AIO APEX
fr
Works well with Claude Opus 4.5 or GPT-5.1 for the structured cross-referencing this requires — both hold attribution steady across long combined source text without losing track of which claim came from which source. Gemini 3 Pro works fine for shorter source sets (under roughly four sources) but can blur attribution once the combined text gets long.A product manager has four analyst reports and a dozen saved articles on whether on-device AI or cloud AI will dominate enterprise deployment by 2027, needs to present a one-page synthesis to the leadership team by Friday afternoon, and does not have time to read everything closely enough to defend a recommendation under questioning in the room.learning

Le Synthétiseur de Recherche : Transformez des Sources Dispersées en une Carte des Accords et des Contradictions

Partager:
Le Synthétiseur de Recherche : Transformez des Sources Dispersées en une Carte des Accords et des Contradictions

Pourquoi ce prompt est important

Without a structured synthesis step, people default to citing whichever single source already matched what they expected to find, and the gap surfaces at the worst possible moment — when a stakeholder points out mid-meeting that another source in the same pile actually argues the opposite. That credibility hit costs far more time and trust to recover from than the twenty minutes a proper synthesis would have taken upfront.

À quoi nous l'utilisons

A product manager has four analyst reports and a dozen saved articles on whether on-device AI or cloud AI will dominate enterprise deployment by 2027, needs to present a one-page synthesis to the leadership team by Friday afternoon, and does not have time to read everything closely enough to defend a recommendation under questioning in the room.

Prompt

Role: Act as a research synthesis analyst who specializes in finding genuine points of agreement, disagreement, and unaddressed gaps across multiple sources covering the same question — not a summarizer who blends everything into one smooth narrative.

Context: I am researching [YOUR RESEARCH QUESTION]. Here are the sources, each clearly labeled:
Source A: [PASTE SUMMARY, ABSTRACT, OR FULL TEXT]
Source B: [PASTE SUMMARY, ABSTRACT, OR FULL TEXT]
Source C: [PASTE SUMMARY, ABSTRACT, OR FULL TEXT]
[ADD MORE SOURCES AS NEEDED, LABELED SOURCE D, E, ETC.]

Task: Produce a structured synthesis that:
1. Identifies the 3 to 5 core claims being made across all sources combined
2. For each core claim, states which sources agree, which disagree, and quotes or closely paraphrases the specific language showing the agreement or disagreement
3. Distinguishes disagreements that are factual (verifiable, at least one source is likely wrong) from disagreements that are interpretive (both sources can be reasonably right, they emphasize different things)
4. Identifies at least [NUMBER] unaddressed gaps — specific sub-questions none of the sources cover, that would need separate research
5. Flags any source that appears to have a conflict of interest or vested stake in a particular conclusion, such as a vendor-sponsored study or an advocacy group

Constraints:
- Do not average disagreements into a vague middle-ground conclusion. State disagreements explicitly, with attribution to the specific source.
- Do not treat source volume as a proxy for correctness. Three sources repeating the same unverified claim is not stronger evidence than one source citing primary data — check whether sources are independently reporting or just citing each other before counting them as separate confirmations.
- If two sources make the same claim but one cites a primary source and the other cites the first source, note this explicitly as one underlying data point, not two.
- Keep the total output under [WORD LIMIT] words.

Output Format:
1. A short summary paragraph (3 to 4 sentences) stating the overall state of consensus or lack of it
2. "Where Sources Agree" — list of claims with supporting sources
3. "Where Sources Disagree" — list of claims with which sources say what, and whether the disagreement is factual or interpretive
4. "Unaddressed Gaps" — list of the open sub-questions
5. "Caveats" — short section flagging any source credibility concerns

Résultat

Overall State of Consensus: Of the three sources reviewed, there is partial agreement that on-device inference will handle a growing share of routine enterprise AI tasks by 2027, but the sources sharply disagree on whether this displaces cloud AI for core workloads or simply adds a new tier alongside it. Source C's central claim appears to rest heavily on Source A's original survey data without independent verification.

Where Sources Agree: - On-device inference costs have dropped enough to handle routine classification and short-context tasks locally (Sources A, B, C) - Enterprise IT buyers cite data residency, not raw cost savings, as the top driver for on-device adoption (Sources A, B)

Where Sources Disagree: - Claim: "On-device AI will capture the majority of enterprise AI spend by 2027." Source A projects 35% share by 2027 based on a survey of 200 IT buyers; Source B projects under 15% share, arguing complex reasoning tasks still require cloud-scale models regardless of hardware improvements. This is a factual disagreement — the two sources make incompatible quantitative predictions, and only one survey methodology can be sounder. Worth checking sample composition before trusting either figure. - Claim: "Cloud AI providers are cutting prices industry-wide to defend market share." Source C asserts this is already happening broadly; Source A documents only one vendor doing this. This is an interpretive disagreement — Source C may be generalizing from a single data point rather than an industry pattern.

Unaddressed Gaps: - No source addresses how data-residency requirements outside the US and EU affect this split - No source quantifies the latency threshold at which enterprises actually migrate workloads from cloud to on-device, versus simply running new workloads on-device from the start

Caveats: Source C is published by a chipmaker with a direct commercial interest in on-device AI adoption and should be weighted accordingly — its 35%-by-2027 figure traces back to Source A's survey rather than independent data.

La façon habituelle dont les gens gèrent des sources contradictoires est de choisir celle qui correspond déjà à ce qu'ils voulaient croire, ou de demander à un outil d'IA un résumé rapide qui mélange tout en un seul paragraphe assuré. Les deux approches effacent discrètement l'information la plus utile du lot : les points précis où des sources crédibles sont réellement en désaccord, et pourquoi.

Pourquoi le Role compte ici

Le Role présente le modèle comme un analyste de synthèse plutôt qu'un résumeur, car ce sont des métiers différents. Le travail d'un résumeur est de compresser ; celui d'un analyste de synthèse est de cartographier le désaccord et de l'attribuer correctement. Demander à un assistant généraliste de « résumer ces quatre rapports » produit presque toujours la version lissée, car le lissage est ce que la synthèse optimise par défaut.

Ce que fait vraiment la section Task

Les cinq exigences numérotées imposent une séquence précise : trouver d'abord les affirmations centrales, puis vérifier l'accord par affirmation, puis classer les désaccords en factuels ou interprétatifs. Cette dernière distinction est celle que la plupart des gens sautent, et c'est celle qui compte le plus. Un désaccord factuel signifie que quelqu'un a probablement tort et qu'il vaut la peine d'examiner la méthodologie. Un désaccord interprétatif signifie que les deux sources peuvent avoir raison tout en mettant l'accent sur des choses différentes — le traiter comme un litige factuel gâche du temps à chercher une réponse « correcte » qui n'existe pas.

La Contrainte qui fait le vrai travail

« Ne moyennez pas les désaccords en une conclusion vague et intermédiaire » est l'instruction qui protège tout l'objectif de l'exercice. Sans contrainte, la plupart des modèles résoudront la tension en trouvant un langage doux et prudent qui accommode techniquement chaque source — exactement la sortie qui prend quelqu'un au dépourvu quand un collègue signale que les sources n'étaient en fait jamais d'accord. La contrainte de comptage des citations compte tout autant : trois billets de blog qui citent tous la même étude originale constituent une seule donnée portant trois tenues différentes, pas trois confirmations indépendantes.

Pourquoi le format de sortie est divisé en quatre sections

Séparer « Où les Sources S'accordent » de « Où les Sources Diffèrent » force les désaccords à rester visibles plutôt que de disparaître dans un récit combiné. La section « Lacunes Non Traitées » existe parce que l'erreur de recherche la plus coûteuse n'est pas de mal lire une source — c'est de supposer qu'une question a été couverte alors qu'elle n'a jamais été posée. La section « Mises en Garde » repère les sources qui ont une raison de préférer une conclusion donnée, ce qui compte plus que le nombre de sources quand on décide quel poids accorder à une affirmation.

Exécutez ceci avant toute réunion, note ou décision qui dépend de la réconciliation de plusieurs sources en désaccord — pas après, quand les lacunes sont déjà devenues l'objection de quelqu'un d'autre.

decision-makinglearningresearchsynthesiscritical-thinking
Partager: