Build Strategic Docs Through Guided Q&A Sessions

Builds durable strategic docs interview-style: the AI never proposes content - it asks one question, the user answers, the AI patches the file, then asks the next question. The file itself is the conversation's output, updated incrementally rather than drafted and handed over.

Use it when building a new single-source-of-truth file (life priorities, life vision, principles, frameworks, ranked lists), filling out a structured doc the user explicitly wants to author themselves, or running quarterly/annual reviews where the user's own words go into the file. It is NOT for day planning (use day-plan instead), task triage (use organize-tasks), or anything where the AI proposes content first.

The loop: create the file with a skeleton (header, sections, "to be filled in" placeholders) using a single write_file call - after that, never overwrite, only patch; ask exactly one concise, specific, open-ended question; wait for the answer without queuing the next question; patch the file with the user's answer in the correct section; then re-ask, either the next question or a follow-up if the answer was incomplete, repeating until the file is complete. Hard rules govern every step: one question at a time, never multiple in a single message; patch, never overwrite, after the initial skeleton; update the file before asking the next question, in that order; treat any list the user gives in response to "which X should we cover?" as an unordered set, never inferring rank or sequence from the order typed - ask explicitly ("Which of these is #1?") if ranking is needed; ask about a referenced person's role or dynamic rather than "who is X?"; keep questions and acknowledgments free of snark, attitude, or filler; and never add speculative sections, edge cases, or "anything else?" prompts unless the user asks. For ranked lists specifically, each rank gets patched in one at a time as the user confirms it, with the user's own words preserved in the entry.

Question design favors domain-discovery over confirmation (asking what wins against everything else, not confirming a guess), surfaces information the AI doesn't already have, uses engine-move framing where applicable ("what's the thing that, if true, makes the rest obvious?"), and prefers concrete framing over abstract ("what's #2, the domain that wins against everything except #1?" over "tell me about your second priority"). The file-patching pattern after each answer: read the relevant section if not already in context, patch with the placeholder or previous entry as old_string and the updated content as new_string, then confirm the diff before moving on. Named pitfalls to avoid: assuming order from an unordered set and writing it as a numbered ranking without confirming; bundling even two questions into one message; overwriting the file instead of patching a specific section, which destroys prior content; filling empty sections with AI-generated content instead of leaving them empty until the user provides it; and skipping the file update between question-and-answer pairs, which lets the doc fall out of sync. It pairs with, but is distinct from, other skills: day-plan uses a different, non-interview task-triage pattern; organize-tasks is Todoist-specific; and memory-management is separate, handling persona and preferences rather than file content.

Users who want to author their own strategic or reflective documents in their own words, one confirmed answer at a time, rather than have an AI draft the content for them to edit afterward.