Harness engineering is the layer above context engineering — you build the system (documentation, standards, quality checks, tool configs) that lets AI run unattended.

I built one over two days for my project book2skills and ended up with a fully automated book-to-skill publishing pipeline.

This post walks through the four steps.

Harness engineering is a term that’s been gaining traction recently. Like most emerging concepts, it’s ahead of widespread adoption — we’re still in the early days, and real-world examples are scarce. That’s exactly why I want to share how I’ve been applying it in my own product.

What Is Harness Engineering?

Earlier this year, OpenAI published a write-up describing how their team built a production app with over a million lines of code — without a single line written by human hands. The engineers weren’t writing code. They were building the system that allowed AI to write code reliably.

That system — the documentation, quality standards, constraints, and feedback loops — is the harness.

The word “harness” comes from horse tack: reins, saddle, bridle. A horse is powerful, but without that equipment it just goes wherever it wants. The AI is the horse. The harness is everything that channels its power in the right direction. You’re the rider building it.

To put it in AI evolution timeline, here’s how harness engineering fits into the broader arc of AI engineering:

- Prompt + workflow: Single calls with rigid, hard-coded pipelines. Predictable, but inflexible.

- Context engineering: Supports longer reasoning chains and more flexible orchestration — but the agent can drift mid-task, and someone usually needs to stay close.

- Harness engineering: Unattended execution. You write the standards, principles, quality checks, and tool configurations — the “playbook” — upfront. When the agent hits a decision point, it consults the playbook. The better the playbook, the less you need to intervene.

Still abstract? Let me make it concrete.

My Case: Building a Harness for book2skills

book2skills is a skills factory — it takes classic non-fiction books and distills the frameworks and decision logic inside them into AI-ready skills, open source and free.

The full production pipeline looks like this:

> Pick a book → Read it → Extract the skill → Publish to GitHub → Generate usage examples → Generate website content

I handle book selection myself — I enjoy it too much to hand it off. Everything else runs automatically.

The whole harness took me about two days to build, across four steps. None of it was generated in one shot — roughly 60% of my time went into iteration and refinement.

The Four Steps to build harness

Step 1: Write Thorough Requirements

Before building anything, I spent about two hours getting the brief right. I had the AI help me research competitors, understand what users actually want, map out implementation approaches, and find design inspiration. The goal was to turn vague instincts into explicit written standards.

Step 2: Build Skills at the Right Granularity

This took three to four hours and went through three phases:

first, get something working end-to-end — even if it’s rough, even if the input is a simple PDF and the output is a basic skill;

then refine each skill individually;

then split any skill that grows too long into two. Skills shouldn’t be too large or too small. Too large and the agent drifts. Too granular and you’re essentially writing code. Finding the right size takes trial and error.

Finally I ended up with a fully automated publishing pipeline: One main skill book2skills-publisher, and then five subskills.

Each step has its own dedicated skill, all orchestrated by the main skill. But the orchestration isn’t rigid like a Coze workflow — it’s written in plain language, loosely structured. The agent can adapt as it goes. That flexibility is precisely what makes skills more powerful than hard-coded automation.

The pipeline goes like

> Pick a book → Read it → Extract the skill → Publish to GitHub → Generate usage examples → Generate website content

Here’s what each skill does:

Book selection: The one step I do myself. Not because AI can’t — I just genuinely love this part.

Reading (read-book-skill): Runs local Python scripts to extract the PDF chapter by chapter, then structures the content into markdown summaries. Claude Code’s native PDF capabilities are the foundation here.

Skill extraction (book-skill-creator): Takes the chapter summaries and distills them into a SKILL.md — following Anthropic’s skill-creator spec, covering skill dimensions, query-response frameworks, and output formats. This is the step that most determines skill quality, and the one most worth continuing to refine.

GitHub packaging (write-skill-repo): Takes the SKILL.md and generates a complete, spec-compliant GitHub folder — cleaned SKILL.md, README, LICENSE — zipped and committed to the local repo.

Usage examples (book-skills-examples): Most skills ship with no real demonstration of what they actually do. I wanted to fix that. This skill has the AI genuinely invoke the skill, combined with web search, to produce real usage examples — two to three rounds of conversation, with real data references. Readers can see exactly what the skill does before deciding whether it’s useful to them.

Website content (write-skills-page-content): Pulls everything together into a publishable page — skill description, examples, install instructions, download link — and deploys it to book2skills.

Step 3: Write an Anchoring Document

Once the pipeline was running and the project had a clear shape, I wrote a “What is book2skills” document before doing any further refinement. Its real purpose wasn’t to explain the product to the outside world — it was to set a boundary for myself, to make sure the skills I was about to build wouldn’t be over-engineered or scope-creep into something they weren’t meant to be. I ended up publishing it as book2skills’ first blog post.

Step 4: Let Skills Iterate on Themselves

This is where things get interesting. If I find a flaw in a skill, I just tell it: “add this rule.” It updates its own file. The change takes effect on the next run. No developer needed.

Over time, each skill accumulates its own references and specs — like a master craftsman’s handbook, full of standards and precedents for the apprentice to consult. The richer that handbook gets, the stronger the harness becomes.

What’s Next

The book-to-skill pipeline is done. Next up: bringing operations, UI, and SEO into the harness as well, that’s why you are seeing more skills as below. I’ll keep sharing as that unfolds.

If any of this resonates, feel free to follow along.

• ARK trimmed HOOD across 6 consecutive quarters (2024 Q1–2025 Q3), reducing from 34.94M to 5.02M shares (-86%) while the stock went from ~$12.70 to ~$143

• ARK cut COIN by ~48.6% in a single quarter (Q1 2024), from ~8.60M to ~4.42M shares, with the stock around $265 at the time

• Both positions recorded positive profit differentials

• COIN’s selling was rule-triggered: ARK’s ETF rules cap any position at 10% of fund assets; COIN crossed that line after surging ~160% in 2023

• HOOD: no public explanation was ever given for the 6-quarter reduction, even while Wood maintained a publicly bullish thesis on Robinhood throughout

• In February 2024, Wood called Coinbase “executing brilliantly” — while ARK was simultaneously selling roughly $100M of COIN

Part 1: Same Fund, Two Completely Different Selling Strategies

ARK Investment Management held both Robinhood Markets (HOOD) and Coinbase Global (COIN) — two companies both classified as “fintech.” But looking at 12 quarters of 13F filings, ARK treated these two positions in strikingly different ways.

For HOOD, ARK started trimming in Q1 2024 and continued for six consecutive quarters, reducing from a peak of roughly 34.94 million shares to about 5.02 million shares by Q3 2025 — a reduction of roughly 86%. During that entire period, HOOD’s stock climbed from about $12.70 to around $143.

For COIN, ARK took a completely different approach: in a single quarter — Q1 2024 — it cut roughly 4.18 million shares, a drop of about 48.6%, nearly halving the position in one move. COIN was trading around $265 per share at the time. It later reached an all-time high of $444.65 in July 2025.

Both trades produced a positive profit differential — meaning the market value declined by less than the number of shares did. But the triggers, the pacing, and what ARK “left on the table” were completely different in each case.

Before diving into the data, it helps to understand the fund behind these decisions.

Part 2: Who Is ARK and Who Is Cathie Wood?

ARK Investment Management was founded by Cathie Wood in 2014, headquartered in Florida. Wood had spent roughly a decade as Chief Investment Officer of global thematic strategies at AllianceBernstein. When she proposed building an actively managed ETF (exchange-traded fund) focused on “disruptive innovation,” the firm turned her down. She left and launched ARK on her own.

ARK’s core thesis centers on five “disruptive innovation” platforms: artificial intelligence and automation, genomics, energy storage (including electric vehicles), blockchain and cryptocurrency, and space exploration. The investment logic: find companies that could fundamentally reshape an entire industry within five years, hold large concentrated positions, and ignore short-term volatility.

That logic produced one of the most extreme performance records in modern fund management:

In 2020, ARKK was the best-performing fund globally among funds with more than $1 billion under management — Bloomberg named Wood its stock picker of the year. In 2022, the Federal Reserve’s aggressive rate increases crushed high-valuation growth stocks, and ARKK fell 67%, one of the worst-performing major equity funds of that year.

As of early 2025, ARK manages roughly $26.2 billion across six thematic ETFs. Understanding this fund’s style — high concentration, high volatility, long-horizon bets on disruptive change — provides useful context for what ARK was doing in these two fintech positions.

Part 3: What These Two Companies Actually Do

Robinhood: The App That Turned Stock Trading Into a Game

Two Stanford graduates founded Robinhood in 2013 with one core pitch: commission-free stock trading. Before that, traditional U.S. brokerages typically charged $7 to $10 per trade, keeping many retail investors on the sidelines. Robinhood eliminated that cost, built an app that looked more like a game than a brokerage, and added confetti animations when trades went through. It quickly attracted a large wave of younger, first-time investors.

“Free” doesn’t mean the company makes no money. Robinhood earns from several sources:

• Payment for Order Flow (PFOF): Robinhood’s largest and most controversial revenue source. To understand it, you need to understand why market makers exist at all.

  U.S. retail investors can trade directly in the stock market — but there’s a practical problem: when you submit an order, there isn’t always someone on the other side ready to trade with you right then. The New York Stock Exchange and Nasdaq are matching systems — they pair buyers with sellers. If you want to buy 100 shares at $50 and someone else is selling at $50 at exactly that moment, the trade happens instantly. If no one is, your order waits — seconds, minutes, or indefinitely. The thinner the market, the less predictable your execution.

  Market makers solve the “trade whenever you want” problem. Think of them like a used-car dealer: if you want to sell your car today, you don’t have to spend months finding a private buyer — you sell directly to the dealer, who takes it and finds a buyer later. The dealer keeps a spread as profit and compensation for holding the risk. Stock market makers (like Citadel) do exactly the same thing: they simultaneously post a buy price (slightly below market) and a sell price (slightly above), ready to be your counterparty at any time, profiting from that spread.

  Robinhood routes its users’ orders to these market makers. The market makers get a reliable stream of retail order flow; in return they pay Robinhood a rebate. That’s PFOF. In 2024, it was roughly 56% of Robinhood’s total revenue.

• Cryptocurrency spread: Robinhood supports Bitcoin, Ethereum, and other crypto trading. It calls this “commission-free” — but it earns through the spread. If Bitcoin’s market price is $100,000, you might see a buy price of $101,750 and a sell price of $98,250 on Robinhood. That 1.75% gap is Robinhood’s income — it doesn’t appear as a labeled fee, but it’s there in every transaction. In Q4 2024, crypto spread revenue grew more than 700% year-over-year to roughly $358 million in a single quarter.

• Gold membership: $5/month for margin loans, higher-yield cash, and premium research.

• Interest on cash balances.

Robinhood’s total revenue: roughly $2.95 billion in full-year 2024 (up ~58%) and roughly $4.47 billion in full-year 2025 (up ~52%). (Robinhood annual reports, 2024 & 2025)

Coinbase: The Regulated Exchange for Crypto

Coinbase was founded in 2012, is the largest regulated cryptocurrency exchange in the United States, and went public on Nasdaq in 2021. Its revenue comes from:

• Transaction fees: ~0.4–1.5% per trade, contributing about 60% of 2024 revenue.

• Staking revenue: Blockchain networks need ongoing transaction validation. Holders can participate by locking up tokens as collateral — like a fixed-term deposit — and receiving rewards (similar to interest) in return. Coinbase manages this process for users, taking a service fee from the rewards. In Q4 2024: roughly $215M/quarter.

• USDC stable coin interest sharing: Coinbase co-issues USDC, a dollar-pegged stable coin, and shares interest income from its reserve assets.

• Institutional custody fees: Coinbase holds Bitcoin on behalf of large ETF issuers like BlackRock, charging custody fees.

Coinbase’s full-year 2024 revenue: roughly $6.2 billion, up ~113% year-over-year, with net income of roughly $2.5 billion. (Fortune, February 14, 2025)

Why Coinbase’s revenue is so tightly tied to the crypto market: It earns primarily from transaction fees — more trading activity means more revenue, and less trading activity means less. The numbers tell the story directly: 2021 bull market, ~$7.7 billion in revenue; 2022 bear market (Bitcoin fell from ~$69K to ~$16K), ~$3 billion in revenue and ~$2.5 billion net loss.

Robinhood, by comparison, has crypto revenue as only one piece of its total income. Its PFOF revenue covers all asset classes. That’s why two “fintech” companies can have meaningfully different risk profiles when crypto markets move.

Part 4: The 12-Quarter Holdings Timeline

Caption: Block size represents position weight — shows the relative size of HOOD and COIN within ARK’s total portfolio

HOOD: Six Quarters of Steady Selling

At the end of 2023, ARK held roughly 34.94 million shares of HOOD at approximately $12.70 per share — market value about $443 million.

Starting Q1 2024, ARK began trimming: down to ~26.20M shares (−25%). Then ~19.15M (another −27%). Then ~12.91M (another −28%). Then ~7.79M. By Q3 2025: ~5.02M shares with HOOD at ~$143, giving a remaining market value of ~$718 million — higher than the $443M value of the full position at the end of 2023.

ARK sold 86% of the position. The remaining 14% was worth more than the original 100%. But if ARK had held all 34.94M shares to Q3 2025 at ~$143, the total would have been roughly $4.99 billion. The selling process also meant forgoing that upside.

(MacroTrends, 2026)

COIN: Nearly Half Gone in One Quarter

In 2023, COIN surged roughly 160% — from ~$67 to ~$174 per share. ARK made modest reductions during this period (~27% of shares trimmed). Then in Q1 2024: ARK cut from ~8.60M to ~4.42M shares — a 48.6% reduction in a single quarter, with COIN at roughly $265.

The documented reason: ARK’s ETF rules cap any single holding at 10% of fund assets. COIN had exceeded that limit after its 2023 surge, triggering systematic forced selling. This was compliance execution, not a change in fundamental view. (CoinDesk, January 4, 2024)

COIN later hit an all-time high of $444.65 in July 2025, roughly 68% above the level at which ARK made most of its cuts.

Two trajectories: HOOD is a slow, steady decline in share count against a rising stock price. COIN is a sharp single-quarter cut, then a more stable, gently declining holding. Both produced positive profit differentials — but the logic behind each is completely different.

Part 5: What Wood Said vs. What ARK Did

Putting Wood’s public statements alongside ARK’s 13F data on the same timeline reveals some things worth examining closely.

COIN: Bullish in Public, Selling in Practice

(Watcher.guru | Crypto.news | The Block)

Wood’s explanation is clear: not bearish, just a rule-triggered profit realization. She maintained she was “very bullish” on Coinbase. That’s logically consistent — but from the outside, calling the stock “executing brilliantly” while simultaneously selling ~$100M of it created reasonable confusion. The explanation came reactively, in response to direct questions. COIN subsequently climbed to $444.65, about 68% above the ~$265 selling level.

HOOD: Selling for Six Quarters, Public Thesis Never Changed

Over the entire six-quarter selling period (2024 Q1–2025 Q3), Wood never publicly stated a reason for reducing the HOOD position. Her public thesis on Robinhood remained consistently positive throughout — “digital wallet,” “financial democratization,” next-generation internet finance.

Yet 13F filings show the position going from ~34.94M shares to ~5.02M shares.

Then, after HOOD pulled back from ~$143, ARK bought back more than $50 million of HOOD stock in December 2025 and February 2026 (TheStreet | Motley Fool). That confirms the thesis was intact — but doesn’t explain why six consecutive quarters of selling happened in the first place.

Three possible explanations, none confirmable from public filings:

• Same 10% cap rule: HOOD surged ~10x, possibly triggering the same limit — but unlike COIN, Wood never said so

• Capital rotation: ARK was sometimes buying COIN on the same days it was selling HOOD (Investing.com)

• Reasons not confirmable from public information

The contrast between these two cases is striking: COIN’s reduction came with a clear explanation, given when pressed. HOOD’s six-quarter selling left no public reasoning at any point.

If COIN’s reduction at least had a rule to point to — the 10% cap was triggered, the selling was a compliance action — then HOOD’s case is harder to explain away: six consecutive quarters, 86% of the position gone, public thesis unchanged, no explanation offered. That gap may be the most practical lesson this data set offers: pay attention not just to what fund managers say, but to whether their actual holdings stay consistent with what they’re saying.

Part 6: What Can an Ordinary Investor Take From This?

A few observations worth noting — keeping in mind these are specific to the context here, not reusable trading rules:

First: an institution selling doesn’t necessarily mean it’s changed its view on the company.

COIN’s selling had a documented rule-based trigger. For ordinary investors, large institutional sells shouldn’t automatically be read as “this fund no longer believes in this company.” Sometimes it’s purely a compliance result.

Second: selling while the stock goes up isn’t the same as selling at the top.

Both statements are simultaneously true for HOOD: ARK generated a positive profit differential from its sales, and it also forfeited an enormous amount of potential upside (~$4.99 billion vs. ~$718 million held). Results alone don’t settle whether the decision was correct.

Third: understanding what a company does helps make sense of institutional decisions.

Coinbase’s revenue is ~60% directly tied to crypto trading volume. Robinhood’s crypto exposure is one part of a more diversified revenue mix. Understanding this helps explain why ARK was sometimes selling HOOD and buying COIN on the same day — they aren’t identical bets on the crypto market.

• Berkshire went from 1.033 billion BAC shares to 517 million (-50%) between mid-2024 and end-2025. BAC’s stock price rose from ~$42 to $50–56 over the same period.

• The 13-year trade made sense from a valuation standpoint — entry at a crisis discount, exit as the premium expanded.

The Counterintuitive Setup

On July 17, 2024, Berkshire Hathaway started selling Bank of America (BAC). Over the next 12 trading sessions, Berkshire sold continuously, raising about $3.8 billion.

— Fortune, Aug 2024

The selling continued for 18 months. By the end of 2025, the position had dropped from a peak of 1.033 billion shares to 517 million — exactly half.

But during those same 18 months, BAC’s stock price went from roughly $42 to $50–$56.

A 13-Year Trade: From Crisis Discount to Premium Exit

As of Q4 2025, BAC represented roughly 10.7% of Berkshire’s portfolio — the third-largest position after AAPL and AXP.

BAC wasn’t Berkshire’s biggest holding, but it saw the largest reduction among core positions. To understand the exit, you have to understand the entry.

2011: Buying Into a Crisis

In August 2011, Bank of America was in trouble. Its stock traded at just 38% of book value (a 62% discount) — the market was pricing in the possibility that hidden losses could bring the bank down.

Buffett invested $5 billion in BAC preferred stock and received warrants: the right to buy 700 million shares of common stock at $7.14 per share at any future date.

— Blackstone Press, 2011

This is exactly how he Buy the Fear during the financial crisis

2017: Exercising the Warrants

By summer 2017, BAC had recovered. The stock was trading at $24.32. Buffett exercised all 700 million warrants at $7.14, generating an immediate paper gain of roughly $12 billion in a single transaction.

— CNBC, Jun 2017

By end of 2021, after additional market purchases, Berkshire’s blended cost basis was approximately $14.15 per share across roughly 1.033 billion shares, for a total cost of about $14.6 billion.

— Motley Fool, Mar 2021

2023–2025: Six Quarters of No Action then sell

There are basically 2 phases during this period.

From Q1 2023 through Q2 2024, Berkshire held its 1.033 billion BAC shares without making a single change. Over that period, the stock climbed from about $28.6 to $39.8 per share, and the position’s market value grew from roughly $29.5 billion to $41.1 billion.

This included March 2023, when regional bank failures (Silicon Valley Bank) sent shockwaves through the sector. Berkshire didn’t reduce.

Starting July 17, 2024, Berkshire began selling. Within 12 trading days, it had raised about $3.8 billion. The selling continued every quarter:

• Q3 2024: down to 797.7 million shares (-22.8%, roughly 235 million shares sold)

• Q4 2024: 680.2 million

• Q1 2025: 631.6 million

• Q2 2025: 605.3 million

• Q3 2025: 568.1 million

• Q4 2025: 517.3 million

— CNBC, Jul 2024

By Q4 2025, BAC was trading at $50–$56 per share. At 517 million shares and roughly $55 per share, the remaining position was worth about $28.5 billion — down 31% from the $41.1 billion peak.

This is exactly how he Sell the Euphoria during the financial crisis

The data points to something specific: Berkshire was selling into a rising market. Why?

Why Sell Into a Rising Stock?

Reason 1: The Valuation Thesis Had Run Its Course

This is the most straightforward explanation.

Berkshire bought BAC in 2011 because it was trading at a steep discount to book value (38% of book — a 62% discount). By 2024, BAC was trading at roughly 40% above book value.

— Motley Fool, Mar 2026

The original thesis — “the market is undervaluing this bank” — had not only expired, it had reversed. The market was now overvaluing it by Berkshire’s entry-price logic. A crisis discount entry naturally pairs with a premium exit.

Reason 2: Interest Rate Cycle Shifting

BAC is the most interest-rate-sensitive of all large US banks. When the Federal Reserve raised rates by 525 basis points between March 2022 and July 2023, no major bank benefited more in net interest income than Bank of America. As the rate cycle neared its end, that advantage was fading.

— Fortune, Sep 2024

So Was the Timing Right?

From a profitability standpoint: absolutely. Cost basis of $14.15 per share, selling in the $40–$55 range. Every sale was profitable.

From a “maximum return” standpoint: not perfect. Selling started at $42, and the stock went on to $55+. If “right timing” means selling everything at the top, Berkshire didn’t do that.

What Investors Can Take Away

1. The exit signal is the expiry of your original thesis, not the price level

Berkshire bought BAC because it was priced at a crisis discount. When that discount turned into a premium, the thesis expired. The question to ask isn’t “how much has this gone up?” — it’s “is the reason I bought this still true?”

Going from a 62% discount to a 40% premium is an observable, data-driven answer to that question.

2. Entry price determines how much margin for error you have on exit

With a blended cost basis of $14.15 and selling between $40 and $55, every transaction was profitable. Low-cost entry creates decision space: you’re not forced to find the perfect exit because almost any exit above cost generates returns. Getting the entry right tends to matter more than getting the exit perfect.

⚠️ Disclaimer: This content is generated by AlphaWiseWin based on publicly disclosed Form 13F data. Holdings data reflects approximately 45-day reporting delays and does not represent current positions. This is not investment advice. Investing involves risk.

Sources

• Fortune: Berkshire generates $3.8B selling BofA stock in 12 days (Aug 2024)

• CNBC: Berkshire sells BAC for ninth straight day (Jul 2024)

• Motley Fool: Buffett sells 50% of BAC (Mar 2026)

• Motley Fool: Buffett sold 45% of BAC (Dec 2025)

• CNBC: Buffett makes $12B on BAC warrants (Jun 2017)

• Motley Fool: Buffett’s cost basis for top holdings (Mar 2021)

• Fortune: Why is Buffett selling BAC? (Sep 2024)

TL;DR: Berkshire accumulated OXY for 12 straight quarters — share count up ~25%, but the position was underwater the entire time (slope gap: -3.68%).

On February 28, 2026, the US and Israel struck Iran. WTI hit $119.48 intraday. OXY gained +16.9% in February alone .

The war is definitely NOT the reason why Berkshire bought, but the basic logic behind the decision is worthy of taking a look.

The Hook

“She(Occidental Petroleum CEO Vicki Hollub) knows how to separate oil from rock — that’s a rare talent, valuable to her shareholders and to her country.”

That’s what Buffett wrote in the 2023 Berkshire shareholder letter about Occidental Petroleum (OXY) CEO Vicki Hollub. When he wrote those words, OXY’s stock had already fallen nearly 30% from its 2022 peak.

Berkshire was still buying.

By March 2026, the dual-axis slope analysis(below) tells a precise story: over the past 12 quarters, Berkshire’s OXY share count grew from 211M to approximately 265–284M, a net increase of ~25% over three years, but the market value fell from 13B to 11B.

Buying more while losing money. Never selling. Never stopping.

Then on February 28, 2026, the US and Israel struck Iran — and the position revalued itself. But what I want to walk through today is not the war. It’s the three years of logic that made the payoff possible.

Historical Position Data

From the full portfolio view, OXY is not Berkshire’s largest position, it take around 4% .

— but I’d argue its status is uniquely significant:

• Berkshire holds ~26.9% of OXY’s outstanding shares — the single highest ownership stake in any company across the entire portfolio ✅ Motley Fool, Dec 2024

• One of the most consistent accumulation targets over the past 12 quarters — zero reduction quarters on record

• One of eight stocks Buffett explicitly placed on the “long-term hold” list in the 2024 shareholder letter, alongside Coca-Cola and American Express ✅ Motley Fool, 2025-01-08

From the historical position changes in the chart below, after rapidly building his position in early 2022, Buffett steadily increased his holdings for the next 12 quarters without any record of reducing his position.

Here’s the timeline that built up to February 2026.

Timeline of OXY positions

• Early 2022 — Russia invades Ukraine. Oil spikes. OXY rises from ~$30 to the $75 range. Berkshire builds its first large position during the geopolitical shock and continues adding as oil briefly touched $130/barrel.

• Full year 2023 — Inflation cools. Oil pulls back to $55–65. Berkshire holds steady — no meaningful reduction.

• 2024 — OXY falls ~17% for the year, down ~32% from historical highs, settling around $46. Berkshire doesn’t cut losses. It adds more in 2024Q4 at an average of $45.55–$46.92 per share. Total position as of the 2025Q4 13F cutoff: approximately 265–284 million shares.

• February 28, 2026 — The US and Israel strike Iran. WTI breaks $100/barrel, reaching an intraday high of $119.48. OXY gains +16.9% in February alone , then adds another ~+4% in early March.

According to Stockcircle, Berkshire’s weighted average cost basis across 12 purchases is ~$51.76/share . At the early March 2026 price of ~$54.19, after the February surge, the position has moved from underwater to in-the-money.

The timeline is clear. But numbers don’t explain motivation — so the more important question is what kept Berkshire in the trade for three years before the payoff.

Buffett’s Two Public Positions on OXY

Buffett has been unusually explicit about why he holds OXY. Two statements stand out:

1. Operator credibility: “She knows how to separate oil from rock — that’s a rare talent.” (2023 shareholder letter) — This isn’t an oil price forecast. It’s a specific assessment of management’s technical capability and operational efficiency.

2. Long-term intent: OXY placed on Berkshire’s “hold forever” candidate list in the 2024 shareholder letter — alongside Coca-Cola and American Express, both of which Berkshire has held for decades.

The Real Bet: Capability Over Price Prediction

What I think is often misread about this position: Buffett’s core thesis on OXY was never “oil prices will go up.” The thesis had three specific legs:

• OXY’s per-unit extraction cost is falling — Hollub’s technical expertise is converting the same asset base into progressively higher output at lower cost. That’s a capability edge, not a commodity bet.

• OXY holds some of the best Permian Basin acreage in the US — no Middle East supply dependency. The asset quality doesn’t disappear when oil prices fall.

• Survive the lows, maximize the highs — At $50–55/barrel, OXY services its debt and maintains operations. At $100+, it generates outsized free cash flow. That asymmetry is the actual position.

This “capability bet” framework explains the 12-quarter losing streak without a single reduction: the thesis wasn’t disproved by lower oil prices. The position was structurally set up to wait — for prices to recover naturally, or for an external shock to force recovery fast.

What You Can Take From This

1. Capability beats price prediction

Buying a commodity stock is not a bet on price direction. The right question is: can this company survive at low prices? And when prices recover, how much leverage does it carry? OXY’s Permian acreage and Hollub’s operational track record are the answers — not the oil futures curve.

To watch: The 2026Q1 13F (~mid-May 2026) will reveal whether Berkshire reduced OXY after the oil spike — the key data point for understanding whether this was always a trade with a planned exit, or a permanent holding in the making.

⚠️ Disclaimer: This content is produced by AlphaWiseWin based on publicly available 13F filings. Position data carries approximately 45-day reporting delays and does not represent current holdings. This is not investment advice.

If you found this useful, consider subscribing — we track institutional portfolios to find where conviction capital is actually moving.

In Q4 2025, Berkshire initiated a $352M position in NYT — one of only two new stocks they’ve added in three years. The other was Google. That pairing is not a coincidence.

Story was different 5 years ago

In 2020, Warren Buffett sold Berkshire Hathaway’s entire newspaper portfolio — 31 papers including the Omaha World-Herald and the Buffalo News — for $140 million. He called the newspaper industry “toast.”

Five years later, in Q4 2025, Berkshire’s 13F (Form 13F is a quarterly holding disclosure required of large institutional investors) revealed a brand-new position: 5,065,744 shares of The New York Times Company (NYT), valued at $352 million at December 31, 2025.

One quarter before that, in Q3 2025, Berkshire had opened its first-ever position in Alphabet (GOOGL) — 17.85 million shares, approximately $4.3 billion.

Two information companies. Two consecutive quarters.

For a portfolio that otherwise moves at 4.7% turnover per quarter ( you can see how stable it is in picture below) This warrants a closer look.

Over those three full years, the firm initiated exactly two positions it had never held before: Alphabet and The New York Times. Any new name here clears an extremely high bar.

What Berkshire Sold in 2020 vs. What It Bought in 2025

This is the core of the story — and why the two decisions aren’t contradictory.

The 31 papers sold in 2020 were regional businesses built on local advertising. Their value depended on car dealerships, real estate agents, and local retailers who had no other way to reach local audiences. Once those advertisers found Facebook and Google, the papers’ revenue base didn’t shift — it disappeared. The decline was structural, not cyclical.

Even at the time of the sale, Buffett drew a line. He said that newspapers with national brands — specifically the Times and the Wall Street Journal — might be different. National brands can charge readers directly, anywhere. They’re not dependent on a local advertiser ecosystem.

The New York Times Company in 2025 looks nothing like a regional newspaper:

• Paid digital subscribers: 12.78 million (+1.4M net new in 2025)

• Digital revenue: $2 billion+ (first time crossing this threshold)

• Digital subscription revenue: $1.4 billion (+14% year-over-year)

• Bundle adoption (news + Games + Cooking + Wirecutter): +24% year-over-year

• Digital advertising: $411 million (+20%)

• Adjusted operating profit: $550 million (+21%)

Digital subscription now accounts for more than 60% of total revenue. The business runs four distinct subscription products — news, Games (Wordle, crossword), Cooking, and Wirecutter (consumer product reviews) — each creating independent recurring habits. This is closer to a software subscription model than a newspaper.

NYT + GOOGL: One Thesis, Two Angles

One quarter before NYT, Berkshire built its first-ever Alphabet position at an average cost of approximately $225 per share.

Put them together and a common logic emerges — a bet on the information supply chain.

• Google controls the discovery layer: where people go to find information.

• NYT controls a trust layer: where people go for information they believe.

Both are dominant players at different points in how information flows from creation to consumption. And both face the same macro tailwind: as AI-generated content floods the internet, the volume of total information increases while the supply of credible, editorially verified information stays relatively fixed. Scarcity of trust could become a genuine competitive advantage.

Position Details

NYT’s Q4 2025 trading range: $54.10 (October 14 low) to $71.23 (December 24 high), closing the year at $69.42. Berkshire’s average cost falls somewhere in this range. As of March 2026, NYT trades around $80.44 — an unrealized gain of over 10% from year-end.

At 0.13% of the portfolio, this is not a large position for Berkshire. But in a fund this concentrated and this selective, the fact that it exists at all carries signal.

What This Framework Teaches

1. The relevant question is never “is this industry in decline?”

The regional newspaper industry was in structural decline. NYT was in the same industry by label, but not by business model. The two situations were never comparable in Buffett’s framework — he said so explicitly at the time of the 2020 sale. Broad industry labels often mask the difference between companies that are exposed to a structural headwind and companies that have already moved past it.

2. In a low-turnover portfolio, every new position carries information

At 4.7% quarterly turnover, new entries are rare. But there’s a structural problem with following them: 13F filings are released approximately 45 days after quarter-end. The NYT position became public knowledge in February 2026, at which point the stock had already moved substantially on the news. The signal has value. The window to act on it at the original price has already closed.

3. Subscription transitions change the valuation framework entirely

When a media business converts its revenue base from advertising (cyclical, advertiser-controlled) to subscriptions (recurring, reader-controlled), the right valuation model changes. Subscription revenue behaves like software-as-a-service — high switching costs, predictable renewal rates, compounding subscriber growth. NYT crossed 60% digital subscription revenue in 2025. At that point, the right peer group is no longer legacy media.

Risks

• 13F lag — Q4 2025 data; Q1 2026 position changes are not yet disclosed.

• Unknown decision-maker — Buffett, Combs, or Weschler; no public disclosure.

• CEO transition — Greg Abel’s portfolio priorities may differ; position could be reduced.

• AI content threat — NYT’s own 2025 10-K flags AI-generated content as a long-term risk to subscriptions.

• Valuation — Entry price-to-earnings ~38x; growth-multiple compression risk if subscriber growth slows.

• Position size — 0.13%; an exit would be invisible at the portfolio level.

Disclaimer: All holdings data is sourced from 13F public filings, subject to approximately 45-day reporting delays, and does not represent current positions. This is not investment advice.

If you found this useful, consider subscribing — we track institutional portfolios to find where conviction capital is actually moving.

Subscribe now