Deposition: Shared Understanding for More Autonomous Work

With GPT-6 Astra, I want to delegate more work to the model. It can investigate code, try approaches, implement changes, and check its results. I want to spend my attention on purpose, consequential tradeoffs, and the quality of the deliverable, while leaving room for methods I would not have thought of.
Delegation first requires some understanding of whether we mean the same thing. A model can diligently complete a request, produce a substantial artifact, and pass its checks. Only when I read the result do I discover that it organized the work around a different purpose. Sometimes I have failed to express my intention. Sometimes I am still finding a clear way to understand it myself.
deposition is a skill I am developing for this collaboration. It asks the model to build a shared mental model with me: investigate the context, explain its understanding in its own words, contribute judgment I can examine and correct, and test that understanding against the intended use before the work depends on it. The starting point can be an incomplete idea, an existing plan, or an unsatisfactory result. A finished specification is not a prerequisite.
Where human attention becomes useful
OpenAI's skills guidance for GPT-6 Astra recommends reviewing accumulated instructions, reconsidering elaborate recipes, and clarifying decision boundaries and completion. It prompted me to think again about where my instructions should direct the model's attention, and how much of that direction a capable model actually needs.
In the second part of her pstack guide, Lauren calls this “The art of supervising someone smarter than you.” She describes two recurring difficulties: insufficiently expressed intent, and insufficient context for doing the work correctly. She wants to communicate the desired outcome while leaving the model freedom to discover how to achieve it.
Her most useful practice for me is asking the agent to restate the problem in its own words. The restatement compresses a noisy conversation into a structured account of the problem. It lets her catch a misunderstanding before any code is written. And because the agent has not simply adopted her assumptions and hypotheses, it can still reach a view she had not considered. Asking someone more capable to explain their understanding in terms you can follow is how you supervise them.
The two difficulties call for different work. A fact available in the repository calls for inspection. An uncertain behavior may need a small experiment. Purpose, priorities, and acceptable costs require human involvement. Turning every contextual gap into a question makes the user do unnecessary investigation. Treating every unclear intention as an implementation detail lets the model settle the direction too early.
What persistent questioning offers
The version of grill-me I first encountered was short. It asked the model to examine a plan relentlessly, work through dependencies between decisions, ask one question at a time, and supply a recommended answer for each question. Questions answerable by inspecting the code called for inspection. That early version takes very little time to read.
Its appeal was practical. I could start with an incomplete idea and discover omissions through the questions. Focusing on one question at a time also made complex work easier to discuss.
But I sometimes found myself accepting a recommendation, then accepting the next one. Each local choice received confirmation, without necessarily making the overall purpose any clearer to me.
Repeated agreement is not itself a failure. When a model has read the code and can explain a technical recommendation, I want to delegate that judgment. The concern is that the model may also have decided how to frame the question and which alternatives deserve consideration. If those choices depend on an unspoken goal, a sequence of reasonable answers can keep carrying that goal forward.
A recommendation answers a question. It does not show me how the model understands the problem the question belongs to. Deposition moves that understanding to the front, so I can respond to it before I start approving the choices that follow from it.
In the model's own words
A restatement can still be shallow. Suppose I ask for a first presentation on a collaborator's spatial transcriptomics data: basic quality control and a slide deck, no deep biological analysis. The model can reply, “You want basic QC and a slide deck, without deep biological analysis.” That is accurate, and it gives me nothing to examine.
A more useful reply takes a position: “I understand this meeting mainly needs to show the collaborators that we have received and understood their data. The QC metrics are evidence for that, not the organizing center of the talk.” It might be wrong. But it exposes a judgment I can react to. Is the deck organized around metrics, or around the collaborators' understanding of where their data stands? I do not need to know the right slide structure in advance to tell whether this is close to what I meant.
This is what I mean by a shared mental model. The model connects the intended use to the constraints, choices, consequences, and evidence that would establish success, and shows how those relationships shape its judgment. Goals, acceptance, verification, and boundaries all belong in that picture:
- Goals describe what the result is for: choosing the next experiment, or deciding whether the data justify further investment.
- Acceptance describes a result fit for that use, which is more than having the expected sections and figures.
- Verification supplies evidence for particular claims and should say what it leaves open.
- Boundaries establish where the model can decide and which choices come back for discussion.
In an earlier version of the skill, I listed these four as the things the model should explain. Listing them turned them into a form: the model filled in each field and stopped thinking about what else the task needed. A real task may turn on something none of them names. The better instruction is to build the connected picture and let the particular problem decide what needs to be seen.
Contributing, not only following
GPT-6 Astra follows instructions well. That is useful for execution and limiting for discovery. If the model only makes my stated requirements precise, it cannot help with what I have not yet said.
So I want the model to take intellectual initiative: surface assumptions and tacit judgments that affect the task, bring relevant knowledge, and propose a better framing when it changes what is worth doing. It should explain what the insight would change, ground it in evidence, and keep it distinct from my decisions. Initiative without that discipline turns into novelty for its own sake.
Tacit knowledge needs particular help. Two situations recur for me. Sometimes I have a judgment I have not managed to articulate: I know the slides are wrong but cannot yet say why. Here a candidate expression helps, something I can recognize or reject: “Yes, that is what I meant,” or “close, but the difference is here.” Other times the relevant knowledge is already mine but has not come to mind: I have read an argument that applies, but have not connected it to the current task. Here the model can propose the connection. In both cases, a contrasting example or a short sketch is often easier to respond to than an abstract question.
Neither case lets the model claim to know what I really want. It offers a grounded interpretation; my response decides whether it caught the thing. When a blind spot needs sustained exploration, I use a companion skill, known-unknowns, and bring what it surfaces back into the same shared understanding.
Testing fitness for purpose
A shared mental model can be internally consistent and still miss. The check I trust most is to walk through a plausible use of the proposed result, especially the part we understand least. Could the work satisfy every stated requirement and still fail the person using it?
In the presentation example, agreeing on a workflow of QC, then figures, then slides does not settle what the slides will say. Agreement on the division of work, or on broad categories, does not mean the actual content is aligned. Walking through what the collaborators would see on each slide, and what they should conclude, exposes whether we agree about the deliverable itself.
Writing gives another example. A conversation may already contain design choices, trial results, and references. A model can organize them in the order they arose and produce an accurate development record. A reader encountering the topic for the first time needs a different starting point: what difficulty they recognize, why it matters, what they could do afterward. A short synopsis and outline, with the audience and argument marked as provisional, let me see what would actually be written before the full draft depends on it.
The discussion can close when the shared understanding supports a useful next action and the remaining uncertainty is resolved, delegated, or explicitly bounded. Exploratory work does not require knowing the answer in advance; agreeing on the question and a point to review the findings is enough.
How understanding enables execution
Intellectual initiative is not execution authority. The model can offer a new framing without asking; turning that framing into changed files, a larger scope, or a new deliverable needs applicable authorization.
When I explicitly invoke deposition, I am asking for discussion before the dependent implementation. Once work is authorized, the model should carry it through without seeking permission for every implementation detail. Suppose a button's wording is settled while the page title remains open. The model can change the button and hold the title question. An unresolved issue pauses only the actions that depend on it.
Between those two points lies a small but persistent source of friction: what does “yes” mean? If the model asks whether its interpretation is right and I say yes, I have confirmed an understanding. If it proposes a concrete change within a task I already authorized and I say yes, it should proceed. In practice, models tended either to treat any agreement as permission to start, or to ask again for permission I had already given.
Two things help. The model should make clear whether it is checking an interpretation or proposing an action, and read my reply against that and against the authorization already in force. I can also use two optional signals: agree confirms the understanding without starting new work, and proceed authorizes the settled next action. They are shortcuts, not stages. Ordinary language such as “yes, make those changes” works too, and proceed does not resolve a scope that is still unclear.
During execution, the shared mental model keeps guiding choices. When new evidence changes a consequential premise, the relevant choices are revisited. When I clarify the goal partway through, earlier choices made under the old goal deserve another look, rather than just a new constraint appended to the old plan.
Apply the method to the skill itself
Deposition needs the same scrutiny. A failure should prompt an investigation into whether guidance was missing or ambiguous, whether the model ignored an existing instruction, or whether necessary context never reached the work. Adding a rule after every disappointing result gradually creates a burdensome process.
Investigating real sessions gave answers I would not have guessed. In one tool, the skill's text had been passed to the model truncated at 8,000 characters, so the rules about reusing authorization never arrived; a correctly installed skill is not necessarily a fully read one. In another session the model had the complete text and still asked twice, reading “pause to resolve a disagreement” as “start the approval over.” And in the presentation work, the model kept earlier figure choices after the purpose became clearer, polishing titles instead of reconsidering them. None of these called for more rules. The last one in particular was not about how much I had said; it was about whether the model kept using the understanding we had built.
Following the Astra guidance, I cut the skill to about a third of its former length. The cuts removed generic method tutorials, repeated completeness checklists, and a fixed two-round approval script. An independent review by another model agreed with most of the cuts but pointed out where they had removed real constraints rather than procedure, such as naming each deliverable without saying what it must contain. Those constraints went back in.
Whether the shorter version reduces friction in practice is not something a text review can establish. The evidence I rely on now comes from real use, and I continue to revise the method as I see where it drifts.
Deposition is a standalone skill in my skill collection. The skill instructions and source notes are public. This is the kind of request I want to be able to start with:
Use deposition to help me think this through. I am not yet sure what good would look like. Look into the context, tell me in your own words what you think this is for and what would make it useful, and point out what I might be assuming without saying. Then we will decide the next step.
What I want from the exchange is understanding sufficient for delegation. I keep a grasp of purpose and consequential tradeoffs; the model contributes its judgment to that understanding, then investigates, chooses methods, and completes the execution.
Deposition:与模型形成共同理解,把执行交给它

使用 GPT-6 Astra 时,我想把更多工作交给模型。它可以调查代码、尝试方案、完成实现,也能检查自己的结果。我希望把注意力放在目标、重要取舍和交付质量上,让模型有空间采用我没有想到的方法。
但把任务交出去,首先需要知道双方理解的是不是同一件事。模型可以认真完成一个请求,输出完整、检查通过,而我读到结果时才发现,它组织这项工作的出发点与我不同。有时是我没有把意图表达出来,有时则是我自己也还在寻找一个清楚的说法。
deposition 是我为这种协作发展的一项技能。它要求模型和我共同建立一个共享心智模型(shared mental model):调查相关情境,用自己的话说明它如何理解这件事,提出我可以审视和纠正的判断,并在工作依赖这份理解之前,用预期用途检验它。起点可以是一个模糊想法、一份已有方案,或一项让人不满意的结果。用户不需要先准备好完整需求。
模型能够执行之后,人需要关注什么
OpenAI 的面向 GPT-6 Astra 的技能指南建议重新检查积累下来的指令,警惕过细的操作流程,并明确决策边界和完成条件。这让我重新考虑:我的指令应该把模型的注意力引向哪里,一个能力很强的模型又究竟需要多少这样的引导。
Lauren 在 pstack 指南第二部分中,把这个问题称为“The art of supervising someone smarter than you”。她指出两类经常遇到的困难:意图表达不足,以及缺少正确完成工作所需的情境。她希望向模型讲清要达成什么,同时给它自由探索解决方法。
她的做法里,对我最有用的是让智能体用自己的话复述问题。这样做有三个好处:它迫使智能体把一段嘈杂的对话压缩成结构清楚的问题陈述;她能在任何代码写出之前发现误解;而且智能体没有直接接过她的假设和猜测,仍有可能得出她没有想到的判断。让一个比自己更有能力的人用你听得懂的方式讲出他的理解,正是监督他的方式。
两类困难需要不同的工作。仓库里查得到的事实,模型应该去查;某个行为是否成立,可以用适当的小实验检查。关于用途、优先级和可接受代价的问题,则需要人的参与。把缺少的资料全部变成问题,会让用户承担不必要的调查;把尚未明确的意图全部视为实现细节,又会让模型过早替用户决定方向。
持续追问为什么吸引我
我最初接触的 grill-me 很简短。它要求模型持续推敲计划,沿着决策之间的依赖逐步追问,每次一个问题,并为每个问题提供推荐答案;能够检查代码解决的问题,就先去检查代码。这个早期版本的原文可以直接读完。
它吸引我的地方很实际:我可以带着一个不完整的念头开始,借助提问发现自己遗漏了什么。一次集中处理一个问题,也让复杂任务更容易讨论。
但我遇到过一种体验:模型给出问题和推荐,我觉得有道理,就回答 yes;下一轮又是如此。局部选择都得到了确认,我却未必更清楚整件事要达到什么目的。
连续接受推荐本身没有问题。模型读过代码,能够解释一个技术选择的依据,我当然希望把这种判断委托给它。值得注意的是,模型可能同时决定了问题怎样表述、有哪些备选方案、哪一个值得推荐。若这些选择依赖一个尚未说清的目标,一连串同意也可能沿着那个目标继续推进。
推荐回答的是一个问题,却没有让我看到模型如何理解这个问题所属的整件事。Deposition 把这份理解放到前面,让我在开始批准由它推出的各项选择之前,先对它本身作出回应。
用模型自己的话
复述也可能很浅。假设我请模型为合作者的空间转录组数据准备第一次汇报:做基础质控(QC),出一份幻灯片,不做深入的生物学分析。模型可以回答:“你要基础 QC 和一份幻灯片,不做深入的生物学分析。”这句话准确,却没有给我任何可以检验的东西。
更有用的回答会表明立场:“我理解这次会议首先要让合作者看到,我们已经接住并理解了他们的数据。QC 指标是支撑这一点的证据,而不是汇报的组织中心。”它可能不对,但它暴露了一个我可以回应的判断:这份幻灯片是围绕指标组织,还是围绕合作者对数据现状的理解组织?我不需要事先知道正确的页面结构,也能判断它离我的意思有多远。
这就是我所说的共享心智模型。模型把预期用途与约束、选择、后果以及能够证明成功的证据联系起来,并说明这些关系怎样影响它的判断。目标、验收、验证和边界都在这幅图景之中:
- 目标说明结果要支持什么:选择下一项实验,还是判断数据是否值得继续投入。
- 验收说明什么样的结果适合这种用途,这不只是章节和图表齐全。
- 验证为具体主张提供证据,也应说明它留下了哪些空缺。
- 边界说明哪些选择由模型决定,哪些需要回来讨论。
在技能的早期版本里,我把这四项列为模型需要解释的内容。列出来以后,它们就变成了一张表:模型逐项填写,然后不再思考这项任务还需要什么。真实任务的关键,可能落在这四项都没有提到的地方。更好的指令,是让模型建立这幅相互关联的图景,由具体问题决定还需要看见什么。
贡献判断,而不只是执行指令
GPT-6 Astra 很擅长遵循指令。这对执行有利,对发现却是一种限制。如果模型只是把我已经说出的要求变得更精确,它就无法帮助我处理尚未说出的部分。
因此,我希望模型在思考上主动:指出影响任务的假设和隐性判断,带来相关知识,在一个更好的问题框架会改变工作价值时,把它提出来。它需要说明这个见解会改变什么,以证据为依据,并与我的决定区分开。缺少这种约束,主动就容易变成为新奇而新奇。
隐性知识(tacit knowledge)尤其需要帮助。我常遇到两种情况。一种是我已经有判断,却还没说出来:我知道幻灯片不对,却讲不清为什么。这时,一个候选表达会有帮助,让我能够辨认或否定:“对,就是这个意思”,或者“接近,但区别在这里”。另一种是相关知识本来就在我的经验里,却没有在此刻被想起:我读过一个适用的观点,却没有把它和眼前的任务联系起来。这时,模型可以提出这个联系。两种情况下,一个对比例子或一张小草图,往往比抽象的问题更容易回应。
这两种情况都不允许模型声称知道我真正想要什么。它提出有依据的解释,由我的回应判断它是否抓住了那层意思。当某个盲区需要持续探索时,我会使用配套技能 known-unknowns,再把发现带回同一份共同理解。
用真实用途检验理解
一个共享心智模型可以内部自洽,却仍然偏离目标。我最信任的检查,是沿着一种合理的使用方式走一遍拟定的结果,尤其是双方理解最少的部分。这项工作有没有可能满足所有明确要求,却仍然帮不上使用它的人?
在汇报的例子里,就“先 QC、再出图、最后做幻灯片”的流程达成一致,并不意味着幻灯片要讲什么已经确定。同意分工或宽泛的分类,不代表实际内容已经对齐。逐页设想合作者会看到什么、应当得出什么结论,才能暴露双方对交付物本身是否一致。
写作是另一个例子。对话里可能已经有设计取舍、试验结果和参考来源。模型可以按材料出现的顺序组织它们,写出一份准确的开发记录。而第一次接触这个主题的读者需要另一个起点:他们遇到什么困扰,为什么值得关心,读完能做什么。一段简短的大意和框架,把读者和论点标为暂定,能让我在全文依赖它之前,看到最终会写出什么。
当共同理解足以支持一个有用的下一步,剩余的不确定性已经解决、委托或明确划定范围时,讨论就可以结束。探索性工作不要求预先知道答案;约定要调查的问题和复核结果的时间点,就足够了。
共同理解怎样支持自主执行
思考上的主动不等于执行上的授权。模型可以不经询问就提出一个新框架;但把它变成修改文件、扩大范围或新增交付物,需要相应的授权。
当我显式调用 deposition,意思是在依赖讨论结果的实施之前先讨论。一旦工作获得授权,模型应当把它做完,不必为每个实现细节重新请求许可。假设按钮用词已经确定,页面标题仍有分歧,模型可以完成按钮修改,保留标题问题。未决事项只暂停依赖它的行动。
在这两者之间,有一个很小却反复出现的摩擦:“对”到底是什么意思?如果模型问它的理解对不对,我说对,我确认的是理解。如果模型在一项已经授权的任务中提出具体修改,我说对,它就应该动手。实际使用中,模型要么把任何同意都当成开工许可,要么为我已经给过的授权再问一遍。
有两点帮助。模型应当说清自己是在核对理解,还是在提出行动,并结合这一点和已有授权来理解我的回答。我也可以使用两个可选信号:agree 确认理解,不启动新的工作;proceed 授权执行已经确定的下一步。它们是快捷方式,不是必须依次经过的关卡。“对,就这样改”这类日常表达同样有效;而 proceed 也不能替尚不清楚的范围作决定。
执行过程中,共享心智模型继续指导各项选择。新证据改变了重要前提,就重新审视相关选择。如果我在中途把目标讲得更清楚,那些在旧目标下做出的选择也值得重新检查,而不是只在旧计划上追加一条新约束。
用同样的方法检查技能本身
Deposition 也需要接受这种审视。一次失败首先要求检查:规则是否缺失或含糊,模型是否忽略了已有指引,还是所需情境根本没有进入工作过程。每次偏离都追加一条规则,会逐渐积累出越来越重的流程。
回查真实对话,得到了一些我事先猜不到的答案。在一个工具里,技能正文传给模型时被截断在 8,000 个字符处,关于复用已有授权的规则根本没有送达;技能安装正确,不等于模型读到了全文。在另一次对话中,模型拿到了完整正文,仍然重复询问,把“暂停以解决分歧”理解成“从头重新审批”。而在准备汇报时,目标变得更清楚之后,模型仍沿用早先选定的图,只继续修改标题和措辞,没有重新考虑图本身。这些问题都不需要更多规则。尤其是最后一个,问题不在于我说得够不够多,而在于模型有没有持续使用我们已经建立的理解。
按照 Astra 指南的思路,我把技能删减到原来篇幅的三分之一左右。删掉的是通用的方法教程、重复的完整性清单,以及固定两轮审批的流程。另一个模型的独立审阅认可了大部分删减,但也指出有几处删掉的是真正的约束,而不是流程,比如只列出交付物的名称,却不说明它需要包含什么。这些约束又被补了回来。
更短的版本是否真的减少了摩擦,文本审阅无法证明。我现在依靠的证据来自实际使用,也会继续根据它在哪里偏离来修改这套方法。
Deposition 是我技能合集中的一项独立技能。技能正文和来源说明保存在公开仓库。开始一次讨论时,我希望能这样提出请求:
用 deposition 帮我把这件事想清楚。我还不确定怎样才算做好。请先了解相关情境,用你自己的话告诉我,你认为这件事是为了什么、怎样才有用,并指出我可能默认了却没有说出的前提。然后我们再决定下一步。
我希望最终得到的,是足以支撑委托的共同理解。我掌握工作的目的和重要取舍;模型把它的判断贡献给这份理解,再据此自主调查、选择方法并把执行做完。