Test whether your written setup definition is clear enough to apply

How it works

1. Pick one setup you trade and copy its written definition from your playbook word for word. If it isn't written down yet, write it first (see the playbook interview entry).
2. Collect 15 to 30 past examples from your own chart screenshots or journal: some you counted as the setup and some you looked at and rejected. Save your label for each in a separate file the agent doesn't see. Crop out or cover the security name and the price scale.
3. Give the agent the definition and the unlabelled examples. For each, it answers qualifies, doesn't qualify or can't tell, quotes the clause of your definition that decided it, and says in one line what it saw.
4. Compare its labels with yours, by hand or with a small script, and list the disagreements. The agent sees your labels only after it has answered.
5. Sort each disagreement: the agent misread the chart (common; ignore it), the definition is silent on something you actually decide by, or the definition says one thing and you do another. Only the last two lead to edits.
6. Edit the definition in your own words, log the change and date in your playbook, and rerun on a fresh set of examples next month.

Copy-paste prompt

You test whether my written setup definition can be applied by someone who only has the text. You do not give
trading advice, you do not say whether any example was a good trade, and you never suggest what to trade.

Definition: {{definition_text}}
Examples: {{examples}} (images or written descriptions, each with an id; names and price scales removed)

For each example, answer qualifies, doesn't qualify or can't tell. Quote the exact clause of my definition that
decided it, and describe in one line what you saw that matched or failed that clause. If my definition doesn't cover
something you needed in order to decide, write "definition silent on:" and name it.
Use only my definition. Do not use pattern names, indicators or rules from anywhere else.

Return a table: id, your label, deciding clause, what you saw, silent-on notes.

Guardrails

• The agent applies only your text; outside pattern definitions are not allowed.
• It never sees your labels before it answers, so it can't simply agree with you.
• Security names and price scales are removed from the examples.
• Models misread charts, so disagreements are prompts for your judgement, not corrections. The agent edits nothing; you change the definition yourself.

What can go wrong

Vision models misread charts: they miscount bars, confuse wicks and bodies, and describe volume they can't see. Treat a disagreement on a visual detail as probably the model's error, and look for the other kind, where the agent says your definition is silent on something. That's the useful output. The other trap is rewriting the definition until the agent agrees with every example. That only makes the text match your past labels, not your future decisions, so always retest on fresh examples.

Source and references