All case studies

LinkedIn · AI-assisted editorial workflow

Connecting search demand, prompting and content QA at LinkedIn

Keyword research, prompt development and a content-quality feedback loop for Collaborative Articles, working with an early, unreleased GPT-4 prototype.

Keyword research, prompt development and content QACollaborative Articles launch programme4 minute read
Search demandkeyword research informing the questions
Prompt designpersonal responsibility for building prompts
Human QAreview findings fed back into prompting

The engagement in context

A contribution to LinkedIn’s Collaborative Articles programme, covering keyword research, prompt development and content QA with an early GPT-4 prototype.

A new way to combine AI and professional experience

LinkedIn introduced Collaborative Articles in 2023. AI-powered conversation starters provided a foundation that professionals could add to with their own experience and advice. The product brought together generated content and contributions from people with relevant knowledge.

My part of that programme sat where search research met content production: identifying the questions to address, building prompts and reviewing the completed content.

The prototype was capable, but less reliable than expected

The biggest challenge was working with an as-yet unreleased GPT-4 prototype. Early access was useful, but it did not mean the output could be accepted without close review. The model was not as reliable as I had hoped.

That made prompt development an iterative editorial task. A plausible-looking result was only a starting point. The important question was whether the completed content served the question that had prompted it, and what the review taught us about the instructions.

My responsibility connected research with the finished content

I was personally responsible for the keyword research and for building the prompts. The research helped establish the subjects and questions people might use in search, giving the generation process a purpose beyond producing more text.

I then handled QA of the completed content. Keeping those responsibilities connected meant I could consider the output in the context of the search question and the prompt that produced it, rather than reviewing it in isolation.

  • Keyword research to identify relevant search questions.
  • Prompt development to turn those questions into generation instructions.
  • QA of completed content.
  • Feedback from that review into further prompt development.

QA became an input to the next prompt

The work did not stop when a draft was generated. I fed the QA findings back into a feedback loop to improve the prompting. Research, generation and review were connected stages of the same process.

The practical value of that loop is that a review can inform the next attempt as well as the current draft. It creates a way to examine the relationship between the instructions and the output, while retaining human judgment over what is useful.

A connected process from search question to content review

The contribution combined keyword research, working prompts, content QA and feedback into the next iteration. It connected what people were looking for with how the content was generated and assessed.

For a team developing an AI-assisted content process, that connection matters. A prompt needs a clear editorial purpose, and the review needs a route back into the instructions. My role covered both sides of that loop.

Build the review process into the workflow

The experience reinforced a lesson that applies well beyond one model: access to a capable system does not remove the need to define the task and review the result. Prompting works better as part of a research and quality process than as a one-off request for content.

For your team, the useful starting questions are practical. What should the output help a reader do? Who can assess it? How will findings change the next run? Those decisions belong in the workflow before automation is expanded.

Connect your content research to a workflow your team can trust.

Get an approach shaped around your evidence, constraints and goals. Share what your team needs to decide and explore what a useful first step would involve.