Markdown edition

# Set up a four-role AI marketing team and review one campaign

> Configure a Paperclip CEO, researcher, marketer and designer around one owned campaign, approve research before creative work, and check the task and run records before using the result.

By BIGCHANGE.AI Editorial

Published: 2026-10-24T20:03:39.077Z
Updated: 2026-10-24T20:03:39.077Z
Canonical: https://bigchange.ai/blog/paperclip-four-role-marketing-team-campaign-guide

![Conceptual charcoal illustration of two hands reviewing an open four-section ring binder, with a pencil poised above a blank ruled page.](https://bigchange.ai/api/media/file/paperclip-private-campaign-binder-review-hero-v3.png)
AI-generated conceptual illustration by BIG CHANGE.

Paperclip can organize AI agents into a reporting chain, attach their work to a project and route tasks through review. For a marketing campaign, the useful setup is a CEO, researcher, marketer and designer working from the same approved brief, with a person checking each handoff. This guide follows [Paperclip's current documentation](https://docs.paperclip.ing/guides/org/org-structure/), checked 24 October 2026. BIG CHANGE did not install Paperclip, configure agents, run a model or produce a campaign.

Bring an actual product and audience, an approved campaign objective, channels and call to action, a deadline, a private workspace you control, source links for product claims, and rights to every proposed asset. You also need authority to use Paperclip and a supported Claude Code sign-in or provider account. Keep customer data, passwords and unpublished claims out of prompts and shared task text.

## The big change

- **What changed:** Paperclip provides a reporting chain, project tasks, run records and review policy that can keep four agent roles on one bounded campaign.
- **Why it matters:** A researcher can pass dated evidence to the copy and design tasks, while a person checks the source, rights and final package before any campaign material is used. The task records make that handoff inspectable.
- **What to watch:** Saved review policy governs completion, while an agent's permitted tools can act during a run. Check actual permissions, run limits and charges before resuming each role.

## 1. Install Paperclip and write one campaign brief

Use the [current installation guide](https://docs.paperclip.ing/guides/getting-started/installation/) for your chosen local or self-hosted route. Its local instructions require Node.js 24.11.0 or newer and pnpm, and warn against running the embedded PostgreSQL setup as an administrator. Check the installed Paperclip version, host, account and model billing. Install Claude Code separately and use a supported authentication route for the [Claude Code adapter](https://docs.paperclip.ing/reference/adapters/claude-code/). Its **Test Environment** performs a real model probe, which can incur provider usage; this article reports no such test. The documentation's example screens and paths may differ on your host.

Create an organization, then add a bounded campaign goal under **Goals**; the [first-company guide](https://docs.paperclip.ing/guides/getting-started/your-first-company/) says the creation wizard asks only for a name. Create or select one [project](https://docs.paperclip.ing/guides/projects-workflow/projects/) and attach an owned [workspace](https://docs.paperclip.ing/guides/projects-workflow/workspaces/). Confirm the actual absolute working directory before work begins. Save a brief with the product, audience, goal, channels, deadline, permitted call to action, approval owner and output paths. For each factual claim or asset, record its source URL, date, owner, usage right and approval status. A private research table, draft copy, design concept or asset plan, and final human decision are a useful completion target; sales or reach are separate outcomes to measure later.

## 2. Configure four roles, then pause all four

Create a CEO agent and three agents titled **Researcher**, **Marketer** and **Designer**, each with the supported Claude Code adapter. In **Harness / Runtime**, confirm the three specialists report to the CEO using Paperclip's [manager field](https://docs.paperclip.ing/guides/org/org-structure/). Give each role a narrow instruction:

1. Researcher: return a dated claim-and-source table, with uncertainty and missing evidence visible.
2. Marketer: draft the specified channel copy only from approved claims, linking each substantive statement to a research row.
3. Designer: produce a concept, layout or asset plan from supplied assets whose rights are recorded; label anything still to be made.
4. CEO: route the bounded tasks, compare the actual outputs, flag conflicts and submit the combined package for human decision.

Before creating or assigning a campaign task, **Pause all four agents**. Read back `paused` for each and check that no campaign run is queued or running. Turning off interval heartbeats alone is insufficient: assignments and comments can wake an agent. [Paperclip's agent guide](https://docs.paperclip.ing/guides/org/agents/) allows a board user to stage work for a paused agent, but an agent cannot assign work to one. The board operator will assign the specialists while they remain paused.

## 3. Set permissions and spending limits before resuming

In **Harness / Runtime**, set `dangerouslySkipPermissions: false` separately for **each** adapter, save and read back all four effective values. The [adapter reference](https://docs.paperclip.ing/reference/adapters/claude-code/) documents a `true` default. For the default ACP engine, keep `nonInteractivePermissions: deny` or choose `fail`; an action requiring interactive approval then stops in a headless run. Inspect the effective tool list, local file access, workspace trust and [Claude Code permission rules](https://code.claude.com/docs/en/permissions/). Project `allow` rules are not active before workspace trust. Give only the local permissions needed for the owned campaign folder and the required task actions. If an essential action is denied, record the exact conflict and hold that task; do not enable permission bypass as a fix.

Prefer no external connector for this private draft. If one is needed, inspect its account and actions, set its Write group **Off** where appropriate, and read the **Allowed / Ask first / Off** summary in [Paperclip's action-permission guide](https://docs.paperclip.ing/connectors/action-permissions/). Those connector controls do not restrict shell commands or credentials exposed to a shell. Use a dedicated workspace and account without unrelated service tokens. Do not grant public posting, email, advertising, publishing or destructive actions. A task review gate controls completion; it cannot undo a tool action already taken.

Set a finite `maxTurnsPerRun` and positive `timeoutSec` for each role, then read back **Advanced Run Policy** with **Continue after max-turn stop** off. The adapter documents a 300-turn default and, for local or SSH work, no wall-clock limit when `timeoutSec` is zero. Twenty turns and 300 seconds could be an owner's small first-run choice; they are not measured requirements. Under **Audit → Budgets**, verify an organization cap and each agent cap. [Paperclip's cost guide](https://docs.paperclip.ing/guides/day-to-day/costs/) says setup does not create every per-agent cap. Check provider billing and allowances separately: a task prompt or Paperclip budget is not proof of a provider bill ceiling. Keep interval heartbeats off and all agents paused until their exact tasks are ready.

## 4. Stage and review research before creative work

A board operator creates the parent campaign task from the approved brief and a first child for research. Assign them to the paused CEO and researcher. In each existing task's **Properties**, set and reopen the parent, project, assignee and human Reviewer or Approver; read back the saved `executionPolicy`. Paperclip's [Tasks](https://docs.paperclip.ing/guides/day-to-day/issues/) and [Execution Policy](https://docs.paperclip.ing/guides/power/execution-policy/) pages support review fields on existing tasks. An `@` mention supplies context; it is not an assignment. Check the exact brief revision, rights, output paths and run limits before any resume.

Resume the CEO only for its staged, bounded routing or strategy task. If it submits a strategy, have the person decide on it before proceeding; pause the CEO again while children are staged. Because the specialist is paused, the board operator directly creates or assigns the research child, then deliberately resumes the researcher. Do not assume a CEO comment or direct message dispatched anyone.

When research finishes, inspect its dated source table and open the cited primary pages. Check unsupported, stale or conflicting claims and the actual run and usage record. A configured review stage sends an executor's `done` to `in_review`; record the human decision and the exact approved research revision. Keep the Marketer and Designer paused until that approval. Then have the board operator create their child tasks from the approved brief and research revision. Set each task's human review participant in **Properties**, reopen and verify its policy and assignee, and only then resume the relevant agent. A creative draft built from an unreviewed research revision remains on hold.

## 5. Reconcile the campaign and record the decision

Compare each marketing claim and call to action with the approved brief and cited source row. For the design output, check asset origin, rights, sizes and whether each visual is finished or a placeholder. Open task comments and **Agents → Runs** transcripts to see what each role actually did. Have the CEO reconcile the research, copy and design outputs only after its parent task is safely staged and deliberately resumed. The person reviews each required task and accepts or rejects the combined campaign with an explicit decision.

If an unexpected wake is queued, use its Cancel control and verify cancellation. If a run has started, Cancel the run, Pause the agent, then verify the run ended before editing its input or permissions. Pause prevents future wakes; it does not establish that an in-flight action stopped. Inspect touched issues, transcript, output files and connector activity for side effects. Correct a failed task in its saved thread, recheck scope and policy, and resume deliberately. Leave denied actions, missing sources, unclear asset rights or unresolved review conflicts as named holds.

The acceptance record should show all four agent IDs and their reporting lines, the project and approved brief revision, each task's owner and saved review policy, the source table and approved revision, copy and asset-plan files, run IDs, observed usage and charges, and human review decisions. Record the Paperclip version, adapter model IDs and permission readbacks from your own environment. Keep unknown fields blank until observed. Completion means a reviewed **private campaign package**; posting, outreach, ad buying and publication need their own authorization.

Use this blank record for your own run. It is a suggested evidence sheet, not a Paperclip configuration file or a report of a run BIG CHANGE performed. Fill fields from saved task, agent and run pages and actual output files; leave an unknown blank and hold the decision.

```yaml
checkedAtUtc: ""
paperclipVersionOrCommit: ""
organizationId: ""
projectId: ""
workspaceAbsolutePath: ""
campaign:
  owner: ""
  approvedBriefPath: ""
  approvedBriefRevisionOrHash: ""
  productAndAudience: ""
  goalAndChannels: ""
  allowedCallToAction: ""
  rightsAndDataBoundary: ""
agents:
  ceo: {id: "", reportsTo: "board", statusBeforeStaging: "", statusBeforeRun: ""}
  researcher: {id: "", reportsToId: "", statusBeforeStaging: "", statusBeforeRun: ""}
  marketer: {id: "", reportsToId: "", statusBeforeStaging: "", statusBeforeRun: ""}
  designer: {id: "", reportsToId: "", statusBeforeStaging: "", statusBeforeRun: ""}
adapterForEachAgent:
  engine: ""
  actualModelId: ""
  authenticationRouteNoSecret: ""
  cwdReadback: ""
  dangerouslySkipPermissionsReadback: ""
  nonInteractivePermissionsReadback: ""
  allowedLocalToolRulesAndTrustReadback: ""
  connectedServicesAccountsAndActionsReadback: ""
  maxTurnsPerRunReadback: ""
  timeoutSecReadback: ""
  continueAfterMaxTurnStopReadback: ""
  organizationBudgetReadback: ""
  agentBudgetReadback: ""
  providerUsageBoundaryReadback: ""
tasks:
  parent: {id: "", assigneeId: "", projectId: "", reviewerIds: "", approverIds: "", executionPolicyReadback: ""}
  research: {id: "", assigneeId: "", reviewerIds: "", executionPolicyReadback: "", approvedOutputRevision: ""}
  copy: {id: "", assigneeId: "", reviewerIds: "", executionPolicyReadback: "", approvedResearchRevisionUsed: ""}
  design: {id: "", assigneeId: "", reviewerIds: "", executionPolicyReadback: "", approvedResearchRevisionUsed: ""}
outputs:
  sourceTablePathAndDate: ""
  checkedClaimIdsAndSourceUrls: ""
  rightsCheckedAssetIdsAndOrigins: ""
  marketingDraftPath: ""
  designConceptOrAssetPlanPath: ""
  ceoReconciliationPath: ""
runs:
  ceoRunIdsAndStatus: ""
  researcherRunIdsAndStatus: ""
  marketerRunIdsAndStatus: ""
  designerRunIdsAndStatus: ""
  observedTurnsTimeoutsUsageAndCharges: ""
  deniedActionsErrorsOrUnexpectedRuns: ""
  cancelPauseReadbackAndSideEffectCheck: ""
  correctiveTaskIdAndDeliberateResumeReadback: ""
humanDecision:
  reviewerIdentity: ""
  taskReviewDecisionIds: ""
  finalCampaignDecisionAndTimestamp: ""
  remainingHold: ""
```

## What's next

Check the final package against the original brief and retain the exact task and run records alongside it. If a later revision changes a claim or asset, route that revision through the same research and human review steps before use. For a narrower creative-file task, our [Pomelli campaign guide](https://bigchange.ai/blog/pomelli-checked-campaign-creatives-guide) covers checked product creatives; the [Replit go-to-market guide](https://bigchange.ai/blog/replit-agent-three-function-gtm-tool-guide) covers a three-function app from one brief.

## Sources & further reading

- [Paperclip installation](https://docs.paperclip.ing/guides/getting-started/installation/) and [first-company setup](https://docs.paperclip.ing/guides/getting-started/your-first-company/) — Checked 24 October 2026. Establish local prerequisites and the organization/goal sequence. Your installed version, host and account still need their own readback.
- [Organization structure](https://docs.paperclip.ing/guides/org/org-structure/), [agents](https://docs.paperclip.ing/guides/org/agents/), [projects](https://docs.paperclip.ing/guides/projects-workflow/projects/) and [workspaces](https://docs.paperclip.ing/guides/projects-workflow/workspaces/) — Checked 24 October 2026. Document manager relationships, pause/assignment behavior and workspace binding. They do not establish that a reader's four agents have been configured or run.
- [Tasks](https://docs.paperclip.ing/guides/day-to-day/issues/) and [execution policy](https://docs.paperclip.ing/guides/power/execution-policy/) — Checked 24 October 2026. Support existing-task Properties review fields and the `in_review` handoff. A saved review policy and an actual human decision must be checked in the reader's tasks.
- [Claude Code adapter](https://docs.paperclip.ing/reference/adapters/claude-code/), [costs and budgets](https://docs.paperclip.ing/guides/day-to-day/costs/) and [connector action permissions](https://docs.paperclip.ing/connectors/action-permissions/) — Checked 24 October 2026. Define adapter defaults, finite run settings, budget controls and connector permissions. They do not measure a campaign run or provider charge, and connector policy alone does not contain shell access.
- [Claude Code permissions](https://code.claude.com/docs/en/permissions/) — Checked 24 October 2026. Documents effective local tool rules and workspace trust. A permission configuration still needs a host-specific readback before work resumes.

## Sources

- [Paperclip installation](https://docs.paperclip.ing/guides/getting-started/installation/) — Checked 24 October 2026. Local Node/pnpm prerequisites and embedded database caution; installed reader environment unverified.
- [Your first company](https://docs.paperclip.ing/guides/getting-started/your-first-company/) — Checked 24 October 2026. Organization wizard and later goal creation; no reader organization was created.
- [Organization structure](https://docs.paperclip.ing/guides/org/org-structure/) — Checked 24 October 2026. Agent reporting relationships and manager field; four-role campaign structure is an editorial application.
- [Agents](https://docs.paperclip.ing/guides/org/agents/) — Checked 24 October 2026. Pause, run and board assignment behavior; no campaign agents ran.
- [Projects](https://docs.paperclip.ing/guides/projects-workflow/projects/) — Checked 24 October 2026. Project task and budget context; actual project remains reader-supplied.
- [Workspaces](https://docs.paperclip.ing/guides/projects-workflow/workspaces/) — Checked 24 October 2026. Workspace binding and absolute working-directory behavior; host-specific path needs readback.
- [Tasks](https://docs.paperclip.ing/guides/day-to-day/issues/) — Checked 24 October 2026. Existing task Properties and assignee readback; no task was created for this article.
- [Execution policy](https://docs.paperclip.ing/guides/power/execution-policy/) — Checked 24 October 2026. Saved review fields and in_review path; actual human decision is required.
- [Claude Code adapter](https://docs.paperclip.ing/reference/adapters/claude-code/) — Checked 24 October 2026. Permission default, cwd, Test Environment, turns and timeout; no adapter test was performed.
- [Costs and budgets](https://docs.paperclip.ing/guides/day-to-day/costs/) — Checked 24 October 2026. Organization and agent budget controls; no charge was measured or provider bill bounded here.
- [Connector action permissions](https://docs.paperclip.ing/connectors/action-permissions/) — Checked 24 October 2026. Connector action controls and their shell-credential limit; effective reader grants need inspection.
- [Claude Code permissions](https://code.claude.com/docs/en/permissions/) — Checked 24 October 2026. Local tool permissions and workspace trust; actual host configuration unverified.