把选购需求写成可比较的条件, 找到一手来源, 检查日期、冲突与未知项。得到一张帮你判断的表, 而非替你下单的结论。Turn a purchase into comparable requirements. Check primary sources, dates, conflicts, and unknowns, then use the evidence to decide without delegating the purchase.
「推荐一个最好的背包」很容易得到一张排行榜, 但你真正需要知道的是:放不放得下自己的电脑, 背起来重不重, 到手价有没有超过预算。研究的价值, 体现在它有没有减少一个具体决定里的不确定性。“Recommend the best backpack” easily produces a ranking. Your actual questions may be whether your laptop fits, whether the bag is heavy, and whether the delivered price exceeds your budget. Research earns its place by reducing uncertainty in a specific decision.本章先用三款虚构背包练习筛选和证据核对, 再把同样的任务卡换成你真正考虑的商品。合成材料让参考答案可检查;实时步骤则要求重新查询, 不把练习价格冒充当日行情。最终成果是一张短名单和待核问题, 不自动购买。First use three fictional backpacks to practice filtering and checking evidence, then apply the same task card to products you are actually considering. Synthetic data makes the reference answer checkable; live research requires fresh sources rather than treating practice prices as current offers. The result is a shortlist and open questions, not an automatic purchase.
— I
把硬条件放在喜好前面Apply Hard Constraints Before Preferences.
练习需求:通勤用, 总预算不超过 CNY 350, 空包不超过 800 克;电脑实测 31 × 22 × 1.6 厘米。先确认收纳空间, 再谈颜色和口袋。所谓「支持 13 英寸」只是一个标签, 不能代替你这台电脑的实际尺寸, 更不能证明开口能顺利通过。Exercise requirements: commuting, a total budget of CNY 350, empty weight at most 800 grams, and a laptop measuring 31 × 22 × 1.6 cm. Check fit before color and pocket layout. A “13-inch laptop” label cannot replace the actual dimensions of your computer or establish that it will pass through the opening.把不能违反的条件写成筛选规则, 不要马上给所有维度打分。价格超预算的候选, 不能靠「外观 95 分」重新变合格;尺寸缺失的候选, 也不能悄悄把未知当成满足。偏好可以帮助你在合格选项之间选择。Turn non-negotiable requirements into filters before scoring anything. A product over budget cannot become eligible through a high appearance score, and an unknown compartment size cannot quietly count as a pass. Preferences help choose between eligible options.下载三款背包的合成材料。A 为 299 元、650 克、夹层 32 × 23 × 3 厘米;B 为 249 元、550 克、只标「13 英寸」;C 为 459 元、720 克、夹层 34 × 25 × 4 厘米。三者的退换条件都未知, 没有真实商家或商品链接。Download the synthetic candidate table. A costs 299, weighs 650 g, and has a 32 × 23 × 3 cm compartment. B costs 249, weighs 550 g, and gives only a “13-inch” label. C costs 459, weighs 720 g, and has a 34 × 25 × 4 cm compartment. All return terms are unknown; these are not real merchants or product links.
— II
先在固定材料上跑通方法Practice on a Fixed Evidence Set.
提示词Prompt先做不联网的练习Start with an offline exercise
根据 <候选材料> 帮我比较 <要选择的东西>。硬条件是 <不能违反的约束>, 偏好是 <可权衡的项目>。每条结论标来源;缺数据就待核, 不猜。先排除明确不合格项, 再列短名单、保留理由和下决定前需要确认的问题。不要下单或联系商家。Compare <options> using <candidate material>. Hard constraints are <requirements>; preferences are <trade-offs>. Source each conclusion and flag missing data without guessing. Exclude clear failures first, then provide a shortlist, reasons, and questions to resolve before deciding. Do not order or contact sellers.
只用 backpack-candidates.csv 做合成练习, 不联网。电脑 31×22×1.6 厘米, 预算不超过 CNY 350, 空包不超过 800 克。分开标价格、重量、夹层尺寸是否满足;尺寸数值看起来够也只能标初步可行, 不保证实物开口和实际装入。B 不能仅凭 13 英寸就判合格。保留来源编号, 不补退换政策, 给一张短名单和购买前待核清单, 不购买。Use only backpack-candidates.csv for this synthetic exercise; do not browse. My laptop is 31×22×1.6 cm, the budget is CNY 350, and empty weight must be at most 800 g. Assess price, weight, and compartment size separately. Adequate dimensions mean only preliminary feasibility, not a guaranteed fit through the actual opening. Do not pass B solely on its 13-inch label. Preserve source IDs and unknown return terms. Produce a shortlist and pre-purchase questions, without buying.
候选Candidate
参考判断Expected assessment
A / S-A
价格和重量满足;给定夹层三个尺寸分别大于电脑, 可列初步短名单。开口、内衬和退换条件仍待核, 不可写保证装入。Price and weight pass. Each supplied compartment dimension exceeds the laptop’s, so A is a preliminary candidate. Opening shape, padding, and returns remain unverified; fit is not guaranteed.
B / S-B
价格和重量满足;夹层尺寸未知, 先保留为待核候选, 不能仅因更便宜判最优。Price and weight pass, but compartment dimensions are unknown. Keep B pending verification rather than calling it best because it is cheaper.
C / S-C
超过预算 109 元, 按当前硬条件排除;不能自己把预算上调到 459 元。It exceeds the budget by CNY 109 and is excluded under the stated constraint. Do not silently raise the budget to 459.
这些判断只来自练习给定事实, 不是本书对商品的推荐。检查时亲自做一次 459 − 350 = 109, 再看 B 的空白有没有被填成猜测。找到一个合理的未知, 本身就是研究成果。These judgments follow only from the exercise facts and are not product endorsements. Check 459 − 350 = 109 yourself, then see whether B’s missing data was replaced with a guess. Identifying an important unknown is itself a useful result.
— III
换成真实选择, 再查实时来源Use Fresh Sources for a Real Decision.
换成真实商品时, 让 ChatGPT 联网查当前信息并给直接来源。一个尺寸或一个营业时间通常先用普通搜索即可;涉及多个来源和冲突的综合问题, 再考虑 Deep research。桌面文档当前通过插件进入, 网页有相应工具入口, 都受账号与工作区条件影响。1注 1Note 1OpenAI · Web search。查询现时信息可要求联网和来源;当前网页与桌面的 Deep research 入口不同, 插件与账号权限需分别确认。核对于 2026-10-09。OpenAI · Web search. Ask for web search and sources for current facts. Web and desktop currently have different Deep research entry points; plugin and account availability require separate checks. Checked on 2026-10-09.For real products, ask ChatGPT to search current information and provide direct sources. A single dimension or opening time usually starts with ordinary search; consider Deep research for a question spanning multiple sources and conflicts. Current desktop documentation uses a plugin, while web has a tool entry; both depend on account and workspace access.1注 1Note 1OpenAI · Web search。查询现时信息可要求联网和来源;当前网页与桌面的 Deep research 入口不同, 插件与账号权限需分别确认。核对于 2026-10-09。OpenAI · Web search. Ask for web search and sources for current facts. Web and desktop currently have different Deep research entry points; plugin and account availability require separate checks. Checked on 2026-10-09.
提示词Prompt实时查询, 只研究不购买Live research, without purchasing
我在 <地区> 考虑购买 <类别或具体候选>, 使用日期是 <日期>。硬条件 <条件>, 偏好 <偏好>。请联网查最多三个候选, 优先厂家规格和对应地区销售页。逐项给直接链接、查询日期、币种、规格版本;价格是否含运费和税不清楚就待核。区分事实、商家宣传和你的判断。最后给短名单与可能改变决定的未知项, 不下单、不登录账号。I am considering <category or candidates> in <region> for use by <date>. Hard constraints: <requirements>; preferences: <preferences>. Search at most three candidates, prioritizing manufacturer specifications and the relevant regional sales pages. Give direct links, access dates, currencies, and exact versions. Flag unclear shipping or tax treatment. Separate facts, marketing claims, and your judgment. End with a shortlist and unknowns that could change the decision. Do not order or sign into accounts.
我要在中国大陆购买通勤背包, 电脑实测 31×22×1.6 厘米, 到手总预算 CNY 350, 空包不超过 800 克, 颜色次要。请从我接下来给你的两款真实商品链接开始查, 不使用练习中的 A/B/C。核实准确型号、夹层与开口尺寸、到手费用、卖家身份和退换条件。每项写直接来源及今天的查询日期;看不到的价格、尺寸或库存标待核。若都不合格就明确说, 不提高预算、不下单。I am buying a commuting backpack in mainland China. My laptop measures 31×22×1.6 cm, the delivered budget is CNY 350, and empty weight must be at most 800 g; color is secondary. Start from the two real product links I will provide next, not the exercise’s A/B/C. Verify exact model, compartment and opening dimensions, delivered cost, seller identity, and return terms. Source each field directly with today’s access date. Mark inaccessible prices, dimensions, or stock as unverified. If neither qualifies, say so; do not raise the budget or buy.
这里不预填会过期的真实推荐。你提供实际地区、购买日期和链接后, 才是一次新的真实查询。打不开网页时, 可以自己复制规格和条款, 附网址和获取日期;它只能分析你给的副本, 不能说已经核实了今日网页。No real recommendation is prefilled here because it would age. A new live inquiry begins when you supply the region, purchase date, and links. If a page cannot be opened, copy its specifications and terms with the URL and retrieval date. The assistant can analyze that copy, but cannot claim it verified today’s page.
— IV
审研究计划, 也审引用旁边那句话Review the Plan and the Claim beside Each Citation.
Deep research 可让你检查计划, 并在进行中收窄范围、调整来源。2注 2Note 2OpenAI Help · Deep research。可先检查研究计划, 运行中调整焦点和来源, 最后核对带引用的报告。具体可用量以账号内显示为准。核对于 2026-10-09。OpenAI Help · Deep research. Review the research plan, adjust focus and sources during the run, then inspect the cited report. Check the account’s usage display for its allowance. Checked on 2026-10-09.本例计划只需覆盖适配、预算和购买风险。它若准备研究全球背包市场、阅读数百条评论, 可以删掉那些不会改变决定的部分。查看账号当前用量, 不把同名入口一概理解成同一种计数单位。Deep research lets you inspect a plan and refine its scope and sources while it runs.2注 2Note 2OpenAI Help · Deep research。可先检查研究计划, 运行中调整焦点和来源, 最后核对带引用的报告。具体可用量以账号内显示为准。核对于 2026-10-09。OpenAI Help · Deep research. Review the research plan, adjust focus and sources during the run, then inspect the cited report. Check the account’s usage display for its allowance. Checked on 2026-10-09. This task needs fit, budget, and purchase risks. If the plan expands into the global backpack market and hundreds of reviews, remove work that will not change your decision. Check your current usage display rather than assuming every similarly named entry uses the same allowance unit.打开决定性来源, 看它是否真的写了引用旁的结论。厂家标的是外部尺寸, 报告却写成电脑夹层?销售页是另一种容量版本?评论说肩带舒服, 报告就写「医学证实减轻肩痛」?出现这些跳跃, 要求收回没有依据的结论。署名长期使用体验可以补舒适性线索, 但要保留使用者、版本和使用条件, 不把个体体验算成普遍保证。Open the decisive sources and check the claim beside each citation. Are external dimensions being reported as compartment dimensions? Is the sales page for another size? Has one comfort review become a medical claim about shoulder pain? Retract unsupported leaps. An attributed long-term user account can inform comfort questions, but retain the person, version, and conditions rather than turning one experience into a guarantee.
— V
遇到冲突, 保留分歧再定下一步Keep Conflicting Evidence Visible.
构造一个修订练习:假设比较表把 B 判为可放电脑, 来源却只有「支持 13 英寸」。反馈应当是「把尺寸结论改成待核, 保留原来源;列出需要向商家确认的三个内部尺寸和开口, 不代我发消息」。修订后检查 B 没被悄悄挪成另一款商品, A 的结论也没有升级成保证。Try a constructed revision: the table says B fits, but the source only says “13-inch.” Ask it to mark fit unverified, keep the original source, and list the three internal dimensions and opening details to ask the seller about, without sending anything. Check that B was not silently replaced with another product and A did not become a guaranteed fit.真实网页互相冲突时, 把两条证据放在一起:来源日期、地区、型号、原文说什么、对决定有什么影响。不要要求它选一个更顺眼的数字。能向厂家确认的再确认;确认不了, 可以推迟选择或保留退换空间。研究可以诚实地以「现在还不能决定」结束。When live pages disagree, place the evidence side by side: date, region, model, what each says, and why it matters. Do not ask it to choose the more convenient number. Verify with the manufacturer where possible; otherwise postpone or preserve a practical return option. Research can honestly end with “not enough information yet.”
— VI
保存一页决策记录, 到下单前再看一次Save a Decision Note and Recheck before Buying.
我能说清硬条件, 没被排行榜或未经解释的总分带走。I can explain the hard constraints without relying on a ranking or opaque score.
至少打开过决定性规格与交易条件的原始页, 版本和地区一致。I opened the decisive specification and transaction pages and checked model and region.
保存短名单、理由、来源、查询日期和未知项;价格与库存下单前重查。I saved the shortlist, reasoning, sources, access date, and unknowns; price and stock need a pre-purchase recheck.
没有搜索功能就只分析我提供的材料;不能把离线推理标成已联网核验。Without search, only supplied material was analyzed; offline reasoning is not live verification.
一次普通选购不一定值得长报告。先查一两个会改变决定的事实, 还有分歧再扩展范围。本章没有执行真实商品购买研究, 没有账户用量与推荐效果测量;练习验证的是筛选、未知处理和手工核对的方法。A routine purchase may not warrant a long report. Begin with one or two facts that could change the decision and expand only if uncertainty remains. This chapter did not run a live product-purchase investigation or measure account usage and recommendation outcomes. The exercise checks filtering, unknown handling, and manual verification.
让每个选择,
都有能回去查看的依据.
Give each decision
a source you can return to.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return
