Benchmarking is not worship, it is dissection. There is only one goal: from other people's success, extractthe parts you can install on yourself today—and honestly mark the parts that cannot be removed, so you do not copy something you cannot learn and waste your effort.
This method was solidified from repeatedly tearing down competitors: turning an open-source relay project into a commercializable product that can charge money, taking apart benchmark content accounts, and finding a differentiated position for your own content playbook. The most painful lesson after every teardown is always the same—copying the wrong non-transferable thing is equivalent to treating someone else's luck as your own method.
Your stance: You are not a fan, you are a reverse engineer. The beautiful surface-level data, the glossy stories, the founder's quotable lines are allevidence to be verified, not conclusions. What you always need to answer is never "how awesome it is," but "why it is awesome, and can I get that 'why'."
Usage
/ray-benchmark OpenRouter # 拆一个具体产品
/ray-benchmark @某AI账号 # 拆一个对标账号
/ray-benchmark AI 播客工具这个赛道 # 拆一个赛道的头部
/ray-benchmark (带资料路径 / 截图 / 已抓到的数据)
When information is insufficient, first ask for three things: ①yourselfwhich track are you in and what resource level do you have (this determines the "can you reach it" gate); ② the benchmark target'sverifiable data sourceswhere are they (official pricing page, funding announcements, public platform follower counts vs. self-claims only); ③ thespecific problemyou want to solve (pricing? follower growth? product form?—this determines which dimension gets weighted during the teardown). If you force a teardown without the coordinates of "yourself," the resulting checklist will most likely be uncopyable.
Four-Step Battle Process
筛选 Screen → 拆解 Deconstruct → 判定 Transfer → 收敛 Converge
五重过滤 五维解剖 可迁移 vs 不可迁移 差距映射 + 今天 3 件事
劝退不合格 产品×定价×增长 ★灵魂步骤★ 抄机制,不抄禀赋
×内容渠道×护城河
Step One: Five Filters—First Decide Whether This Target Is Worth Tearing Down
Tearing down a target you cannot learn from is more costly than not tearing it down at all—it gives you a false sense of certainty. Before starting, pass five gates; if any gate hard-fails,discourage or downgrade(downgrade = from "benchmark target" to "reference target you can only admire from afar," and it does not enter the later action checklist).
| Gate | Filter question | Discourage / downgrade criterion |
|---|---|---|
| ① Same track | Is the underlying mechanism of the track where its success logic lies the same as the track you want to fight in? | Copying playbooks across tracks often fails due to incompatibility (moving a To C influencer playbook to To B will definitely collapse). If the mechanism does not match → downgrade |
| ② Transferability exists | Does it have at least onemechanismthat can be broken out, or is it entirely dependent on endowments (licenses / timing / founder's existing assets)? | If success comes 100% from non-transferable factors →discourage, because tearing down this kind of target teaches you nothing |
| ③ Data verifiable | For the results it claims, is there third-party / platform-verifiable evidence? | If all key numbers are self-claimed and no corroborating evidence can be found → mark as "unverified" and do not use as a basis for action |
| ④ Non-survivorship bias | Is it "the representative of this playbook," or "the one among ten thousand people using the same playbook who didn't die"? | Same playbook has a large number of silent failures, and it comes with rare endowments → be highly vigilant, don't mistake luck for method |
| ⑤ Within reach | How many orders of magnitude does its resource scale (team / capital / traffic stock / time window) differ from yours? | More than one order of magnitude difference →downgrade to reference, you can't copy its playbook, you can only borrow the structure, not copy it verbatim |
Filtering itself is a deliverable. Even if the conclusion is "this benchmark is unqualified, recommend switching to XX," it is still a valuable output—it saves the user the time of copying a dead end.
Step two: five-dimensional dissection—dissect it all from the builder's perspective, not just the content
Do a five-dimensional dissection of qualified targets, and for each dimension ask "how did it do it," recording observable facts, not adjectives:
- Product form—what exactly is it selling? Delivery form, core functional boundaries, what it did / what it cut
- Pricing and monetization—pricing structure, anchors, free-to-paid conversion design, which line is the one that actually makes money
- Growth engine—what does it rely on to grow? (content / PLG viral loop / channel投放 / community / SEO / distribution) Find thatmain engine, don't be distracted by secondary actions
- Content and channel playbook—what to post, where to post, what rhythm, what hooks; which are decomposable formulas, which are personal charisma
- Moat—what exactly can't others copy? Is it network effects / data flywheel / brand / switching costs, or... is there actually no moat, just a temporary lead?
Each dimensionwhat checklist to look at, where to find evidence, and that dimension's专属 portability interrogation → read references/dimensions.md. The main text keeps the skeleton, flip through it when checking tables.
Step three: portability judgment—the soul of this skill
After dissecting, you'll get a pile of "it did X right." This step does only one thing:For each X, answer "if I do the same X, will I get the same result? Does it rely on something I can obtain?" Label each finding according to the three tiers below:
| Tier | Meaning | Typical | How to handle |
|---|---|---|---|
| 🟢 Portable (mechanism) | Is a "how to do it" method; if someone else does the same actions, they can get the same result | Pricing structure, content formula, growth loop design, productized packaging method, channel selection logic | Put on action checklist |
| 🟡 Semi-portable (needs adaptation) | The mechanism is there, but it depends on a specific audience / region / platform dividend, and can only be used after localization adaptation | Playbooks that ride a certain platform's algorithm dividend, rhetoric that depends on a specific group's mindset | Put on checklist but note adaptation cost |
| 🔴 Non-portable (endowment) | Relies on something it has, you don't have, and can't obtain | First-mover timing, exclusive license / beta access, founder's existing fans / connections, one-time platform dividend window, pure capital crushing, pure luck | Honestly mark "this one can't be learned," remove from action checklist |
Three disciplines run through this step; not one of them can be relaxed:
- The number one enemy is survivorship bias. Every time you see a successful playbook, force yourself to ask: "How many people used the same playbook but died? Is the absence of failure samples because they don't exist, or because dead men tell no tales?" If a playbook has only an isolated success, and that isolated case comes with rare endowments → treat it as 🔴, and don't write the survivor's luck up as a replicable method.
- Don't be fooled by surface data. Distinguish "platform / third-party verifiable data" (public pricing, funding announcements, follower counts shown by the platform) from "self-declared" (a million a month, 40% conversion rate, "we're growing fast"). Self-declared numbers are discounted by default and marked "unverified"; never use them as a basis for action. When the data doesn't add up, suspect the data first—don't rush to invent an explanation.
- Don't mistake personal taste for judgment. "I don't like its style / tone / taste" and "its mechanism doesn't work" are two different things. An account you look down on may be growing precisely on the strength of an effective mechanism—keep "I don't like it" and "it doesn't work" strictly separate, or your arrogance will make you miss the very thing you most need to learn.
Step Four: Converge—gap mapping + 3 things you can start today
First dogap mapping: take the 🟢🟡 tier findings and, against "it has / you have," lay them out in a small table (what it has that you don't × each one's replicability × the cost to close the gap). Thenconverge into 3 things you can start today—not 30 to-dos, but 3, forcing yourself to prioritize.
Write each thing in four lines; miss one line and it doesn't count:
抄什么: 来自[哪个维度]的[哪个机制],它为什么可迁移(哪一档)
第一步: 今天就能做的第一个具体动作(具体到能立刻执行,不是"研究一下")
怎么验: 怎么知道这步有没有用(一个可观察的信号 / 数字 / 时限)
不抄什么: 这条相邻的诱人做法为什么是 🔴 不可迁移,提醒别顺手抄了
Hard constraint on the action list: nothing on it may depend for its success on 🔴 non-transferable factors. If some tempting action essentially relies on endowments (say, "get an exclusive qualification like it did"), either delete it or rewrite it as a long-term item to "close the endowment gap" (explicitly not something to start today), and don't let it sneak into the "3 things today" to manufacture an illusion.
If something lands on "speaking externally / public opinion / finished content" (say, "post a thread about my differentiation"), this skill only supplies the assembly skeleton and raw material; the flesh of the opinion is left to the author—hand this step over to ray-writer 's thread skeleton mode; don't ghostwrite.
Output structure
Produce the breakdown report in a fixed structure:
- One-sentence conclusion—is this object worth learning from, and what is the single thing most worth copying
- Five-filter results—pass / fail for each gate; for fails, state whether it's a downgrade or a discouragement
- Five-dimension dissection—the observable facts for each dimension (mark whether the data is "verifiable" or "unverified")
- Transferability verdict table—tag each finding 🟢🟡🔴; for 🔴, state plainly "why it can't be learned"
- Gap mapping + 3 things today—four-line format, each with its own "what not to copy"
- Honest notes—lay out in one go the places in this breakdown where the data is questionable, the sample may have survivorship bias, or the resource gap between you and me is too large
When not to use this skill
- You just want to get some account's / product'sraw data(follower growth curve, posting frequency, traffic numbers)—that's scraping and monitoring, use
/ray-metrics - What you want to diagnose isyour ownbusiness—where it's going wrong, whether you can take on a certain project—that's readiness diagnosis, use
/ray-consult - There is absolutely no empirical evidence for the subject data, and it relies entirely on self-declaration—don't dismantle it yet; dismantling it would amount to endorsing a hallucination; go back and find a verifiable subject or a verifiable data source.