Tried it the lazy way once — pasted a messy day and asked for "summary with PnL." Model invented a win rate that would've made me proud if it were true. It wasn't. CSV was sitting right there. Never again without a ledger join.
### Safe workflow
1. Export fills from broker CSV. Source of truth.
2. You write the tags: setup, mistake codes, emotion codes.
3. LLM summarizes **text fields only**. Numeric PnL comes from CSV joins. Not from the model's imagination.
### Prompt guardrails
Code: Select all
Summarize the trader notes below.
You MUST NOT invent PnL, win rate, or trade counts.
If a number is not present, write "NOT_IN_NOTES".
Return:
- Recurring mistake codes
- Setups attempted
- Process breaches (risk rules)
- Questions for Sunday review
Pasting screenshots and asking "how much did I make?" Letting the model invent psychological deep reads you then treat as fact. Weekly summary without the CSV attached — that's a campfire story.
Use the model to compress narrative. Keep the ledger boring and external.
Do you join CSV → notes in a sheet, or a small script?
Which mistake codes does the model actually surface that you were blind to?
Anyone tried screenshot→LLM for journals and regretted the numbers?
Local vs cloud for journal text — where do you draw the line?