AI drafts. People decide.
Scope a useful AI drafting task, preserve source evidence, validate the output, and keep editorial authority with the reviewer.
Read the guide07 / REVIEWABLE INTELLIGENCE
AI dapps.Useful drafts. Human judgment.
Start with an inspectable task: a draft, summary, glossary check, or explanation. Keep model suggestions separate from publication and execution authority.
Choose a result a person can evaluate
A release-note summary is a better starting task than “manage our documentation.” Define what the output must preserve, what it must not invent, and what the reviewer should do when the source is incomplete. Keep the original material visible beside the draft so evaluation does not depend on fluent wording alone.
Treat retrieved text as evidence
Documents can contain instructions that are irrelevant or hostile to the intended task. OWASP describes this class of problem in its prompt injection guidance. Keep source material separate from the authority to choose tools, change destinations, or publish. Review the paths through which input can become markup, links, or action parameters.
Validate before publication
Check facts, links, identifiers, and examples against the reviewed source. Separate style edits from technical approval. Preview safety-relevant instructions beside the interface action they describe. A correctly formatted draft is not proof that its claims are correct, and a confident tone is not evidence of compatibility or permission.
Evaluate the workflow when it changes
Keep representative examples: complete sources, incomplete notes, contradictory wording, and irrelevant embedded instructions. Repeat the evaluation when the model, prompt, source collection, or output handling changes. Assess unsupported additions and omitted facts as well as editing effort. Narrow the task when the reviewer cannot reliably verify the result.
OWASP’s prompt injection guidance describes risks from instructions embedded in model inputs and recommends layered controls.
Read the prompt injection guidanceQuestions to settle early
Does AI assistance require an onchain model?
Not for the editorial workflows described here. First identify the task and its evidence requirements; then decide which components, if any, need network execution.
When does assistance become an agentic workflow?
When the system gains capabilities to act through tools, such as publishing or changing configuration, review the additional permissions and authorization boundary separately.
READ NEXT / DAPP CMS LAB
All Lab notes Go deeper in the Lab
Give agents boundaries
Separate plans, proposals, approvals, and execution. Give agents narrow tools, external authorization checks, and a tested stop control.
Read the guide
Model content, not transactions
Structure guides, network references, and notices with clear permissions, review states, validation, and an exportable publishing workflow.
Read the guide