Check exit-rule adherence after the close

How it works

1. Write each exit rule as a structured row: rule_id, type (close_below_stop | close_below_ma | time_stop), threshold or ma_length, applies_to. Use simple, unambiguous conditional statements with explicit numeric thresholds.
2. A Python script reads {{positions_csv}} plus a daily-close file you already export (symbol, date, close). A broker position CSV does not include moving averages. The script computes days_held and any required MA, compares each rule, and writes {{eod_values_csv}} with columns symbol, rule_id, value, threshold, met (true/false), missing (true/false). The script, not the model, sets met.
3. The agent receives the pre-computed values and your exact rule statements, then lists per position which rules are met or not met.
4. You review and decide; the agent has no order access.
5. The agent shows the journal draft; you fill in the decision and append it yourself. The agent drafts a journal entry recording the decision you state and shows it to you as a preview; nothing is written to your journal until you confirm.

Copy-paste prompt

Using {{eod_values_csv}} (met/not-met already computed; do not recompute, do not infer missing numbers) and {{exit_rules_file}} for labels only, list each open position and each rule as met, not met, or unchecked if missing=true, with value and threshold copied from the file. Do not add rules, do not interpret chart patterns, do not recommend an action. Then draft a one-line journal note per position with rule ids met and a blank decision: field. End with: Review complete. Decisions are yours. Do not write the journal until I confirm this draft.

Guardrails

• Reports only; no order access.
• Only your written rules are checked.
• Your decision is logged, so rule-following can be measured.
• Rules must be written as simple, unambiguous conditional statements with explicit numeric thresholds (e.g. "close < initial_stop" or "days_held > 10").
• You must spot-check every agent report against the raw CSV numbers before logging any decision; treat the agent output as a first draft only.

What can go wrong

Plain-English rules such as 'trail it if it looks extended' cannot be checked. When those rows were missing, the model filled in a guessed met/not-met. Keep only threshold rules the script can compute, and treat missing as unchecked, not as not-met.

Source and references