I Built a Trading Journal in Obsidian That Claude Analyses Every Week. Here Is What It Found.

I started the journal in January with one question: am I actually as disciplined as I think I am?

Six months and 34 trades later, Claude ran the first cross-session analysis. I fed every journal entry into a single prompt, six months of entries, notes, conviction scores, thesis statements, exits, and the post-trade reasoning I wrote in the days after.

The finding that stopped me: in 71% of my losing trades, there was a note already in my Obsidian vault that contradicted the entry thesis. Captured before I entered the position. In my own words. Timestamped before the trade.

I had the counterevidence. I did not check it before trading.

That is not a market insight. That is a behavioral pattern. And it would have been invisible without something that could read across six months of sessions simultaneously and ask whether the person who entered those trades was reading their own research before acting on it.

Why Obsidian Instead of a Spreadsheet

The standard trading journal is a spreadsheet. Entry price, exit price, P/L, maybe a notes column. The spreadsheet is good at arithmetic. It is bad at everything else.

What I needed the journal to do was not calculate returns. I already know my returns. What I needed was the ability to run a natural language analysis across the behavioral and qualitative layer: why I entered, what I believed at the time, how that belief compared to what I had previously captured, and what pattern was sitting across all of that that I could not see from inside any individual trade.

A spreadsheet has no context window. Claude does.

Plain markdown files in Obsidian are readable by Claude directly. Every journal entry is a note. Every note has a date, an asset, a thesis, a conviction score, a position size, and a post-trade reflection. Six months of those notes, pasted into a single prompt, gives Claude enough to find the patterns that survive across sessions.

The patterns that survive across sessions are the ones that matter. A single bad trade is noise. The same mistake appearing in 71% of losing trades across six months is a system failure.

The Journal Structure

Each trade gets its own note in a 03-Projects/trading-journal folder. The note has seven fields and nothing else.

Trade Note Template

[TICKER] [LONG/SHORT] [DATE]Entry price: Position size (% of portfolio, not dollar amount): Conviction at entry (1-10):Thesis at entry: [One paragraph. What I believe and why. What would have to be wrong for this trade to fail.]Exit price: Exit date: Why I exited: [Did the thesis change, or did I exit on price action? If thesis changed, what changed it.]Post-trade reflection: [Written 48 to 72 hours after exit. Not during the emotion of the trade. What was right, what was wrong, what I would do differently.]Vault notes that were relevant at entry: [Any notes in 01-Sources or 02-Ideas that touched this asset or thesis at the time of entry. Listed by file name.]Last published content about this asset: [date and title, or None]

Position size must be recorded as a percentage of total portfolio, not a dollar amount. Comparing a $500 altcoin position to a $2,000 NVDA position in the same analysis is meaningless. The conviction-sizing inversion finding only became visible because every position was expressed in the same unit. Use percentage throughout.

If you trade both crypto and stocks, add these asset-specific fields after the main template fields.

For crypto positions:

Chain or protocol: DEX or CEX: Liquidity note (was exit size constrained by available liquidity?):

For stock positions:

Sector: Days to next earnings at entry: Stop loss pre-set at entry (yes/no, and at what level):

One template covers both as long as the irrelevant fields are left blank. The analysis prompt handles mixed entries correctly because it reads content, not structure.

The last field is the one that required discipline to complete honestly. It means going back to the vault before entry and listing what was already there. The 71% finding emerged from comparing what I listed in that field against the trade outcomes. When I listed a vault note that contradicted the entry thesis and entered anyway, the trade had a significantly higher failure rate than trades where the vault notes supported the thesis.

The journal documented my own confirmation bias in real time. I just had not read the data across all the entries at once until Claude did it for me.

The Weekly Analysis Prompt

Every Sunday evening, I run the week’s entries through Claude alongside the full journal history.

Weekly Journal Analysis

I am going to give you my complete trading journal as a set of markdown notes. Each note covers one trade: the entry thesis, conviction score, position size, exit reasoning, and post-trade reflection.Read every note before producing any output. Do not summarise individual trades. Look for patterns across the full set.Produce four sections:1. Behavioral patterns: what consistent behaviors appear across my entries that I have not explicitly named? Look specifically at the relationship between conviction scores and position sizes, the gap between stated exit rules and actual exit behavior, and any pattern in the timing of entries relative to external events.2. Accuracy of thesis: in trades where I documented the entry thesis clearly, how often did the thesis play out versus fail, and is there a pattern in which types of thesis statements proved more or less reliable?3. Vault usage: in entries where I listed vault notes at entry, did the presence of contradicting vault notes correlate with worse outcomes? Was there a pattern in whether I checked the vault before entering?4. One honest observation: the single most uncomfortable pattern in this dataset that I have probably been avoiding naming.Do not validate. Do not encourage. Find what the data actually shows.

The fourth section is the most important one. Claude produces it once it has read everything else, and it is the section I read first.

How many entries before the analysis produces something real.

Ten trades is the minimum for the weekly prompt to return anything meaningful. Below that, Claude identifies individual trade characteristics rather than patterns. Patterns require repetition.

Twenty entries is where behavioral patterns start to emerge with enough consistency to act on. Thirty or more is where the uncomfortable ones appear, the ones that have survived long enough to become habits.

If you are starting from zero, run the prompt anyway from trade one. The early outputs will be thin. That is useful too. Knowing what the analysis cannot yet see is different from not running it at all. By the time you hit twenty entries, the pattern history from earlier prompts gives Claude something to track against.

What the Six-Month Analysis Found

I ran the full analysis in late June after accumulating enough data for the patterns to be statistically meaningful. 34 trades across AI and tech equities, BTC, ETH, and a small set of altcoin positions. The four findings in order of significance.

Finding 1: The counterevidence pattern.

71% of losing trades had a contradicting note in the vault captured before the entry date. The contradiction was not subtle in most cases. The vault notes specifically named the risk that materialized. In nine of the eleven worst-performing trades, the vault contained a source note or thesis note that would have been grounds to either size down significantly or not enter at all.

I had done the research that told me not to enter the trade. I entered it anyway. Not because I had not read the research. Because I had not read it immediately before entering. The research was in the vault. The impulse to trade was in the session. They never met.

Finding 2: The content publication effect.

In trades where I had published content making a bullish case for the asset in the previous 48 hours, entries underperformed by a significant margin versus trades with no recent content connection. Claude flagged eight instances where the entry date was within two days of publishing something bullish about the position.

The pattern: publishing a bullish take on an asset generates positive responses. The positive response increases apparent conviction. The increased apparent conviction drives a position add or a new entry. The add or entry is not based on new research. It is based on the social feedback loop of the content.

The journal documented this. I had not noticed it because each individual entry felt justified in the moment.

Finding 3: The conviction-sizing inversion.

My highest conviction entries (rated 8 to 10) were consistently smaller than my lower conviction entries (rated 5 to 7). The pattern held across asset classes.

The mechanism Claude identified from the post-trade reflections: high-conviction calls tended to be earlier-stage or less-liquid positions where sizing up felt psychologically riskier. Lower-conviction calls tended to be larger, more liquid assets where the size felt safe regardless of conviction. The result was that I sized my positions inversely to my stated research confidence.

Finding 4: Exit behavior versus exit rules.

My stated rule is to exit when the thesis changes. The data shows 78% of exits were triggered by price action, either a stop loss or a target being hit, not by a documented thesis change. In most cases, the post-trade reflection was written after the exit and contained the thesis update reasoning that should have preceded the decision.

I was documenting thesis changes retrospectively and calling it thesis-driven trading.

What Changed After Seeing the Data

Two behavioral changes happened immediately.

The first: mandatory vault check before any entry. Not a scan. A deliberate search of 01-Sources and 02-Ideas for the asset and thesis before the trade note is opened. If a contradicting note exists, it gets listed in the vault notes field and read in full before sizing is decided. The counterevidence pattern was only possible because the vault check was optional. Making it mandatory removed the mechanism that produced 71% of losing trades.

The second: a 72-hour cooling period after publishing any bullish content about a position. The period is long enough to let the social feedback loop decay before entering or adding. Three instances have been caught since implementing it.

The enforcement mechanism is built into the template. The last field, “Last published content about this asset,” forces acknowledgment at entry time. If the date is within 72 hours, the field itself is the flag. The act of filling it in and seeing a recent date requires a deliberate decision to proceed rather than an unconscious one.

The same pattern is documented in CLAUDE.md under Hard Rules: “If I published bullish content about an asset in the last 72 hours, any entry or add to that position requires a second vault check and a one-sentence justification written in the trade note before the position is sized.” The rule is in the file that opens every Claude session. It does not prevent the trade. It requires the acknowledgment to be explicit rather than convenient.

The conviction-sizing inversion and the exit behavior finding are both ongoing. They are in the Open Contradiction section of my research methodology note. The data is clear. The behavioral change is harder.

How to Build This

Step 1: Create the journal folder

In Obsidian, create 03-Projects/trading-journal Every trade gets one note in this folder Name the note with the format: YYYY-MM-DD-TICKER-LONG or YYYY-MM-DD-TICKER-SHORT

Step 2: Use the template for every entry

Copy the template above into every new trade note The vault notes field is mandatory, not optional Complete the post-trade reflection 48 to 72 hours after exit, not immediately

Step 3: Get the notes into Claude

Two methods depending on how many entries you have accumulated.

Under 20 entries: open a new Claude conversation, select all notes in the 03-Projects/trading-journal folder, copy the full content, paste it directly above the weekly analysis prompt. A full month of detailed trade notes runs roughly 5,000 to 10,000 tokens, well within Claude’s context window in a single paste.

Over 20 entries: create a Claude Project at claude.ai and upload the journal folder as project knowledge. Every weekly session inside the Project starts with Claude already having the full journal history. You only need to paste the new entries from that week and run the prompt. The Project accumulates the history automatically.

Either way, run the prompt once per week, not daily. Daily is too little data between runs. The weekly cadence gives you enough new entries to see whether the patterns from last week are improving, stable, or worsening.

Do not read the validation sections first. Read section four first.

Step 4: Build the monthly pattern note

After each weekly analysis, create a note in 02-Ideas/trading-patterns.md Record any new pattern the analysis surfaced This note becomes the input for the monthly behavioral review

The monthly behavioral review prompt:

Read my trading-patterns.md note and all journal entries from the last 30 days. For each behavioral pattern I have documented: is it getting better, staying the same, or getting worse based on the last month’s data? Be specific. Reference entries.

The value compounds with time. Six weeks of data produces surface-level patterns. Six months produces the ones that are uncomfortable to name.

What to Do When the Analysis Flags a Live Position

The weekly analysis is built around closed trades. But occasionally it will surface a pattern that clearly matches something still open.

The counterevidence pattern is the most common version of this. Claude identifies that 71% of your losing trades had contradicting vault notes at entry. You have two open positions. You did not run a vault check before entering either of them.

Do not exit immediately based on pattern recognition alone. A pattern is a statistical tendency, not a certainty about any individual position.

Do this instead:

Run the vault check now, not at entry. Search 01-Sources and 02-Ideas for anything touching the asset and thesis. If you find a contradicting note, add a flag to the open trade note: “Pattern match flagged by weekly analysis [date]. Contradicting note found: [filename].”

Then schedule a deliberate review within 24 hours. Re-read the original entry thesis. Re-read the contradicting note in full. Decide explicitly: does the contradiction invalidate the thesis, or is there a reason the exception applies here? Write the answer in the trade note.

If the contradiction invalidates the thesis, that is a thesis change, not a pattern panic. Exit on the thesis change, not on the pattern flag. The distinction matters because it keeps the exit logic clean for the next month of analysis.

If the contradiction does not invalidate the thesis, document why not and leave the position unchanged. The pattern match becomes a data point in the next weekly analysis rather than a reason to act.

A trading journal in Obsidian is not a trade management system. It does not execute anything. It does not replace the judgment calls of when to enter or exit.

It is a behavioral audit trail that runs on your own data, in your own words, and surfaces the gap between what you say your process is and what your process actually was. That gap is where most losses live. Not in bad market reads. In the repeated behavioral patterns that good research should have prevented and did not.

The journal found that I had the research. I just was not reading it at the right moment.

That is a fixable problem. A spreadsheet cannot find it. Claude and six months of your own notes can.

This article is personal experience and financial commentary, not financial advice. Do your own research before making any investment decision.

Follow @damidefi on X for daily Claude AI tools, crypto analysis, and the full journey to 100K. Bookmark this. Share it with one person who tracks their trades but has never analysed the behavioral layer underneath them.