Build a private pre-market scan digest from your own written rules

How it works

1. Write your screening criteria in plain English in a rules file: your own trend criteria, distance from the 10, 20 and 50-day moving averages, relative volume, minimum liquidity.
2. Point a script at your own universe and licensed data source. The script must compute every value and output explicit boolean columns (meets_rule_trend, meets_rule_volume, etc.) for each rule.
3. The agent only reads those pre-computed boolean flags to group; it performs no rule interpretation.
4. It adds a chart link for each name so you can check it yourself, and notes any data gaps (holidays, missing bars).
5. It sends the digest privately to you (email or chat) before the open. Nothing is posted publicly.
6. Start each digest with the rules-file version and the data timestamp so you can see what produced it.

Copy-paste prompt

You are my pre-market scan assistant. You do not give trading advice and you never recommend positions.
Inputs: {{rules_file}} (my written screening rules) and {{scan_output_csv}} (produced by {{scan_script}}).
1. Use only values from {{scan_output_csv}}. Never estimate, round differently or type a number yourself.
2. Group rows using the pre-computed boolean rule columns in the CSV and nothing else. Do not re-evaluate, interpret, or apply any rules yourself. If a boolean column is missing, note it and do not guess.
3. List any row with missing data separately.
4. For each row add the chart link column from the CSV. Do not add commentary about future price moves.
5. Start the digest with the rules-file version and the data timestamp.
6. This digest is private to me. Do not format it for social media.

Guardrails

• All numbers computed in code, never typed by the model.
• Script pre-computes every rule match as boolean flags; model never interprets rules or applies logic to raw numbers.
• Output is a private review list for you only; never publish it.
• Rules live in a versioned file so you can see what changed.
• Check your data provider's licence before storing or sharing anything derived from it, and confirm your specific provider explicitly permits derived data to be sent to third-party LLMs.

What can go wrong

A common failure is pasting raw prices into a chat and asking the model to work out moving averages: it returns confident, wrong numbers. Another is stale data on market holidays producing an empty or duplicated scan. A third is the model silently mis-grouping names because it re-interpreted the English rules instead of using script-provided boolean flags; always verify the script outputs explicit rule-match columns. Compute in code and print the data timestamp at the top.

Source and references