Idea Generation

A funnel, not a screen dump. A name that clears a filter is a candidate: it has earned the next hour of work, nothing more. The output is a short ranked list where each name says why it surfaced, what would have to be true, and what would kill it, plus the names that screened well and were rejected, which is where most of the learning sits.

Evidence labels, source tiers, staleness, the readiness posture and the intake limits: .agents/skills/research-conventions/SKILL.md, read before the first deliverable.

Step 1: Mandate and universe

Read the investor-mandate memory first per .agents/skills/research-conventions/references/intake.md, then ask only about the forks it leaves open. Record every slot, with a disclosed default where nothing fixes it:

Slot	Default when unstated
Mandate and vehicle	long-only fundamental
Objective	absolute return over a multi-quarter horizon
Direction	long, plus shorts only where the mandate allows them
Horizon	two to eight quarters, which is what the screens below are calibrated for
Geography and listing	the user's home market, primary listings
Market-cap range	above the liquidity floor, otherwise unrestricted
Liquidity floor	a minimum daily traded value the position size could clear in a few days
Benchmark	the mandate's index, or none stated
Instruments and constraints	common equity; restricted lists, borrow availability and position limits taken from the user when they exist
Theme	none, unless the request names one

Done when every slot carries a value or a default that the delivery message discloses.

Step 2: Validate the universe before anything is ranked

Screen metrics lie on a universe nobody cleaned: