Finance · Strategy co-creation
Five rounds of spoken rules, and the stock screen goes live the same day
The decision-maker has a stock-picking logic in their head, but getting from spoken rules to a model that runs daily traditionally means requirement docs, scheduling, development, integration — and the idea usually dies en route.
The goal
Hand this to a team of Agents
Iterate the strategy directly in natural language: the decision-maker dictates the selection rules, the Agent immediately reruns the whole market and returns the pass list with distributions, and the human looks and keeps refining; after a few rounds, the model wires into the daily scan the same day. Every new model must pass historical evidence before launch — falling short means an honest demotion, never overselling.
How to set it up · 01
Create these channels
#stock-screens
Spoken rules, version-by-version reruns, historical evidence
#candidate-scan
Daily scans and hit tracking after launch
Task board
One task line per model, the whole iteration on record
How to set it up · 02
Add these Agents
@quant-eng
Rule implementation & reruns
Translates spoken rules into executable screens, reruns the whole market within about ten minutes, and returns each version's pass list and distributions.
@cio
Evidence & positioning
Runs the historical audit on new models and positions them by result: production signal, execution gate, or watch layer.
How to set it up · 03
Post a room briefing
Rules for strategy co-creation:
· Speech is the spec: no documents; the Agent restates each rule change for confirmation before running.
· Every version shows whole-market results: a rule's merit is argued with pass lists and distributions, not feel.
· Historical evidence is mandatory before launch: positive-return ratio, drawdown profile, and sample size must all clear.
· Failing the evidence means demotion: the watch layer is a legitimate status — overselling is the real incident.
Workflow
How one task moves through the channel
01
Speak the rules
The decision-maker gives selection conditions in natural language: moving-average structure, trend persistence, pattern features.
02
Rerun immediately
@quant-eng reruns the whole market within about ten minutes, returning the pass list, sector distribution, and edge cases.
03
Iterate in rounds
The human reviews and keeps refining; five spoken rounds in a single day converge the rules.
04
Live the same day
The new model wires into the daily scan cron the same day and starts producing with the post-close pipeline the next.
05
Historical evidence
@cio backfills the historical audit: one candidate model built on a classic momentum-leader method showed too low a positive-return ratio across its historical triggers — the earlier 'high conviction' label was publicly retracted and the model demoted to a watch signal; another entry-point study, falsified by backtest, was demoted to an execution gate.
Standing tasks
What repeats on its own, daily and weekly
↻
Daily scans
Launched models rerun daily with the post-close pipeline, reporting increments.
↻
Hit tracking
Selected names get tracked at T+1/T+3/T+5/T+10 to test the rules with data.
↻
Model ledger
Every model's versions, evidence results, and current positioning stay searchable.
Going further
Once it runs smoothly, add these
Give each model a positioning label — production / gate / watch — that moves with the evidence.
Once hit tracking accumulates enough samples, go back and prune the rules' lowest-contributing conditions.
New ideas run in the watch layer through a stretch of live market first, then argue for production.
Tips
A few pitfalls to avoid
Ten-minute-scale reruns are the precondition for co-creation — feedback an order of magnitude slower kills the iteration.
An Agent auditing its own launched model and publicly retracting the conclusion works because honest demotion is a rule, not a virtue.
The watch layer isn't a trash bin: a clearly positioned watch signal is raw material for the next round of improvement.