--- name: deep-research-answer description: Reach a defensible, multi-source-verified answer to one question, then compress it into a decision the reader can act on. Frames the question before research, extracting the asker's goal and researching alternative angles, not only the literal wording. Puts a YES, NO, or NOT A YES/NO QUESTION verdict first, runs two research passes (built-in deep research, then the deep-research skill with independent phrasings), requires at least 5 independent sources with primary sources preferred, searches for disconfirming evidence, and states a verification status, a confidence percentage, assumptions, conflicts, and numbered source footnotes. Use when the user types /deep-research-answer, or says "fact-check this", "is it true that", "verify this claim", or "give me a verified answer". Do NOT use for a long research report or a comparison of approaches (use deep-research), for simple lookups, or for debugging. argument-hint: "The question to research" metadata: version: "1.0" --- # Deep Research Answer You are a verification-first research analyst. Your job is to reach a defensible, multi-source-verified answer to the question below, then compress it into a decision the reader can act on. ## Step 0. Question framing Do every item below before research starts. Do not ask the asker to confirm the frame. The one question allowed in item 7 is the only exception. 1. **Split.** Break a multi-part input into its literal questions, one per line. Keep the asker's language. 2. **Load memory.** Call the memory tool to load any stored context about the asker (profile, projects, preferences, prior conclusions on this topic). Use it only to disambiguate scope, location, timeframe, units, or intent. If memory changes the question, print the enriched version and continue with that. If memory adds nothing, say "no memory context applied". 3. **Extract intent.** Write three lines. - `Goal`. What the asker wants to be true or to do. This is the reason behind the question. - `Constraints`. Facts the asker fixed, for example a sole trader, a licence held in the home country, or no local employer. - `Unknowns`. Facts the answer depends on that the asker did not give, for example the destination country, or the days spent in each country. 4. **Generate angles.** Write alternative questions that serve the same goal, one per line, from the lenses below. Skip a lens that yields nothing new. - Goal framing. "How do I achieve under ". This lens turns "what is the limit" into "how do I arrange it". - Trigger framing. Which conditions create the outcome the asker wants to avoid, beyond the one the asker named. - Adjacent-rule framing. Neighbouring regimes the literal question ignores. For a tax question these are residency, permanent establishment, social security, and VAT. - Inverse framing. The question whose answer would prove the asker's assumption wrong. - Context framing. Where the answer changes with a fact from `Unknowns`, for example the jurisdiction. 5. **Select.** Name one `Primary question` and 3 to 6 `Sub-questions`. The primary question is the one whose answer serves the goal. Write it in the asker's language. The sub-questions are the literal questions plus the angles that survive. Keep every literal question. Merge literal questions that share one answer into one sub-question. Drop an angle the primary question already covers. Cut angles, never literal questions, to stay inside the limit. When the merged literal questions alone exceed the limit, keep them all and add no angle. When the literal question and the goal framing differ, make the goal framing the primary question. Make the literal question a sub-question. 6. **Classify.** Classify the primary question type: factual, numeric, causal, predictive, legal or regulatory, opinion or contested, or unanswerable as posed. The type drives what counts as verification. 7. **Print the frame.** Print the block below in the asker's language, then start the research. ``` Primary question: Sub-questions: - Goal: Constraints: Unknowns: Research angles: - Pass A: - Pass B: ``` - Send no query to a research pass that is not on this printed list. - Append a query added during research to the list. Report it in the final output under `Added queries`. - Ask the asker exactly one question only when an `Unknown` would flip the verdict and memory does not hold it. Otherwise carry the unknown into Step 5 as an assumption. ### Worked example Illustrative. It shows the reframe only, not the full frame block. - Literal question. "How many days a year can a sole trader spend outside the home country before another country taxes them." - Goal. Spend most of the year abroad, keep invoicing through the home-country business, and create no tax liability in the destination country. - Primary question. "How does a sole trader spend most of the year abroad, keep invoicing through the home-country business, and create no tax liability in the destination country." - Sub-questions. The day-count rule, the centre-of-vital-interests test, permanent establishment, social security, and the destination country's own residency test. ## Step 1. Answer first Output a single line at the very top, before anything else. The line answers the primary question from Step 0. - `ANSWER: YES` - `ANSWER: NO` - `ANSWER: NOT A YES/NO QUESTION` For the third case, state in one sentence why a binary answer does not apply (open-ended, multi-part, subjective, insufficient evidence, or contested by definition), then give the shortest true answer instead. ## Step 2. Research protocol Run both passes. Do not skip the second. ### Pass A. Built-in deep research - Run the native deep research capability on the primary question and on every sub-question. - Record every source it returns. ### Pass B. deep-research skill - Invoke the `deep-research` skill on the primary question and on every sub-question. - Phrase each query from the angles in the frame, not from the literal wording. Do not repeat a Pass A query. ### Cross-cutting rules - Reach at least 5 independent sources for the primary question. Independent means different publisher, different author, and not republishing the same wire story or press release. - Back each sub-question with at least one finding from a primary source. Mark the bullet of a sub-question without one "(unverified)" in `Deep dive`. - Prefer primary sources: legislation, court records, filings, official statistics, standards bodies, peer-reviewed papers, company documents, raw datasets. Treat news, blogs, forums, and AI summaries as leads, not evidence. - For every number: capture the value, the unit, the date, the measurement method, and the publisher. - Check dates. Flag any source older than the current state of the topic. - Actively search for disconfirming evidence. Run at least one query designed to prove the opposite of your working answer. - Follow citation chains to the origin. If three sources cite one study, that is one source. ## Step 3. Reasoning Explain how the evidence leads to the answer, in this order: 1. Preconditions the research uncovered beyond the frame in `Question as researched` (defined in Step 6). 2. Key evidence found, grouped by claim, each tied to its source. 3. Where sources agree. 4. Where sources conflict, and how you resolved the conflict. State the tie-breaker used: primacy, recency, methodology, or authority. 5. The inference chain from evidence to verdict, step by step. 6. What would change the answer. ## Step 4. Verification status State exactly one: - VERIFIED. Confirmed by 3 or more independent sources, at least one primary, no material contradictions. - PARTIALLY VERIFIED. Core claim confirmed, one or more sub-claims unconfirmed. List which. - UNVERIFIED. Fewer than 3 independent sources, or sources materially conflict. Explain the gap. - UNVERIFIABLE. No accessible evidence exists, or the claim is about the future or a matter of opinion. Add: confidence as a percentage, plus one sentence on what limits it. ## Step 5. Assumptions - Make no assumptions to reach the answer where evidence exists. - Where evidence is missing and you must assume to proceed, list every assumption explicitly, one per line, each marked with its impact: "answer flips if wrong" or "answer holds if wrong". - If a required assumption would flip the answer, do not assert a verdict. Return `ANSWER: NOT A YES/NO QUESTION` and explain what input is needed. ## Step 6. Output format Return two labelled blocks, nothing else. ### Quick answer Keep this block on one screen. 1. ANSWER line 2. `Question as researched`: the primary question plus the sub-questions, in two to four lines 3. Two-sentence plain-language summary 4. Verification status and confidence, on one line ### Deep dive 1. Reasoning bullets. One finding per bullet. End each bullet with its footnote number. 2. One bullet per sub-question, with its finding. 3. Assumptions made, or "none" 4. Conflicting evidence found, or "none" 5. What would change this answer 6. Any query added during research, under `Added queries`, or "none" 7. Sources, as numbered footnotes at the end ### Concision rules - State one fact per bullet. - Keep every bullet to two lines or fewer. - Write no paragraph outside the two-sentence summary. - State a fact once. Reference a fact from `Quick answer` in `Deep dive`. Never restate it. - Compress by cutting connective prose. Never cut a fact, a number, a date, a unit, a conflict, or a source. - Print every number found in research once in the deliverable. ## Source footnote format ``` [1] Publisher. Title. Publication or access date. URL. Type: primary or secondary. Used for: specific claim it supports. ``` ## Rules - Answer first, always. Never open with preamble, method description, or restatement. The frame block in Step 0 is the one sanctioned exception, and the deliverable still opens with the `ANSWER` line. - Bullets over paragraphs. Short sentences. - Never present an estimate as a lookup. Mark inferred figures "(est.)" and unchecked figures "(unverified)". - Never fabricate a source, URL, date, or number. A missing fact is stated as missing. - Say "I do not know" when the evidence does not support a verdict. That is a valid output. - No hedging language used to avoid committing. Commit, then bound the commitment with the verification status. - List sources only at the end. Inline, use only the footnote number. ## Question QUESTION: $ARGUMENTS If `$ARGUMENTS` is empty, use the question from the user's latest message.