The problem is an inconsistent source of truth
Independent prompts can produce contradictory campaigns even when every draft sounds plausible. A content workflow may describe a feature that has not shipped; a social workflow may target a different buyer from the landing page. Identify which statements are approved, where they came from, who owns them, and when they should be reviewed. A modest, maintained brief is useful before building a large knowledge system.
Separate durable facts from campaign choices. Product capabilities, pricing, security claims, and customer proof need accountable owners. Audience hypotheses and channel experiments may change more often. Each specialist should receive a task-specific slice of context, relevant brand rules, and facts it must not invent. Giving every step every scraped page increases noise and hides stale information.
| Context layer | Example | Review owner |
|---|---|---|
| Approved facts | Features, pricing, legal wording | Product or operations |
| Audience evidence | Interviews and objections | Research lead |
| Campaign choices | Offer, segment, channel | Marketing lead |
| Output history | Draft, source links, edits | Publishing owner |
Design handoffs as contracts
Define what each specialist receives and returns. Research can provide an evidence table with URLs and dates. Content can turn that into an outline with claims mapped to evidence. Social can adapt approved points to the channel. A reviewer checks tone and links. Include an explicit “unknown” value so a missing fact gets flagged rather than filled with a guess.
For a launch, approve the product brief, extract buyer objections, draft the landing page, create channel adaptations, run claim checks, then publish. If the offer changes midstream, mark prior drafts stale. Versioning can begin with a document revision and timestamp. What matters is that downstream work can identify the facts it used.
Evaluate the system
Sample outputs across channels. Do they agree on the buyer, offer, product limitations, and proof? Track time spent fixing unsupported claims, the share of drafts requiring major correction, and the age of source facts. This is more useful than claiming “zero drift” from a few examples. A good workflow makes disagreement visible early and gives a human a clear way to correct the source.
Decide what belongs in shared memory
The common brief should be small enough to audit. Begin with a product fact sheet, approved positioning, audience evidence, current offer, exclusions, and source links. Each item needs an owner and a freshness rule. A permanent product constraint may be reviewed on release; a campaign claim may expire at the end of the quarter. Store experimental insights as hypotheses so they cannot silently become approved facts.
When a specialist begins work, assemble a task packet rather than dumping the entire memory into its prompt. A landing-page writer needs the buyer problem, approved claims, proof, and page objective. A technical auditor needs URLs, performance criteria, and environment details. Both need the same rules for identifying unknowns and attaching evidence, but they do not need all the same source passages. This makes output easier to review and reduces irrelevant context.
Test the handoff with failure cases
Run a tabletop exercise before trusting the workflow at scale. Change a product feature in the source brief and check which drafts become stale. Remove a citation and see whether the next step flags the unsupported claim. Give two agents conflicting audience assumptions and see whether the supervisor asks for resolution. These tests reveal more than a demonstration in which every input is clean.
Record a small audit trail: source revision, task request, generated draft, human edits, and publication decision. Reviewers should be able to answer why a statement appeared and who approved it. If a downstream asset cannot be traced, pause that asset and repair the handoff contract. Shared context is useful when it makes correction cheap and visible, not when it makes all channels repeat the same mistake.
Questions to settle before expanding the agent system
What happens when two sources disagree?
Do not ask downstream writers to choose whichever sounds better. Mark the claim unresolved and route it to the accountable owner. Product documentation may govern availability, while a campaign brief governs framing; a dated source can supersede an older one only when its authority is clear. Record the resolution and invalidate outputs that used the old fact. The system should prefer a visible pause over a polished contradiction.
How much context should each specialist receive?
Enough to complete the task and defend its claims, but not every document in the workspace. Start with approved facts and relevant source passages, then include the audience, goal, exclusions, and output format. Ask the specialist to cite its source facts and list unknowns. Reviewers can request more context when the draft exposes a gap. This is easier to audit than a huge prompt whose useful facts are buried among unrelated pages.
Copy this handoff contract
Write task owner, requested output, audience, approved-facts revision, source links, prohibited claims, required evidence, unknowns, and reviewer. Define what the next step may change and what it must preserve. On completion, store the draft with the source revision and approval decision. If the brief changes, search for all dependent assets and notify their owners before any stale copy reaches a channel.