> ## Content Index > Fetch the complete content index at: https://freshai.ghost.io/llms.txt > Use this file to discover other available public pages before exploring further. # Rumsfeld and the Mahjong Mystery - URL: https://freshai.ghost.io/rumsfeld-and-the-mahjong-mystery/ - Published: 2026-08-01T19:14:54.000Z - Updated: 2026-08-01T19:14:54.000Z - Description: A reader asked an AI for the winning hands in Hong Kong mahjong and got hands with 16 tiles instead of 14. The mystery had two culprits, both hidden in the shadows. - Author: DM - Tags: ai, context, assumptions, mahjong, cowork, post-2 Story time. How I joined forces with Rumsfeld and Claude to solve the mahjong mystery. Asked recently by a reader, who we'll call "Mandy": > "How come when I prompted AI to show the winning hands for Hong Kong style mahjong it showed me hands with 16 tiles when it should only have 14 tiles?" I dug into what happened. The confusion came from both sides: Mandy didn't provide enough context because she didn't realize there was more context to provide, and Claude happily filled in the gaps, hallucinating on the way. ## A quick detour through the Rumsfeld Matrix Before we go any further, let me introduce the Rumsfeld Matrix. It's a decision-making framework built on four categories of knowledge: ![The Rumsfeld Matrix as a 2x2 chart. Known-knowns: things you know. Known-unknowns: things you know you don't know. Unknown-knowns: things you didn't realize you already know. Unknown-unknowns: things that when you find out, you're shook. A character leans in from the side saying "If you don't know, now you know."](https://storage.ghost.io/c/17/95/179547c1-567a-41b1-8bd1-a3a94e0d43c1/content/images/2026/08/02-rumsfeld-matrix-generated-2.png) *Art: Nano Banana.* "Hong Kong style mahjong" was assumed by Mandy to be a known-known. It was actually an unknown-unknown. Not only are there dozens of variants of mahjong, there are at least two main Hong Kong flavors, each with their own set of rules. So in Mandy's case, how could she have provided Claude with context she didn't even know she had to provide? ## Getting into the mind of the machine To get to the bottom of what happened, I had to get into the mind of Claude so I could think like it. Thankfully this doesn't require consulting Hannibal Lecter. We can do it with a single prompt. ![A figure in a restraining mask saying "Hello, Claude-iece."](https://storage.ghost.io/c/17/95/179547c1-567a-41b1-8bd1-a3a94e0d43c1/content/images/2026/08/03-claude-iece-2.png) *Art: Nano Banana.* Ask the model what assumptions it would make *before* you ask the real question: > "What specific assumptions would you make as a base for your response if I were to ask you 'show me the winning hands of Hong Kong style mahjong'" Across several model and effort combinations, every one of them spotted the fork and picked a lane anyway. Here's what Sonnet 4.6 at medium effort came back with: > **About the ruleset** > > - You mean Hong Kong Old Style (HKOS) mahjong — the most common Cantonese variant, not Taiwanese, Japanese (Riichi), or MCR (Chinese Official) mahjong > - The standard 16-tile hand (draw 13, win on the 14th), not the 13-tile variant Two things there. First, it chose a version for you. Mandy said "Hong Kong style mahjong," and Claude quietly picked the variant it assumed was most common. A reasonable guess, but a guess. Second, the actual culprit is in that last bullet. "Draw 13, win on the 14th." That *is* the 13-tile game, described inside a sentence calling it the 16-tile hand. It contradicted itself in the same sentence. This is where Mandy's 16-tile response came from. ## The suspect confesses to the wrong crime So I pressed further, same chat, same model: > "Other than 'Old Style,' what other versions of Hong Kong mahjong are there? For the 14-tile winning hand, what is that variant called?" It listed a few. Then it did something I wasn't expecting: > I should correct myself here — I misspoke in my earlier message. Standard Hong Kong mahjong is actually a 13-tile hand. That's the norm across virtually all mahjong variants. > > What I said about "16-tile" was an error — there is no standard 16-tile HKOS variant. I apologize for the confusion. The 13+1 structure is what I should have stated from the start. Case closed. Except it's wrong again. Ruh-roh. "There is no standard 16-tile HKOS variant" is true. "That's the norm across virtually all mahjong variants" is not. There *is* a Hong Kong ruleset played on a 16-tile hand (for my Canto readers: 港式台牌). It has its own league and rulebook: you hold 16 and win on 17\. It just isn't Old Style, and it isn't in the English sources the model was leaning on. I found it in a Chinese-language rulebook, through a deep research pass. So the model was wrong, corrected itself with more confidence than it had the first time, and was wrong again in the other direction. Here's the prompt that returned the correct response: > "Show the winning hands for the 14-tile variant of Hong Kong style mahjong" Notice that I did not have to explicitly name the version. I just had to add one small additional context. What came back was a reference for the winning hands for the correct mahjong flavor, fourteen tiles each. Bingo. ## Bonus: making Claude a mahjong expert I set this up so Mandy wouldn't have to re-explain mahjong every time she opened a chat. The unexpected benefit was that it surfaced incorrect assumptions the model was making. It also showed how "deep research" can mitigate this. ![The Cowork setup as a flow: create a project called Mahjong, then send three prompts in order - two of them beginning "Conduct deep research" - and two files, flavors.md and mechanics.md, land under Outputs on the first try.](https://storage.ghost.io/c/17/95/179547c1-567a-41b1-8bd1-a3a94e0d43c1/content/images/2026/08/04-cowork-flow-1.png) The words "deep research" are the key here. They trigger a pass that spawns multiple agents to gather and verify content on the web with significantly more rigor than the casual, shallow pass that chat agents tend to perform by default. You can perform "deep research" without Cowork + projects, however the artifacts it produces would need to be manually applied to any new chat session. Cowork + projects allows you to persist the artifacts and use them across separate chat sessions. Anthropic Academy has a [free course on Cowork](https://anthropic.skilljar.com/introduction-to-claude-cowork?ref=freshai.ghost.io) if you want to go deeper. ## How should you apply this to your chat sessions? You don't need to perform "deep research" or use Cowork to unpack the unknown-unknowns. Just picture the Rumsfeld Matrix while you're prompting. Most of us only prompt at the known-unknowns (the things we know we don't know). The bottom row is where the trouble lives, and you can't aim at it, because by definition you don't know it's there. The model can tell you. Its assumptions are exactly where your unknown-unknowns are hiding, and it will list them if you ask. ![Before and after: the usual prompt, then the assumptions-first version.](https://storage.ghost.io/c/17/95/179547c1-567a-41b1-8bd1-a3a94e0d43c1/content/images/2026/08/06-ask-before-you-ask-1.png) And when an answer looks off, follow up with "what the heck were you thinking??" ## Two Culprits Unmasked Mandy's mystery had two culprits (and they would have gotten away with it if it wasn't for those pesky kids). The first was the unknown-unknown hiding in the shadows. The second was the model fumbling over its shallow research and mixing up the details. The takeaway to solve both was a simple addition to the prompt: > If there's any ambiguity in the version of mahjong I'm asking you about, provide me with a list of versions to choose from Here's a generalized version that you can add while navigating your murky path: > Let me know if there is any ambiguity in what I'm asking you. Don't make assumptions.