An AI Trading Journal — Reviewing Your Own Decisions Without Flattery

A language model produces what plausibly follows the text in front of it. If your input contains the sentence "the setup was valid but the market went against me", then everything that follows is generated in a context where the setup was valid. The premise has been established by you and the model has no independent access to the chart, the order book or your account. It is completing your document, not auditing it.

Layered on top is the training. Reinforcement from human feedback rewards answers people rate well, and people rate validation well — particularly on emotionally loaded material. The tuning that makes these tools pleasant to use is the same tuning that makes them useless as a critic of your own work. Both properties come from the same place.

This produces a closed loop that feels like reflection. You write a rationalisation, the model reflects it back with better vocabulary and a framework attached, and you read your own excuse in an authoritative register and update your beliefs towards it. The excuse has been laundered. You now hold it more firmly than before, and the process felt like rigour.

The tell is emotional rather than analytical. If reading a trade review leaves you feeling better about a loss, be suspicious. A genuine review of a genuinely bad decision is uncomfortable. That discomfort is the signal that something outside your own framing was actually applied.

The review is only as good as what goes into it, and most trading journals are written in a way that makes an honest review structurally impossible. They are narratives. "Took a breakout, it failed, choppy market." That is a conclusion with a few facts embedded in it, and any model reading it will analyse the conclusion.

A reviewable entry is a record, written before you know the outcome or at least written as though you do not. Instrument and segment. Date and time of entry. Price paid. Position size in rupees, and that size as a percentage of your capital. The stop, written as a price. The target, written as a price. The reason for entry, stated as an observable condition rather than an opinion — "closed above the level that had capped it for eleven candles, on volume higher than the prior five sessions" rather than "looked strong".

Then, separately and clearly fenced off, the outcome: exit price, exit time, exit reason, realised result in rupees. And separately again, in its own section that the review prompt will be told to ignore, whatever you felt about it. The feelings section has real value for you. It has no place in the data the model reasons over.

The discipline that makes all of this work is writing the entry before the exit. A rationale recorded after you know the result is not a rationale, it is a reconstruction, and it will be the single most contaminated piece of text in your journal. If you only have post-hoc entries, mark them as such — they still support review, but you must know that the reasoning field is unreliable.

The single most effective intervention is also the simplest, and almost nobody does it: do not tell the model what you think happened. Give it the fields and nothing else. No verdict, no "I think I exited too early", no "this was a good trade that went wrong". The moment a conclusion is present in the input, everything downstream is generated inside that frame.

The second intervention is nearly as strong: withhold the outcome for the first pass. Give it entry price, size, stop, target and stated rationale, and ask it to evaluate the decision as it stood at the moment of entry — before revealing what happened. A model that knows the trade lost will construct an explanation for why the decision was flawed. A model that knows the trade won will construct an explanation for why it was sound. Both explanations will be fluent, and both are contaminated by hindsight in exactly the way your own memory is.

Then run a second, separate pass with the outcome included, and compare. Where the two assessments differ is the interesting territory — it isolates how much of your own after-the-fact reasoning is driven by the result rather than by the decision. That gap, tracked across thirty trades, is a more useful thing to know about yourself than any individual review.

The third intervention is anonymisation at scale. When you are reviewing a batch, strip the instrument names. "NSE large-cap, ten trades" rather than a list of tickers removes the model’s ability to lean on whatever it absorbed about those companies, and it removes your ability to read the output through your own attachment to a name.

Once your input is clean, the second half of the work is output structure. The reason "be critical" fails is that agreement is still an available shape for the answer — the model can produce a critical-sounding paragraph whose content is "this was fine". The fix is to demand an output structure in which agreement is not one of the shapes.

Assign a role that has a job, not a temperament. "You are reviewing this record on behalf of someone who must decide whether this process is repeatable" gives the model a task with a criterion. "Be harsh" gives it an adjective. The first changes what gets produced; the second changes only how it sounds.

Force the strongest case against. Ask explicitly for the best argument that this decision was poor, stated as strongly as an intelligent opponent would state it — and then, separately, the best argument that it was sound. Two mandatory sides means the critical case has to be constructed regardless of which way the model would have leaned. Requiring it to state which case is stronger and why turns a symmetric exercise into a judgement.

Then close the escape hatches. Models reach for a small set of comforting phrases when asked to judge a trade: markets are probabilistic, no process wins every time, risk management is what matters. Each is true and each is a way of avoiding the question. Ban them by name in the prompt. And ban the two moves that matter most: never say the decision was reasonable without naming the specific field in the record that supports it, and never say a loss was bad luck without first ruling out every process explanation you can see.

Finally, add the question most reviews never ask: what is missing from this record? A model that can only see six fields will happily analyse six fields. Asking what a seventh field would have revealed is often the most valuable line in the output, because gaps in the journal are the reason most reviews are shallow.

Single-trade reviews are where people start and they are the less valuable half. One trade contains almost no information about your process; it is one sample from something with a great deal of randomness in it. The patterns are in the batch, and a batch is where a model is genuinely useful, because reading thirty of your own entries carefully is tedious in exactly the way that causes people to skim.

So take a quarter’s worth of entries, strip the instrument names, and ask for pattern-level observations. Not "which trades were good" — that question invites scoring, and a score on your own trades is meaningless. Ask instead: which of my recorded reasons recur, and are they conditions or opinions? Where does position size vary, and does the variation correspond to anything I wrote down? Are my stops honoured more often in some circumstances than others? What did I write in the entries where the plan was abandoned?

Those are answerable from your own text and they are descriptive. They tell you about the consistency of your process — which is the only thing a journal can honestly measure. They say nothing about whether your approach is any good, and no amount of AI review can tell you that, because thirty trades is not a sample from which anyone can draw that conclusion.

A note on the arithmetic. Do not ask a model to compute your totals, your average result, or any statistic across the batch. It will produce numbers and some of them will be wrong, and wrong figures about your own trading are worse than no figures. Compute those yourself in a spreadsheet and paste the results in if the review needs them. The model reads language; the arithmetic is yours.

A note on privacy. Your trading journal is unusually sensitive — it reveals positions, capital, risk appetite and behaviour. Before any of it goes into a third-party tool, know what that tool does with your input and whether it is retained or used for training. Strip account numbers, broker identifiers and anything that identifies you personally. Position size as a percentage of capital carries all the analytical information that rupee capital does, without disclosing your capital.

A journal review reads your records. That is its entire scope, and the boundary is worth stating plainly because the tool is fluent enough to sound like it is doing more.

It cannot tell you whether your approach works. That is a question about a sample, a market regime and a time horizon, and thirty records from one quarter cannot answer it — neither can a model reading them. Any output that drifts towards "your strategy appears effective" has left the evidence behind and should be discarded along with whatever else that pass produced.

It cannot see what you did not write. A review is a function of your journal, and a thin journal produces a thin review dressed in confident prose. The "what is missing from this record" section exists precisely because that gap is invisible from inside the output. If your entries do not record what you passed on, no review can tell you anything about your selection.

It cannot tell you what to do next. No output from any of these prompts is a recommendation, and prompts that invite one — "should I take this setup again?" — should not be run at all. The prompts above are deliberately written to refuse that question. Your decisions, your risk and your capital remain entirely yours; nothing here is investment advice.

And it cannot be trusted with your arithmetic. Every figure in a trading review must come from your own spreadsheet or your broker statement. A model asked to total a column will produce a total, and you will have no way of knowing whether it is right without doing the sum yourself — at which point the model added nothing except a risk.

What it can do is real and worth having. It reads thirty of your own entries without getting bored, it notices that the same adjective appears in nineteen of them, and it constructs the case against a decision you were not going to construct on your own. That is a genuine addition to a review process. It is not the process.