions. An AI coordinator needs to know both what belongs to each assignment and what must be available before it starts.

Reliable assignments support delegating more of their coordination. Ask the hub to preserve the roles, shared inputs, dependencies, and checks as a reusable workflow. Test that workflow on another project before relying on it. Later revisions can address the handoffs that failed or the context a specialist lacked.

Give each assignment a purpose and route its findings back to the goal

Separate responsibilities when they need different checks. Evidence gathering asks what the sources say. Recommendation asks what follows from that evidence and your goals. Review asks whether the proposal satisfies its criteria. These can be separate phases in one conversation or assignments in different contexts.

Separate the source material from the reasoning from the deliverable

Source material, reasoning, and presentation suggest possible assignments. Sources include documents, notes, and figures. Reasoning connects that material to conclusions. Presentation explains the conclusions for the audience. Use the distinctions where they help you specify or check a contribution; they need not become three conversations for every task.

Check a recommendation before relying on its presentation. A proposal may look professional while its estimate depends on an untested assumption. Require the assumption and supporting evidence to be stated, whether the assistant works in one session or several. A dedicated analysis pass gives that question a defined place.

Keep dependencies explicit when assigning the layers. Research can supply a source summary; analysis can use it to compare options; drafting can explain the selected approach. That sequence is staged, even if separate conversations do the work. Parallelism is available within it when two assignments can proceed without each other's results.

Hold the goal in a hub and give each spoke one assignment

A workflow gives coordination a clear home. One conversation holds the overall task, while each spoke receives an assignment. The pattern is useful when the hub can integrate the returned findings and the spokes can work from well-defined inputs.

The hub holds the current plan. Give it the goal, audience, constraints, and sources. Ask it to identify assignments, dependencies, and how each result will be checked. It can draft a starting proposal if that helps; later synthesis must account for the evidence returned by the spokes.

Each spoke needs a self-contained assignment. In a manual trial, paste the shared brief and the spoke's prompt into a new conversation, then supply the relevant source material there. Ask it to report any source it cannot access. Return its findings, source references, and unresolved questions to the hub. If you later automate the handoff, check that each assignment still receives the intended inputs and returns the evidence needed for synthesis.

Specify what each assignment may decide. A research spoke may report uncertainty; it should not quietly change an agreed budget. A drafting spoke may improve wording while preserving commitments. The hub reconciles their results against the project goals and brings unresolved decisions back to whoever has authority to make them.

Add conversations when a project needs independently verifiable work or a separate review context. Define each assignment and its inputs before opening it. The coordinating conversation reconciles the results. Keep work together when splitting it adds briefing and integration effort without a useful gain.

Follow a proposal from source review to synthesis

One client proposal split into a hub and three spokes

Shared brief for an illustrative proposal: 'A warehouse team spends an hour each morning copying order counts from a spreadsheet into a planning sheet. The operations manager wants to test whether this can be automated. The pilot budget is $3,000, existing tools must stay in place, and any trial must run alongside the manual process. Prepare a proposal for a two-week internal pilot. We have meeting notes and sample spreadsheets; we have no measured estimate of time savings yet.' Ask the hub to identify assignments that can use these inputs independently, explaining the boundaries and required checks.

Spoke 1, source review: 'Read the meeting notes and sample spreadsheets. List the manual steps, the fields copied, and any missing information. Cite where each fact appears. Do not infer how long an automated version will take.' An appropriate finding might be: 'The notes report an hour of copying; the samples show order counts by date and warehouse. Neither source measures time spent correcting errors.'

Spoke 2, audience review: 'From the meeting notes, identify the operations manager's stated decision criteria. Separate quoted concerns from your inferences.' A source-backed concern might be: 'Keep the manual process available during the trial.' The spoke should not invent resistance to automation simply because that concern sounds plausible.

Spoke 3, pilot risks: 'Review the brief and sample sheets for conditions the pilot must test. Return the condition, the evidence for it, and a proposed check.' For example: 'A column name may change between files. Test an altered sample and require the workflow to flag the mismatch instead of copying the wrong column.' This assessment can run before a draft exists because it evaluates the inputs and proposed task.

Hub synthesis: 'Combine the findings into the proposal. Preserve the $3,000 budget, two-week pilot, and parallel manual process. Require a comparison of copied counts and measured staff time before recommending rollout. Flag any unknown that changes the scope.' A resulting commitment could be: 'The pilot will measure copying accuracy and staff time; it does not promise a particular saving before the trial.'

The benefit is visible in the proposal's commitments. A reported hour of manual work has not become a promised hour of savings, and a technical possibility has not become an approved rollout. The separate assignments supplied checks that informed the final scope. They also cost coordination; compare that cost with the benefit when deciding whether to reuse the arrangement.

Check shared facts, assignment boundaries, and integration

Parallel work can fail in the shared brief, the assignment boundaries, or the integration. The following examples help you diagnose those problems. Each points to a concrete, testable change.

Family	Failure mode	What happens	Prevention
Brief	Context drift	Different conversations receive different versions of the project brief and begin solving different problems.	Give each spoke the current shared brief. If a decision changes, update the affected assignments before using their results.
Brief	Parallel wrongness	Several conversations move quickly, but they all work from a bad assumption in the shared brief. Speed amplifies the error.	Check consequential assumptions against sources, or assign their investigation before dependent work begins.
Boundary	Thread soup	A conversation mixes incompatible instructions or loses track of which result is being checked.	Separate the conflicting responsibilities or add a checkpoint where the task changes.
Boundary	Authority leakage	A spoke changes a shared decision that its assignment only authorized it to report or explain.	Tell each spoke explicitly what it may and may not decide.
Integration	Integration debt	Each spoke produces useful output, but the outputs do not fit together because the format or scope was underspecified.	Define the output format for each spoke before it begins.
Integration	Synthesis collapse	The hub receives findings without a method for resolving disagreements, missing evidence, or incompatible recommendations.	Always return outputs to the hub. Synthesize before finalizing.

Clear assignments, current shared facts, and deliberate integration address these failures. They do not eliminate them automatically. Ask the hub to check for conflicting versions, unsupported assumptions, and commitments that changed during a handoff before finalizing the result.

A separate review can test what a production pass missed. Give it the draft, criteria, and sources, then verify its findings. A different conversation or model can still share the original error; separation helps only when the check contributes useful evidence.

Reconcile conflicting findings before drafting the recommendation

When findings disagree, ask the hub to compare their sources and assumptions. In the illustrative warehouse pilot, a spoke might estimate savings from the reported hour of copying, while another flags correction work that the records never measured. Presenting either assessment as measured savings would overstate the evidence. The hub can propose measuring both copying and correction time during the trial. Check whether that proposal resolves the uncertainty that matters to the decision, and supply any relevant constraint the hub lacks.

Give the hub the facts that should change its recommendation. A budget limit or commitment already made to a client can determine which proposal is acceptable. Ask the hub to show how those constraints affect its synthesis, then check the reasoning against the source material.

Stay in one conversation when a task fits there

Stay in one conversation when it can complete and check the task reliably. A short email or simple summary often does not justify separate assignments. A focused second review may still help when it catches a consequential error. Let the task and the evidence determine the amount of coordination.

Clarify enough of the goal to assign useful work. Ask the hub to identify the uncertainty, the evidence needed to resolve it, and which investigations can proceed independently. A vague request to "research everything" needs a more specific question before extra conversations can help.

Uncertain inputs need investigation before they support a conclusion. Parallel source checks can help investigate them; parallel drafts that repeat the same unverified assumption will not resolve the uncertainty. Also check whether extra assignments are postponing a decision that already has enough evidence.

Test a coordination plan on independently verifiable project work

Design workstreams around your project’s actual dependencies

Open a coordinating conversation in Claude and bring the project goal and sources. Ask for justified assignments and their handoffs, then test the plan in separate conversations where useful.

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.

Problem: A project may benefit from separate research, analysis, drafting, or review assignments, but their inputs and results need coordination. Intent: I want to design a hub-and-spoke workflow for my project, using this conversation as the hub and adding separate conversations where they serve a clear purpose. Ask me for the project, intended audience, goal, constraints, and source material. Use the information I provide and ask focused follow-up questions where a missing fact would change the assignments. Distinguish confirmed decisions from assumptions and open questions. Do not invent my priorities or source findings. Once you have enough context: 1. Draft a SHARED PROJECT BRIEF with the goal, audience, constraints, confirmed decisions, source references, and required final deliverable. State which sources each assignment will need to receive or access. 2. Propose SPECIALIST or REVIEW assignments where separate context would help. For each, state the purpose, required inputs, output format, checks, and decisions it may make. Explain what separation adds. Keep work in this conversation when a separate assignment offers no clear benefit. 3. Identify dependencies. State which assignments can run concurrently and which must wait. A review of a draft needs that draft; an investigation of the brief's assumptions may begin earlier. 4. Write a self-contained opening prompt for each justified spoke. Include the shared brief and its specific assignment. If required source material is missing, have the spoke request it before drawing dependent conclusions. Tell me what to paste or attach in each conversation and what findings, source references, and unresolved questions to bring back here. 5. Draft a SYNTHESIS PLAN. Explain how this hub will compare returned evidence, resolve disagreements where sources allow, preserve confirmed commitments, and identify decisions or missing information that still prevent a conclusion. Include a check for changed assumptions or outdated versions of the brief. 6. Draft a WORKSTREAM MAP recording the roles, inputs, dependencies, output requirements, and checks. Give me a trial plan for checking the handoffs and comparing the combined result and total effort with a one-conversation approach. Return the plan and prompts here. Do not launch the assignments or draft the final deliverable. Treat the workstream map as untested until I supply trial results. Use those results to propose improvements, and ask for confirmation before saving a reusable personal workflow.