Stop publishing job descriptions that demand LLM fine-tuning, RAG pipelines, MLOps, data engineering, React, Kubernetes and 'strong product sense' in one human, because that posting doesn't attract the unicorn, it attracts the people willing to claim they're one. Real experts read a ten-skill wish list, correctly infer that you don't know what the job is, and close the tab. Bluffers apply to everything. Your unicorn JD is a bluffer filter running in reverse.
What the unicorn JD actually selects for
- Experts apply an honesty discount: seeing ten requirements and holding eight, they pass, precisely the calibration you wanted to hire.
- Bluffers apply a confidence premium: holding four, they claim ten, and your screening now depends on catching them.
- The requirements list usually smuggles in three different jobs, an ML engineer, a data engineer and a product engineer, priced as one salary.
- Vague maximalism also signals chaos: strong candidates read 'we haven't scoped this role' as 'you'll be doing everything, appreciated for nothing'.
How to rewrite it
- 1Write the 90-day sentence first: 'This person will have shipped X.' If you can't write it, you're not ready to hire.
- 2Pick the one core strength that 90-day outcome requires, e.g. 'has shipped an LLM feature to production users'.
- 3Cap must-haves at three; move everything else to 'useful context', explicitly optional.
- 4Name the team, the stage, and how decisions get made, experts choose environments, not adjective lists.
- 5Cut every requirement that's really a different role, and decide whether that role is hire #2 or a fractional need.
