The Setup

I've been writing quarterly reflection journals since 2018. Raw, timestamped entries where I track what I'm working on, how I feel about it, what I'm planning next. Career decisions, indie project attempts, emotional states, financial analysis. 25 files. 2.1 megabytes of honest self-talk.

I'd never read them all together. Who would? They're 70–100 KB each, written in a mix of Swedish and English, spanning eight years. I'd review the last quarter's file at the start of a new one, maybe glance at the year-ago entry. But reading the full arc — the longitudinal view — was never practical.

I asked Claude Code to analyze all 25 files for recurring patterns across the full period.

Short answer: It's genuinely useful — not life-changing, but practical in a way that makes it a no-brainer to keep doing. Here's what happened.

What I Was Working With

The journals use a personal time-tracking format I've maintained in Obsidian since 2018. Each quarter gets its own markdown file with YAML frontmatter and timestamped entries — essentially a work diary mixed with life coordination notes.

• 25 quarterly files covering Q1 2018 through Q1 2026
• ~2.1 MB of text — a mix of Swedish and English
• Unstructured entries — some are a single line about a meeting, others are multi-paragraph reflections on career direction, family dynamics, or financial anxiety
• No consistent format — early files are sparse; later ones are dense with detailed timestamps

The material was already digital and in plain text, which made it immediately accessible to AI. No scanning, no OCR, no transcription step. Just markdown files on disk.

What AI Is Genuinely Good At: 7/10

Holding the full arc in memory. This is the main capability that matters. I can remember roughly what I was thinking last quarter, maybe last year. The AI read eight years simultaneously and returned patterns I literally could not see:

• A specific career cycle (build → almost ship → take consulting contract → abandon indie project) repeating three times with nearly identical emotional arcs
• The same strategic insight ("distribution is the bottleneck") articulated 11 times across 7 years in almost identical language — without me ever noticing I was repeating myself
• An emotional sine wave (anxious planning → productive burst → hopeless trough → pragmatic acceptance) that follows the same shape every year, with troughs correlating to financial anxiety

Cross-referencing across time. The AI connected a journal entry from Q3 2022 to one from Q1 2024 and another from Q4 2025 — all expressing the same realization about content marketing being essential, each time as if it were a fresh insight. Seeing "you've said this before, here's when" with exact dates is qualitatively different from vaguely knowing you have a pattern.

Quantifying intuitions. I knew I had a tendency to buy domains instead of shipping products. The AI found 100+ domain registrations across the corpus and presented it as a named pattern: "domain hoarding as procrastination." The difference between "I think I do this" and "you did this 100+ times, here are the names and dates" is the difference between a hunch and evidence.

Parallel processing at scale. Claude Code dispatched four parallel agents — one per era (2018–2020, 2021–2022, 2023–2024, 2025) — plus a fifth that cross-referenced my career strategy documents. Each agent processed hundreds of pages and returned structured findings. The full analysis took about 10 minutes. Reading those 25 files manually would take days.

Being unflinching. I can rationalize my own behavior. The AI just reported what it found. No softening, no "but you've also grown a lot!" — just patterns, dates, frequencies, and quotes.

What AI Is Bad At: 5/10

Interpreting emotional context. The AI can identify that I wrote "det känns hopplöst kring indiehacking" ("it feels hopeless about indie hacking") in May 2023, but it can't fully grasp what that meant in context — the exhaustion, the family situation, the financial pressure. It reports the data point accurately but misses the weight.

Distinguishing signal from noise. Some "patterns" the AI surfaced were just repeated mentions of the same project across quarters, not meaningful recurrences. The human work is deciding which patterns matter and which are just artifacts of how I write.

Actionable recommendations. The AI can tell me I've identified distribution as a bottleneck 11 times without acting on it. It can't tell me why I don't act on it or what would actually work this time. The gap between pattern recognition and behavior change is entirely human territory.

Language mixing. My journals shift between Swedish and English mid-sentence. The AI handled this well for pattern detection but occasionally misinterpreted Swedish idioms or lost nuance in translation when summarizing.

The Result

A structured report identifying seven longitudinal patterns across my life and career, each with specific dates, quotes, and frequencies. Patterns I'd been vaguely aware of were now named, quantified, and undeniable.

The most significant finding: an insight I thought was fresh ("I need to focus on distribution") had been restated in nearly identical words every year since 2020. The AI didn't just find the pattern — it showed me I'd been in a loop without knowing it was a loop.

Verdict: Would I Do It Again?

Yes — this is now a quarterly skill. I built a /reflection command that will re-analyze the journals each quarter and check whether the identified patterns are still active.

Could I have done this without AI? I could have re-read my journals manually. But 2.1 MB across 25 files in two languages — realistically, I never would have. The longitudinal view only exists because AI made it effortless.

I wrote about the personal side of what the AI found — the loop, the gravity well, the domain graveyard — in a separate post: 8 Years of Building, Hoping They Will Come.

The honest takeaway: If you have years of reflection logs sitting around, asking AI to find patterns across them is an obvious thing to do. It's not magic — I was vaguely aware of most of the patterns it found. The value is in precision: there's a difference between "I think I have this tendency" and seeing it dated, quantified, and quoted back at you across seven years. That's useful. It's not the most impressive thing I've used AI for (that would be biography writing from voice memos — a separate post), but it's a solid, repeatable use case.

The Toolchain

• Obsidian — where the quarterly journals live as markdown files with YAML frontmatter
• Claude Code — dispatches parallel agents to read the full corpus, cross-reference patterns, and generate structured findings
• Remember This — captures voice memos and daily context that feed into the quarterly journals
• Plain text — the entire system works because everything is markdown files on disk, immediately accessible to AI without any export or conversion step