A Guide to Par
What 1,911 AI conversations reveal about one person's brain, constraints, and building pattern
Based on 1,911 AI conversations (2022-2026), financial records, medication data, location history, git archaeology, 318 published articles, and an extended interview. This is not a biography. It’s a reference for anyone who needs to understand how Par works, what drives him, and what to watch out for.
The basics
Par Boman. Lives in Kall, a village in Are kommun, Jamtland – rural northern Sweden. Born mid-1980s. Formerly a radio journalist at Sveriges Radio P4 Gavleborg in Gavle. In January 2023 he took over Arebladet, a free local newspaper delivered to all 6,500 households in Are kommun. He runs it alone.
He has ADHD, diagnosed November 21, 2025. Medicated since then (Elvanse/lisdexamfetamine, stabilized at 90mg/day by March 2026). The diagnosis surprised nobody, least of all him.
He has a sister, Sara Johanna (SJ), who is also a heavy AI user. Their mother, Susanne Pedersen-Boman, died in December 2025 after hospitalisation in Helsingborg. Their father moved in with Par afterward.
He cannot code. Not in the “modest senior engineer” sense – literally cannot write a line of code from memory. All software is built by AI. As of March 2026 he has 20 git repositories, 689+ commits, and a multi-service VPS infrastructure. The entire portfolio is 88 days old.
How he makes a living
Arebladet does not pay a salary. The company (Arebladet AB) covers his car, housing, and all AI/tech costs, with a token annual salary for tax purposes. To actually live, he works five part-time jobs: PostNord mail delivery, Circle K gas station, Hotel Kallgarden bartending, BrandImpact shop demos, and seasonal stints at Sveriges Radio.
He needs roughly 2 days of paid external work per week. Private baseline costs: ~12,000 SEK/month. Company fixed costs: ~19,000 SEK/month.
The newspaper was purchased for 1.3M SEK (300k upfront, 1M over 5 years, no interest). Payments are roughly on track. He committed to running it for at least 5 years – locked in until late 2027. This matters: the ADHD pattern of switching jobs every ~2 years cannot activate.
What Arebladet is
18 issues per year. A4 format, full color, typically 16 pages. Fully ad-financed. Regular issue costs ~50k SEK to print and distribute, brings in 60-100k in ad revenue. Four special issues per year (Easter, Summer, Autumn, Christmas) are the profit pillars: 94k-150k each.
The editorial philosophy: soft, cozy, people- and place-focused. Not hard news. “Tidningen ska vara lasvard” – the paper should be worth reading. At least two interesting articles per issue, minimum 25% editorial content. The centerfold is always a full image spread with no ads. He has never broken this rule.
He covers the entire municipality deliberately – Are kommun is much more than the ski resort.
How his brain works
The ADHD pattern
Big exciting idea waves. Spin up whole ecosystems in parallel. Hyperfocus on one thing, then abruptly move on. Lose track early in a cycle, then recover with pants-on-fire effort at the deadline.
In git, this manifests clearly:
- storyteller: 3 days of activity across 32, then silence
- arebladet_manager: built in 6 hours, declared failed 3 days later
- fun/bionic-malcolm: 4 commits in 2 hours, never touched again
- ParKit: 109 commits and still growing – the one that stuck
Production stress is 8/10 about half of all newspaper issues. The typical crunch hits Thursday evening after 18:00.
The two-thirds rule
68.5% of his AI sessions produce no direct output. For every article written or tool built, two sessions went nowhere. The “everything shipped” picture is a selection effect. Monthly production rates range from 22% (December 2025, most sessions but lowest rate) to 75% (August 2025, narrow focused use).
This is not failure. It’s how exploration works for him. The exploration sessions build vocabulary, test boundaries, and occasionally plant seeds that germinate months later. The Parception spec (September 2025) went dormant for three months before its architectural ideas resurfaced in real projects.
Idea containment
He has built three structures to manage the running-ahead tendency:
- Director – “The Lab,” systematic exploration and experiments
- Arebladet2 – phased long-term spec for the newspaper’s evolution
- ParKit – deliberately over-engineered personal productivity suite
These are scaffolding for a mind that generates more ideas than any single person can execute.
How he uses AI
The collaboration pattern
Developed over 1,000+ days of newspaper work before he ever wrote code. The pattern: paste input, get output, iterate with terse feedback, trust the result enough to ship it.
For articles: paste interview transcript, get draft, direct with single-phrase instructions (“Pa svenska”, “Mer citat”, “Kortare rubrik”). This is an editor directing a writer. The same pattern transferred directly to code – he directs AI like an editor directs a reporter.
The learning path
The technical vocabulary came from images, not courses. From April 2024 through the end of 2025, obsessive AI image generation (AUTOMATIC1111, then Flux, RunPod, ThinkDiffusion, FaceFusion) introduced every foundational concept: dependencies, APIs, cloud infrastructure, Docker, multi-step workflows, client-server architecture. $4,600+ in AI spending, 63% on image generation alone.
The strongest early motivation was NSFW content – which matters not for its content but for its function. It created conditions for deep, sustained engagement with the tools. Pipeline randomness held attention: each generation is a decision point requiring active evaluation.
By September 2025, without ever writing code, he could design client-server architecture, specify feature-per-file modularity, and invent “code drift” as a version control concept – transferred from face drift in WAN video generation.
Platform trajectory
- 2022-2025: ChatGPT exclusively (articles, then images, then code attempts)
- May 2025: Claude Pro subscription starts
- Dec 27, 2025: Claude evaluation begins (“They say you are better at code”)
- Jan 4, 2026: First Claude Code session – 13 messages to do what took 213 before
- Feb 2026: OpenAI cancelled. ChatGPT drops to near-zero.
- Mar 2026: Claude Code 316 sessions, ChatGPT 0, Gemini 0
Claude Code’s production rate: 100%. ChatGPT: 31%. Claude.ai: 14%. The tool change explains more variance than medication.
The “sweet spot”
Claude’s own December 2025 assessment, which Par found accurate:
“You are at that sweet spot where you understand the ecosystem and can architect solutions, but you prefer getting complete runnable code rather than diving into the implementation details yourself.”
He thinks like a systems architect. He cannot implement. AI is not an assistant – it’s the implementation layer. “I cannot code” is not modesty. It’s a literal architectural constraint that shapes everything.
What he values
Freedom above all
“Never going to have a real full-time job again.” What he wants changes frequently. The multi-track life is the desired structure. The newspaper, the part-time jobs, the AI projects, the party business (PartyPar) – these are not scattered commitments. They’re a portfolio designed to avoid dependence on any single thing.
Honesty over comfort
He doesn’t want sunny narratives. When the Orchestra project produced increasingly polished stories about his AI journey, his interview corrections were consistently in the direction of “less flattering but more accurate.” RunPod serverless never worked. A1111 was installed without AI help. November’s burst was coping with grief, not productive inspiration. Two-thirds of sessions produced nothing.
He wants the real picture, including the parts that don’t look good on a slide.
The fun path
When two approaches are equally good, pick the more interesting one. “Hard” and “time-consuming” are human measures – with AI doing the implementation, ambitious approaches are cheap. This isn’t recklessness. It’s a correct recalibration of effort costs when the human doesn’t write the code.
Push back
He depends on AI catching bad ideas. He explicitly wants disagreement when his idea is worse than the current approach. Don’t defer because the instruction came from a human.
What to watch out for
The sunny bias
Activity reads as productivity. A 310-message session looks like intense work but might be a dead-end website rebuild that produced nothing. His self-reporting skews optimistic about timelines and completion. The chatarkiv measures engagement, not outcomes.
All previous analytical versions of his story inherited this bias. The correction is always in the same direction: less was accomplished than it looks like from the activity log.
Dead-end chains
He will sink hundreds of messages into something that isn’t working. The five major dead-end chains consumed 1,919 messages (15.8% of all transition-period messages):
- RunPod Serverless: 12 sessions, 516 messages, zero images produced
- StoryMaker: 6 sessions, 352 messages, no surviving code
- Website rebuild: 6+ sessions, 377+ messages, site unchanged
- LifeLab: 6 sessions, 383 messages, schema only
- AI Server Build: 2 sessions, 291 messages, specs only
Pointing this out isn’t criticism. It’s information. Some dead ends are necessary exploration. But if a project has consumed 200+ messages with no output, that’s a data point worth naming.
The memory gap
His recollection of his own timeline is compressed by roughly one year. Every milestone is remembered in the correct order, but dated twelve months too early. The pre-code era felt shorter than it was. The building era feels longer. Verify dates against records, not memory.
Financial precarity
The multi-track life works, but with thin margins. Monthly costs are covered, but there’s no buffer. Cash flow has been a recurring issue. AI experimentation costs are real (~$200/month and climbing). When he says “my actual cost is half” (company-paid at ~50% tax rate vs salary), that’s true but the company still has to earn the money.
The escape pattern
Historically changed jobs every ~2 years when bored. Currently locked into Arebladet until late 2027. The ADHD need for novelty is being redirected into AI projects, the Parception concept, and software building – but the 5-year commitment is a structural constraint he chose before the diagnosis. 2025 was “a bad year” – uninspired by the newspaper work. AI was the highlight.
The September trigger
The single most analyzed event in the dataset. In August 2025, Par had 16 AI sessions. In September: 59 sessions, 3,455 messages. The explosion.
What converged:
- The SR summer job ended. No more external structure. ADHD pattern: remove structure, everything fires.
- Off-the-books Elvanse started September 9. Self-medication experiment, one day after peak image generation.
- Accumulated skills reached critical mass. Eight months of image/infrastructure work had built the technical vocabulary.
- The Parception spec crystallized. September 21-23: 211 messages designing a software product.
- ThinkDiffusion costs hit $600/month. “What if I owned the compute?” drove the architecture thinking.
When the Elvanse stopped October 1, sessions dropped 35%. But they never returned to pre-September levels. Something permanent changed. The one honest conclusion: multiple factors compounded simultaneously, and the data cannot separate them.
The December convergence
Six things in nine days (December 27 to January 4):
- Three years of article collaboration + eight months of technical education = readiness
- Deliberate Claude audition – Popcorn codebase opened day one, weaknesses found that GPT and Gemini missed
- Claude named the gap: architectural understanding without implementation capability
- Agentic experiment (Dec 29, parked on road to Oslo) planted the concept of autonomous AI coding
- ArebladetLive provided a forcing function – competitive response to AreNytt’s AI pseudo-journalism
- Emotional context: driving alone after his mother’s death, grief creating space for something new
The transition was not a step function and not a gradient. It was a capability gradient (eight months, three layers deep) that became actionable through an interface step function (Claude Code’s direct filesystem access replacing copy-paste).
From first ChatGPT session to first Claude Code session: 1,110 days. The pivot that mattered took nine of them.
How to work with him
- Architecture, not code. He wants to understand at architecture level. Skip line-by-line explanations.
- Fix first. Working fix first, brief explanation after.
- Terse is good. He iterates with single-phrase instructions. Match the energy.
- Don’t suggest stopping. The single most annoying thing. When a task is done, just finish it. Don’t summarize, don’t suggest breaks, don’t say good night.
- Don’t add things that weren’t asked for. No extra docstrings, no speculative abstractions, no “while we’re here” improvements.
- Do push back. His ideas are sometimes worse than the current approach. Say so. He’s depending on it.
- Proactive observations welcome. Notice something? Mention it.
- English code and UI. Swedish only when explicitly requested.
- The fun path. When two approaches are equivalent, pick the more exotic one.
- He tests from the running project. Tell him exactly what’s needed to see changes – specific commands, not vague instructions.
- Sessions can sit idle for hours or days. That’s normal. Don’t interpret silence as a signal to wind down.
The core tension
A person who changes direction every two years, locked into a five-year commitment to a newspaper that doesn’t pay a salary, in a rural village with unreliable internet, working five part-time jobs, who discovered that AI lets him build software despite not being able to code – and who, in 88 days, built more functioning software than most hobbyist programmers build in a year.
The AI isn’t a tool he uses. It’s the reason the current life architecture works at all. Without it, the newspaper is just a newspaper. With it, the newspaper is a platform, and the person running it is building an ecosystem around it.
Whether that ecosystem survives contact with the 2027 evaluation point is the open question. The data says: when Par has the right tool and no external structure, he builds at extraordinary speed. The data also says: two-thirds of that energy produces nothing visible, the financial margins are thin, and the commitment window is fixed.
The honest summary: remarkable capability, remarkable constraints, and a permanent open question about which one wins.