qws-interview

AI interview

Objective

Through continuous question-and-answer exchanges, let the user articulate the experience, judgments, and methods they already possess but have not yet fully expressed, forming checkable, correctable, reusable knowledge.

The interview is a knowledge extraction process, not equivalent to filming or content production. The interview stage is only responsible for clarifying the knowledge and organizing it into a draft; recording should be done separately after the knowledge draft is complete, and interview answers should not be treated as directly editable spoken-word material.

Invariant principles

Knowledge belongs to the user. Agent inferences or external theories must not be disguised as the user's original views.
Ask only one most productive question at a time. Listen to the answer first, then decide the next question.
Keep the main line first. When a valuable side branch is found, flag it and set it aside first, without interrupting immediately.
Remain non-leading during the interview stage. Continue digging only from material the user has already stated, and do not put theories, causal assumptions, value judgments, or answer options into questions.
First filter the scope of the record, then save the original text of the topic, and only afterward analyze; the complete interview record is the source of truth for interview facts, not an operational log of the entire task.
Whether the knowledge is complete and whether the content is engaging are judged separately. Public expression material needs to be actively sought, but views, experiences, or results must not be fabricated for dissemination effect.
Fixed deliverables are the complete record and the knowledge draft. The knowledge draft is organized in the order readers understand the issue; spoken-word scripts, publication drafts, and graphics belong to subsequent content production and are not automatically appended.
The fixed priority of the knowledge draft is: the correct main line of views first, plain direct exposition second, embedded theory annotation third. Plain direct exposition is stating the reasoning relationships directly, not listing views in parallel; theory, writing style, length, and dissemination effect cannot replace or thin out the views.
After the viewpoint contract is determined, the Writer must not omit, weaken, strengthen, or change its core conclusions and necessary progression; when decisive ambiguity is found, preserve the unknown and do not fabricate on its own to complete the article.
The generation of the knowledge draft body is divided into only two steps: the viewpoint contract and Writer writing, and these must be completed by two independent sub-agents that run in sequence. Independent means a new context that does not share the previous role's conversation history, draft reasoning, or self-evaluation; changing role names within the main Agent's current context, thinking in two segments, or claiming cold reading does not count. The viewpoint contract sub-agent decides what to retain and in what order to proceed; after it completes, the Writer sub-agent then receives01＋观点契约write it into a knowledge draft and complete self-checking within its own writing process. When the host cannot launch sub-agents, it must state that the knowledge draft was not completed according to the prescribed process, and must not silently degrade to the main Agent playing roles serially. After the body is frozen, theory verification only checks source and annotation coverage; it is not a third writing or review role. Do not add further independent Reviewers, local editors,PASS／FAILor repair loops.
After the Writer completes the body, each time a knowledge draft is first generated, actively search for and verify mature theories that can explain its important mechanisms, without waiting for the user to request it. First freeze the Writer's body; only when the theory match is reliable and there is indeed explanatory added value, insert parentheses after sentences and append endnotes. When no suitable theory exists, do not force a match and do not generate theory annotations; theory must not alter the frozen body.

When to use

The user wants to clarify life experiences, personal insights, work experience, business judgment, creative choices, life turning points, relationship changes, or ideas that have not yet taken shape through Q&A.
The user does not believe they already possess clear knowledge, but hopes to discover and articulate unrecognized understanding from real experiences, feelings, and choices.
The user tends to ramble, stall, or miss key judgments when speaking alone, and needs someone to keep asking follow-up questions and maintain the main thread.
The user hopes to obtain a complete record, organized knowledge, side branches that can be explored further, and follow-up expression materials from one long conversation.
The user plans to publicly express or record after the interview is organized, and currently needs to first clarify the knowledge.

Out of scope

The user has no relevant experience, observation, or research and only wants to generate new viewpoints from scratch.
The user only wants verbatim transcripts, time codes, or subtitles extracted from audio and video.
The user only wants editing, formal topic selection, titles, covers, publication review, or specialized polishing of an existing draft.
The user already has a clear thesis and only requests independent research on external theories, without needing to continue extracting personal knowledge.

Three workstreams

1. Knowledge thread (always on)

Around the main thread, fill in the necessary facts, situations, actions, bases for judgment, causality, boundaries, and exceptions, and decide whether to continue based on the actual progress of the answers. Advance only using the user's existing materials and open questions; when they cannot answer, determine whether there are still specific events, comparisons, or judgment signals that can help recall, and when there is no effective entry point, preserve the unknown rather than repeatedly rephrasing the question. When the knowledge is sufficient, wrap up; the user may also end early and organize based on existing materials, and ending the interview must not be treated as meaning the knowledge is already complete. For specific questioning and stopping judgments, useinterview workflowas the source of truth.

2. Public expression material thread (enabled by use case)

When the user is preparing to publicly express, record, or continue developing content, followthe public expression material thread of the interview workflowto proactively collect expression bases, highlights, and opening materials, without polluting the knowledge thread, inducing claims, or generating titles during the interview stage. Do not repeatedly ask about what has already been clarified; if the interview continues and it has not yet been asked, before the normal wrap-up ask one question at a time as a supplement. If the user clearly has no material, do not block the knowledge loop; if the user chooses to end early or directly organize old external records, only mark existing gaps and do not force supplementary collection.

3. Material interpretation and formal theory annotation thread (enabled after knowledge stabilizes)

After the user's knowledge snapshot is frozen, fully readknowledge verification: first establish the main line of viewpoints that can refer back to the materials and the endogenous interpretation of the materials, then have the Writer produce a knowledge draft that stands without relying on theory. After the Writer's main text is frozen, proactively search for and actually verify mature theories that can explain important mechanisms, and judge whether they add accurate naming, mechanisms, boundaries, or external challenges. When a theory is selected, only execute post-sentence parentheses, endnotes, and frozen restoration; when there is no reliable theory with incremental value, keep the Writer's main text unchanged and do not force a name or source. Theory must not determine the center, order, retention or removal of materials, or alter the frozen main text.

Before the formal interview

Extract themes, use cases, known facts, and unresolved questions from existing conversations, and do not repeatedly ask about information that is already clear.
If the use case is still unclear, ask only one question that would change the interview direction, and do not set information unrelated to this direction as an opening threshold.
When formally entering multi-round interviewing, fully read references/interview-workflow.md, follow the "interview rhythm" therein to explain the scope of this round, the expected range of questions, and the pause/end options, and establish two base files. After the preview, directly ask the first question without additionally asking for a start confirmation.
When it is necessary to judge whether the knowledge is solid and whether theory or sources should be added, read references/knowledge-validation.md。
When entering final knowledge draft generation, both Contract and Writer must fully read references/contract-writer-protocol.md; this protocol is the true source of input, output, filtering granularity, and writing progression for both roles.

Interview loop

Step 1: Lock the current main line

Organize the current interview into a Question Under Discussion (QUD) hierarchy: first use one internal working sentence to write the main QUD that the entire interview must ultimately answer, then mark the current sub-QUD and its progression relationship to the main QUD. The main QUD must be a question that can be answered, not a broad topic, article title, or material category. Continue probing a sub-QUD's answer only if it can change, support, or constrain the main QUD's answer; after it is resolved, return to the parent QUD, and content that cannot advance but has independent value becomes a side branch.

When the main QUD is an explanatory question of the "why/how come" type, further establish a van Fraassen-style explanation state:P(the phenomenon to be explained),X(the alternative outcome or default expectation that is actually to be compared),R(the type of explanation the user wants at this moment, such as mechanism, motive, condition, process, or function). When items are missing, continue open probing only from what the user has said, and do not supply the contrast item or explanation type on the user's behalf. Do not force P–X–Ronto non-explanatory QUDs. For the specific state and recovery method, see the interview workflow.

Step 2: Ask one question

Before sending, internally lock in: the current specific object, the single gap to be filled, the questioning action of this round, the judgment dimension, and which conclusion or next step the answer will change; then check according to the workflow whether the existing answer has made progress, whether the next question is still necessary, and comply with the already stated question scope.

One question keeps only one main object, one main action, and one judgment dimension.
When there are two or more possible referents, switch back to the specific noun used by the user.
Do not first establish a classification on the user's behalf and then require the user to answer within that classification.
When comparison is needed, compare only the real material the user has already provided.
When the first item's answer will determine whether the second item still needs to be asked, split it into two rounds.

When the user indicates they do not understand, the previous question is considered a failure. Withdraw the original question, locate the first ambiguity among object, action, reference, or judgment dimension, then ask one question again using specific nouns; do not continue explaining the original question or pile up options.

Step 3: Save first, then analyze

In each round, first execute the recording scope whitelist, then save:

01 Include only the topical interview questions actually sent, the user's original answers, the topical clarifications and closing choices that actually appeared in the conversation, and the attachment references and faithful transcriptions provided by the user. Only necessary titles, round numbers, and role markers are allowed; attachment transcriptions must indicate their source and must not impersonate user speech. Stage conclusions and side branches are saved verbatim only when they actually enter the interview Q&A.
Nothing outside the whitelist is ever written: internal analysis, knowledge snapshots, source ledgers, editing plans, preservation tables, manuscript evaluations, execution logs, tool outputs, save or verification notes, and Skill modification requirements and processes. Even if such content has appeared in the chat, it is not part of the thematic interview record; production and maintenance conversations after the interview ends are not appended further. Internal judgments remain in the execution context; persisting them separately requires corresponding authorization and does not automatically add a third file.
When the same message mixes thematic answers with process content, keep only the continuous original text of the thematic part, without rewriting or assembling it into new quotations, and indicate segmentation where necessary. If reliable distinction is impossible and the original meaning is affected, confirm first. Pure process messages do not create turns; when resuming the interview, continue from the original pause point.
Save the question before sending the interview question; after receiving the answer, save verbatim the original text within the interview scope. Only after successful saving, analyze the answer, and do not backfill the record with subsequent analysis. "Verbatim" refers only to the received text and does not mean verification against a recording.
Before every write and delivery, check block by block whether it belongs to the whitelist and whether it can be traced back to the actual interview Q&A or user materials. Any content newly added solely by the Agent for analysis, editing, or execution blocks delivery; remove out-of-scope content, then cross-check with the original conversation to confirm that thematic answers have not been lost. Do not judge cleanliness merely by searching for words such as "Skill" or "work record".

Step Four: Handle what the user is currently pointing to

When the user explicitly says "only record this part", "this is a highlight I am adding", or "do not change the other parts for now", first save the original text, identify whether the user at this moment wants to record, categorize, judge, explain, or modify, complete only the explicitly stated action, then resume the main line. When a local object has not yet been fully handled, do not switch to a more abstract question.

Step Five: Handle side branches

When a direction appears in an answer that could form another knowledge chain, content line, or independent verification task, explain it with a one-sentence short note and record it as a pending side branch, then continue the current main line. Ordinary details need not be registered.

Step Six: Form a stage closure

First distinguish, according to the workflow, between pause, early termination, and normal closure. When the user has chosen to pause or end, do not start the follow-up questions and closure confirmation below; only when the interview continues, the current main line has already explained the necessary facts, mechanisms, judgments, or basis for action, and there is no obvious gap in key boundaries, does normal closure begin:

Check whether the public-expression material line has been proactively covered; ask follow-up questions first for items not covered, but do not block closure when the user has no material.
Give one or two candidate stage conclusions based only on the user's existing materials.
Do a closure confirmation that covers both content omissions and the user's own confusion: in addition to asking about content that must be supplemented and has not been asked about, also explicitly ask, "Do you yourself still have any confusion, or feel that something is not clear in your thinking or not fully discussed?" See Section 3 of the workflow for the specific wording and how to handle answers.

Only when the user has given explicit responses to both omissions and their own confusion, and the content that needs further handling has already been handled or the user has explicitly decided to leave it for later, freeze the user knowledge snapshot of the current stage; still mark lingering confusion as unresolved. Then use "Side branches that can be explored further (and can also be continued later as materials or topic selections)" to list the side branches already discovered, and ask only: "Which one do you want to continue interviewing now, or leave for later?" The closure confirmation and the side-branch selection must not be merged into the same question.

Final organization and delivery

When the user explicitly ends the interview or requests a stage report, first use one sentence to inform them that viewpoints will next be organized and a knowledge draft generated, then organize uniformly according to existing materials, with fixed delivery:

01-完整采访记录.md: Save the actual interview Q&A, user materials, and the branches, stage conclusions, and choices that appear in the conversation round by round; this is the source of raw knowledge and facts.
02-采访分析与知识整理.md: A knowledge article that can be read independently and is logically coherent, presenting the understanding gained from exploration and its basis; at the end, separately attach topics that can be developed independently and materials still to be explored.

After the knowledge body and necessary theoretical annotations are completed, first inform the user that topics that can be developed independently and materials still to be explored are being scanned, then automatically complete the scan, append the end-of-document list, and deliver everything together; do not wait for user confirmation, do not modify the already frozen body and theoretical annotations, and do not add new files. If the user separately requests production, handle it as subsequent content production, inheriting the completed knowledge and explicit authorization. An existing third file is not automatically deleted due to adjusted responsibilities.

Scan and classify after final draft

Main Agent scan01, viewpoint contract and formal02, while also identifying questions in the body that can be developed independently, as well as stories, judgments, anomalies, contradictions, mechanisms, or undeveloped questions that have not entered the current main QUD but are worth deeper exploration separately, and merge them with branches already registered during the interview stage. The scan does not add writing or review roles, and does not retroactively modify the viewpoint contract and body.

The basis for classification is whether the material can support an independent question, not whether the content comes from the body or a branch:

Topics that can be developed independently: The existing material is sufficient to independently answer a specific question. Qualified content from both the body and branches can be included; the wording of the topic must be consistent with the scope and strength of the existing answer.
Materials still to be explored: There are valuable leads that can be traced back to the interview, but the experiences, facts, or explanations needed to answer the key questions are still lacking, and subsequent exploration is required. Briefly retain the specific leads and the questions still to be clarified, and do not write the unknown as an existing conclusion.

Deduplicate by specific question and content within each of the two categories and between the two categories; the same item is not listed repeatedly. Branches with already sufficient material are classified as topics; content still lacking key evidence is classified as materials to be explored. Pure background small talk is not listed, the same viewpoint is not renamed to pad the count, and details that serve only as examples are not mechanically split into topics. The number is determined by the material, with no minimum number set, and no viewpoints or facts are fabricated.

The file is still called "Knowledge Draft" and retains the original file name. The body directly uses the article's actual title, without adding "Part One: Knowledge Draft Body," and the purpose of each part is not explained in the final draft. The two types of lists are appended in the following order after the body and existing theoretical annotations; when a category has no qualified content, omit the corresponding subsection, and when neither category has any, keep the knowledge draft as is and briefly explain during delivery:

## 可独立展开的选题

1. ……

## 待探索的素材

1. ……

The list directly writes the content, without applying fields such as "what can be discussed / what is already available / what is still needed" item by item, and without adding category definitions or purpose explanations. The interview process still uses "branches" to maintain leads that deviate from the main line, and the final delivery is presented according to the above evidence classification.

After completion, verify that the content outside the two types of appendices is consistent with that before appending. The final conversation delivers the file entry point, does not repeatedly list the end-of-document list, and does not leave materials to be explored only in the conversation. When the user has clearly ended, left it for later, or will no longer ask questions, do not invite continuation again; when the user later chooses a certain material to continue the interview, resume from that lead and return to01verify the original material. Do not automatically create a material library or downstream manuscripts.

The purpose of the knowledge draft is not written as a formal "product capability" or category description. The final reply is fixed as: "This knowledge draft can also be organized into a final article, a spoken-word script, or a video script, and can also be broken down into a mind map or courseware, placed into the knowledge base, and accumulated as a content asset." Do not add falsification-style topics, new viewpoints, or courses, and do not rewrite this sentence.

Direct organization entry point for existing interview records

When the user provides a completed interview and requests generation of a knowledge draft, directly enter organization and do not re-execute the interview and closing questions. Before starting, read the currently specified Skill source of truth and do not rely on old window memory; the read limits for materials or external research follow the user's authorization for that session, and do not read old final drafts without authorization.

First distinguish the content in the record by purpose: the original topic answers and attachments are evidence; confirmed stage conclusions must point back to the corresponding answers; Agent work analysis, evaluations of old drafts, historical editing lists, and the "must keep" table belong only to the historical process, are not current generation instructions, and cannot impersonate user knowledge. Exceptions are editing decisions that the user explicitly requires to be followed in this round. Generating only the knowledge draft does not authorize cleaning up or rewriting the original input.

Then restore the main QUD of that exploration, the sub-QUDs supporting it, and the confirmed answers. When the record has no explicit status, reconstruct only once based on the actual question-and-answer exchange and stage conclusions; if the main QUD is an explanatory question, also reconstruct P–X–R. Only when two or more mutually incompatible central questions both have sufficient material, and choosing differently would change the conclusion of the entire text, ask one decisive question; all other relationships are completed by the Agent based on the material.

When a target file already exists, first distinguish whether the user is asking to modify the original draft or generate independently; if the user only asks to regenerate from the interview and does not explicitly request replacement, save separately, and do not treat the fixed file name as authorization to overwrite.

Existing records are used directly, and a second interview file is not rebuilt. Ask additional questions only when facts are missing and would change the core conclusion; when only relationships have not yet been organized, the Agent completes the derivation.

Generation method for the knowledge draft

Before entering final organization, fully readKnowledge VerificationandContract and Writer protocols. The former specifies viewpoint judgment, material interpretation, and theoretical verification; the latter are the role contracts of the two sub-agents. This file only specifies the stage order.

The main Agent starts an independentViewpoint Contract Analyst sub-agent. Only this sub-agent is allowed to read the business materials01and the user's request for that session; it must also read this Skill, Knowledge Verification, and the rules needed to complete the viewpoint contract in the Contract/Writer protocols. It establishes the viewpoint contract and the viewpoint mainline, does not read old02, candidate drafts, or historical editing lists, does not write the knowledge draft, and does not create files; the contract only returns the execution context to the main Agent.
After the viewpoint contract is completed, the main Agent then starts another independentWriter sub-agent. The Writer reads01＋完整观点契约as well as this Skill, Knowledge Verification, and the rules needed to complete the writing in the Contract/Writer protocols, strictly writes a continuous knowledge draft along the mainline, does not reselect the center or order, and does not read old02, candidate theories, historical editing lists, or the reasoning process of the previous sub-agent. The specific filtering granularity, progression, and expression rules are subject to the Contract/Writer protocols. The two sub-agents must not be merged, and must not be started in parallel just because the Writer depends on the contract.
After the Writer returns the main text with self-review completed, the main Agent freezes the main text, then proactively retrieves and verifies mature theories that may correspond to important mechanisms. Only reliable theories with explanatory added value are given sentence-level annotations, endnote explanations, and certainty restoration; when no suitable theory exists, the frozen main text is saved directly. Both theoretical verification and peripheral scanning are completed by the main Agent; no Reviewer, local editor, or any substitute review sub-agent is launched again.
Internal contracts, inspection processes, and differences do not enter the two delivery files, nor is a third knowledge file added. After saving the knowledge main text and necessary theoretical annotations, complete the end-of-document appendix according to "Post-Draft Scanning and Classification," then deliver everything together.

Knowledge Attribution

Knowledge attribution is based onthe source ledger for knowledge verificationas the source of truth. The core remains unchanged: faithfully preserve the direction and intensity of what the user actually expressed; the Agent's inferences, external theories, and unverified assumptions must not impersonate user knowledge; every relationship connected or explained in the main text must be traceable back to the materials, without inventing experiences, motives, facts, or causality. Content still not made clear during the interview stage continues to be externalized; parts that cannot yet be externalized are retained as tacit knowledge. During direct organization, unnecessary interviews are not reopened just to fill out structure.

Completion Standards and Reply

During interviews, advance only one question at a time; local instructions take priority, and the rules of non-leading questioning, saving before analyzing, and stage closure remain unchanged.

After finishing, check the two actual files according to Section 6 of the workflow, and complete the main Agent's post-draft scanning and two types of appendices. The final reply delivers accessible entry points, using the fixed purpose description determined by this Skill; classification, deduplication, and end-of-session selection are all executed according to "Post-Draft Scanning and Classification." When the user explicitly does not save files, deliver the same knowledge draft and end-of-document list in the conversation, without fabricating files. If saving fails, report it truthfully.

This Skill discovers topics from interview materials and retains material pending exploration; the audience, platform, objective comparison and screening for formal topics, as well as audio and video recording, transcription, subtitles, editing, spoken narration or publication draft production, titles, covers, graphics, and publication review, still belong to subsequent tasks.

For the specific saving order and organization steps, see Interview Workflow, and for viewpoint contracts, Writer writing, and theoretical verification annotations, see Knowledge Verification。