What is trading?

Trading is risky. Your capital is at risk.

CROSS-CUTTING: COURSE 1 | LESSON 3

Your AI trade-journal analyst

Learning objectives

Prepare journal data in a format an AI assistant can analyse reliably (and know what to leave out).
Run a structured journal-review session: patterns by setup, session, instrument and rule adherence.
Judge which AI observations are grounded in your data and which are decoration to discard.

The most under-used resource in retail trading

Ask any experienced trader what separated their good years from their bad ones and "I reviewed my trades honestly" appears with suspicious frequency. Yet most journals die within a month, and most that survive are never analysed — because staring at 60 rows of your own mistakes is tedious and mildly painful.

This is the single best use of an AI assistant in trading. Not because the AI knows markets — it doesn't — but because journal review is exactly the kind of work LLMs are built for: taking structured text you provide, grouping it, counting it, and asking uncomfortable questions about it. You bring verified data; the AI brings tireless pattern-spotting and zero ego-protection.

Note the phrasing: you bring the data. Your journal comes from your platform history and your own notes. The AI never supplies a single number — it only rearranges yours. That division of labour is what makes this use case safe.

Step 1 — format your journal for analysis

AI assistants handle tables pasted as plain text or CSV well. A workable minimum per trade:

date, instrument, direction, session, setup_name, planned_risk_%, actual_risk_%,

result_R, followed_plan (Y/N), note

2026-06-02, EURUSD, long, London, pullback-to-MA, 1.0, 1.0, +1.8R, Y, clean entry

2026-06-03, XAUUSD, short, NY, news-fade, 1.0, 2.3, -2.3R, N, doubled size after loss

2026-06-05, GBPUSD, long, London, pullback-to-MA, 1.0, 1.0, -1.0R, Y, stopped, no complaints

Three rules for the export:

Results in R (risk multiples), not money. "-2.3R" analyses identically whether your account is $500 or $50,000 — and it means you're not pasting account balances into a chat tool. Privacy and better analysis in one move.
No identifying data. No account numbers, no broker login, no name, no balance. The AI needs none of it (Lesson X1.6 formalises this).
Include the followed_plan column. It is the highest-value field in the whole journal, and most templates omit it.

Twenty trades is a workable minimum; 50+ starts producing patterns worth trusting. Below 20, tell the AI the sample is small and treat everything as hypothesis, not finding.

Step 2 — run the structured review

Don't ask "what do you think of my trades?" — that invites a polite essay. Brief it with the four-part pattern from Lesson X1.2:

Example prompt: "You are a trade-journal analyst. You never give buy/sell recommendations and never predict markets. Below are 42 of my trades in CSV. Tasks: (1) group performance by setup_name, session and instrument, showing trade count, win rate and average R for each group; (2) compare trades where followed_plan = Y against N; (3) list the three patterns most worth investigating, each citing the specific rows it rests on; (4) ask me five hard questions a performance coach would ask. If a column is missing data, say so — do not estimate. This is analysis of my past behaviour, not financial advice."

Condensed example response: "By setup: pullback-to-MA: 18 trades, 61% win, avg +0.42R. news-fade: 9 trades, 33% win, avg −0.51R. breakout: 15 trades, 47% win, avg +0.10R. Plan adherence: followed_plan=Y (31 trades): avg +0.31R. followed_plan=N (11 trades): avg −0.62R. All five of your worst losses (rows 7, 12, 19, 30, 38) are N-trades, and four of the five carry actual_risk_% above planned. Worth investigating: (1) news-fade loses money and 6 of its 9 trades are in NY session after a losing London trade — rows 7, 12, 19… (2) … Questions: What happens in the hour before your N-trades? Row 12's note says 'doubled size after loss' — how often is oversizing a response to a same-day loss? …"

Notice what makes this output trustworthy: every claim points at rows you can check. Win rates you can recount, row numbers you can open. And notice what the assistant did not do — it didn't tell you to stop trading news-fades or to go long anything. It surfaced patterns and asked questions. Whether "news-fade" is a bad setup or just a bad sample is your judgment call.

Step 3 — separate grounded findings from decoration

AI journal reviews reliably produce two kinds of sentences, and you must learn to tell them apart:

Grounded: "Your 11 non-plan trades average −0.62R across rows 7, 12, …" — arithmetic on your data. Verify a couple by hand (spot-check the counting; LLMs do make arithmetic slips), then take it seriously.
Decoration: "This suggests you may struggle with discipline under volatile conditions" — a plausible narrative the model layered on top. It might be true. It is not in the data. Treat it as a question to explore, never a diagnosis.

A good closing message for any session: "List every claim you made that is directly computable from my rows, and every claim that is interpretation." Strong models handle this cleanly, and the split is often eye-opening.

Two more habits: run the review on a schedule (monthly, or every 20 trades) rather than after emotional days, when you'll unconsciously fish for comfort; and keep the AI's five questions in your journal — answering them in writing next month is where the actual improvement happens.

The reusable version

Everything above is packaged as the Trade-Journal Analyst skill in the skills library (skills/trade-journal-analyst.md): a complete system prompt with the input format, output sections, and safety rails pre-written. Copy it once into a project or custom-instructions slot, then each month just paste your new rows. The companion Post-Trade Review Coach skill does the same job for single trades on the day you close them.

Key takeaways

Journal analysis is the best trader use of AI: you supply verified data, the AI supplies grouping, counting and hard questions — never numbers of its own.
Export trades in R multiples with a followed_plan column, and strip all account and identity data before pasting.
Demand that every conclusion cites the specific rows it rests on, and spot-check the arithmetic.
Separate grounded claims (computable from your rows) from narrative decoration (plausible interpretation) — act on the first, merely explore the second.
The AI reviews your past; it must never be asked what to trade next.