Voici une préconisation structurée et pragmatique de l’utilisation des Anthropic Agent Skills (au sens Agent Skills / MCP-compatible) pour la couche 1 “Canaux usager” de Synergia, en respectant strictement votre principe clé :

La couche 1 n’exécute pas de décisions métier.
Elle capte l’intention, structure les signaux, collecte les preuves.

Je distingue donc clairement :

Skills d’instrumentation (capture / structuration),
Skills d’analyse légère (sans décision),
Skills interdits à ce niveau (à réserver à la couche 2 Orchestration).

1. Principe général d’utilisation des Agent Skills Anthropic dans Synergia

Règle d’or

Dans la couche 1 :

✅ Skills stateless, non décisionnels
✅ Sorties JSON strictes, typées
❌ Aucune orchestration métier
❌ Aucun accès direct à Odoo en écriture

Les skills Anthropic servent ici de :

parseurs d’intention
extracteurs multimodaux
normalisateurs de signaux faibles

2. Cartographie des Skills Anthropic par sous-canal

A. WebApp Nurse (smartphone – IDEL) > Transformer des interactions humaines (checklists, notes, photos, vitaux) en événements cliniques structurés.

Objectif

Transformer des interactions humaines (checklists, notes, photos, vitaux) en événements cliniques structurés.

Skills Anthropic recommandés

Usage	Skill Anthropic	Rôle exact	Sortie
Texte libre / notes	intent_extraction	Extraire intention clinique sans juger	intent_type, confidence
Checklists	form_normalizer	Normaliser checklist UI → JSON	visit_event
Photos (plaies, pilulier)	vision_feature_extraction	Décrire visuellement sans diagnostic	visual_observation[]
Bouton “personne à risque”	risk_flag_parser	Structurer le signal + justification humaine	risk_flag{reason, free_text}
Vitaux saisis	medical_value_parser	Validation syntaxique + unité	observation{value, unit}

Skills explicitement interdits

Diagnostic
Priorisation
Décision d’alerte

👉 Le skill ne fait que rendre la donnée exploitable.> c'est pour passer la "balle" aux médecins ( voir avec @Pascal )

B. WebApp Caregiver (aide à domicile) @antonio> Capteur de signaux faibles non médicaux.

Objectif

Capteur de signaux faibles non médicaux.

Skills Anthropic recommandés

Usage	Skill	Rôle	Sortie
Texte incident	event_classification_light	Classifier sans gravité	care_event_type
Besoin exprimé	need_structuring	Structurer demande	need_request{category}
Journal quotidien	daily_log_normalizer	Structuration narrative → événements	care_event[]

Point clé

Les skills ne savent pas ce qu’est une urgence.

Ils préparent la matière pour la couche 2.

C. App / Portail Famille @Armel> Canal émotionnel + consentement, sans médicalisation.

Objectif

Canal émotionnel + consentement, sans médicalisation.

Skills Anthropic recommandés

Usage	Skill	Rôle	Sortie
Message texte	message_intent_tagging	Tagging sémantique simple	family_message{topic}
Photo envoyée	image_captioning_safe	Description neutre	media_post{caption}
Consentement	consent_statement_parser	Vérifier complétude	consent_update{scope, explicit}

Important

Aucun enrichissement clinique
Pas d’inférence médicale sur les images

D. RCS / SMS / WhatsApp / WhatsUs @antonio > réponses ultra-courtes en événements fiables.

Objectif

Transformer des réponses ultra-courtes en événements fiables.

Skills Anthropic recommandés

Usage	Skill	Rôle	Sortie
OUI / NON	binary_reply_parser	Désambiguïsation stricte	reply_yesno{value}
Confirmation RDV	appointment_confirmation_parser	Parsing contextuel	appointment_confirm{id}
Accusé réception	acknowledgement_parser	Marquer réception	alert_ack{timestamp}

Note critique

Ici, la robustesse prime sur l’intelligence.

Skills très simples, très déterministes.

E. WebRTC (voix / visio) > @Finary + @Antonio Transformer une interaction riche en preuves exploitables.

Objectif

Transformer une interaction riche en preuves exploitables.

Skills Anthropic recommandés

Usage	Skill	Rôle	Sortie
Voix → texte	speech_to_text	Transcription horodatée	call_transcript
Analyse conversation	intent_candidate_extraction	Proposer intentions possibles	intent_candidates[]
Pièces évoquées	evidence_reference_extractor	Lier propos → preuves	evidence_snippets[]
Résumé	neutral_call_summary	Résumé factuel, non interprétatif	summary_text

Règle stricte

Le résumé :

❌ ne conclut rien
❌ ne recommande rien
✅ cite uniquement ce qui a été dit

F. IoT / Capteurs (mmWave, pilulier, wearables) @gabriel 2026

Objectif

Normaliser des flux temps réel → événements lisibles.

Skills Anthropic recommandés

Usage	Skill	Rôle	Sortie
Événement capteur	sensor_event_normalizer	Unification formats	sensor_event{type}
Détection chute	pattern_labeler	Étiquetage probabiliste	fall_suspect{confidence}
Observance	adherence_event_builder	Écart planifié / réel	adherence{delta}

Important

Le skill ne déclenche jamais d’alerte
Il fournit un indice probabiliste

3. Tableau de synthèse – Skills autorisés par type

Type de skill	Autorisé couche 1	Justification
Intent extraction	✅	Instrumentation
Vision captioning	✅	Observation factuelle
Speech-to-text	✅	Preuve primaire
Classification légère	✅	Pré-tri
Résumé neutre	✅	Lisibilité
Recommandation	❌	Décision
Orchestration	❌	Métier
Priorisation	❌	Responsabilité
Appel API métier	❌	Couche 2 uniquement

4. Positionnement stratégique (important)

Skills Anthropic dans Synergia ≠ “plugins”

Ils sont utilisés comme :

drivers cognitifs
capteurs sémantiques
normalisateurs multimodaux

👉 La valeur métier commence après, dans la couche Orchestration Synergia.

Proposition opérationnelle de Skill Manifest “Anthropic ↔ Synergia” (format YAML)

Ci-dessous une proposition opérationnelle de Skill Manifest “Anthropic ↔ Synergia” (format YAML) incluant :

permissions déclaratives (policy layer)
schémas JSON exacts (JSON Schema Draft 2020-12) pour inputs et outputs
enveloppe d’audit standard (trace_id, actor, channel, evidence_refs)
mappings Odoo (ce que la couche 2 écrira ensuite, sans l’autoriser en couche 1)

Vous pouvez l’utiliser comme base d’un registre interne et l’exposer via une interface MCP / Agent Skills.

0) Conventions Synergia (obligatoires)

0.1 Enveloppe d’entrée standard (SkillInputEnvelope)


{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://synergia.example/schemas/SkillInputEnvelope.json",
  "title": "SkillInputEnvelope",
  "type": "object",
  "additionalProperties": false,
  "required": ["meta", "payload"],
  "properties": {
    "meta": {
      "type": "object",
      "additionalProperties": false,
      "required": ["trace_id", "timestamp", "channel", "actor", "tenant_id"],
      "properties": {
        "trace_id": { "type": "string", "minLength": 12, "maxLength": 128 },
        "timestamp": { "type": "string", "format": "date-time" },
        "tenant_id": { "type": "string", "minLength": 2, "maxLength": 64 },
        "channel": {
          "type": "string",
          "enum": ["webapp_nurse", "webapp_caregiver", "portal_family", "rcs_sms_whatsapp", "webrtc", "iot"]
        },
        "actor": {
          "type": "object",
          "additionalProperties": false,
          "required": ["type", "id"],
          "properties": {
            "type": { "type": "string", "enum": ["human", "service", "device", "agent"] },
            "id": { "type": "string", "minLength": 1, "maxLength": 128 },
            "role": {
              "type": "string",
              "enum": ["idel", "caregiver", "coordinator", "family", "senior", "system"],
              "default": "system"
            }
          }
        },
        "context": {
          "type": "object",
          "additionalProperties": true,
          "properties": {
            "senior_id": { "type": ["string", "null"], "maxLength": 128 },
            "visit_id": { "type": ["string", "null"], "maxLength": 128 },
            "locale": { "type": "string", "default": "fr-FR" }
          }
        }
      }
    },
    "payload": { "type": "object" }
  }
}

0.2 Enveloppe de sortie standard (SkillOutputEnvelope)


{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://synergia.example/schemas/SkillOutputEnvelope.json",
  "title": "SkillOutputEnvelope",
  "type": "object",
  "additionalProperties": false,
  "required": ["meta", "result"],
  "properties": {
    "meta": {
      "type": "object",
      "additionalProperties": false,
      "required": ["trace_id", "skill", "version", "status", "duration_ms"],
      "properties": {
        "trace_id": { "type": "string" },
        "skill": { "type": "string" },
        "version": { "type": "string" },
        "status": { "type": "string", "enum": ["success", "error"] },
        "duration_ms": { "type": "integer", "minimum": 0 },
        "warnings": { "type": "array", "items": { "type": "string" } }
      }
    },
    "result": { "type": "object" },
    "error": {
      "type": "object",
      "additionalProperties": false,
      "required": ["code", "message"],
      "properties": {
        "code": { "type": "string" },
        "message": { "type": "string" },
        "details": { "type": "object", "additionalProperties": true }
      }
    }
  }
}

1) Permissions déclaratives Synergia (Policy Layer)

1.1 Vocabulaire de permissions (exemple minimal)

pii.read (lecture données identifiantes)
media.read / media.write (accès média)
transcript.write (stockage transcription)
classify.intent (extraction intention)
vision.describe (description visuelle non diagnostique)
sensor.normalize (normalisation IoT)
notify.send (envoi messages, couche 2 seulement)
odoo.write:* (interdit couche 1)

Règle : couche 1 = aucune permission odoo.write:* ni notify.send.

2) Skill Manifest (YAML) — “Anthropic ↔ Synergia”

Ce manifeste décrit des skills “d’instrumentation” utilisables par le runtime (Anthropic/Agent Skills) sur la couche 1.
La couche 2 consomme les sorties et écrit dans Odoo.

manifest_version: "1.0" namespace: "synergia" registry: "internal" defaults: output_envelope_schema: "https://synergia.example/schemas/SkillOutputEnvelope.json" input_envelope_schema: "https://synergia.example/schemas/SkillInputEnvelope.json" audit: required: true log_level: "info" redact: - "$.payload.raw_text" - "$.payload.transcript_text" - "$.payload.image_base64" skills: - id: "synergia.intent_extraction.v1" title: "Extraction d'intention (texte)" description: "Transforme un texte libre en intention normalisée (non décisionnelle)." channels: ["webapp_nurse", "webapp_caregiver", "portal_family", "rcs_sms_whatsapp", "webrtc"] permissions_required: ["classify.intent"] permissions_denied: ["odoo.write:*", "notify.send"] policy: require_human_approval: false pii_handling: "minimize" io: input_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "payload"] properties: meta: $ref: "https://synergia.example/schemas/SkillInputEnvelope.json#/properties/meta" payload: type: "object" additionalProperties: false required: ["text"] properties: text: { type: "string", minLength: 1, maxLength: 4000 } hints: type: "object" additionalProperties: false properties: domain: { type: "string", enum: ["senior_care", "admin"], default: "senior_care" } allowed_intents: type: "array" items: { type: "string" } output_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "result"] properties: meta: $ref: "https://synergia.example/schemas/SkillOutputEnvelope.json#/properties/meta" result: type: "object" additionalProperties: false required: ["intent_type", "confidence", "slots"] properties: intent_type: type: "string" enum: - "visit_event" - "observation" - "risk_flag" - "task_request" - "care_event" - "need_request" - "family_message" - "consent_update" - "reply_yesno" - "appointment_confirm" - "alert_ack" - "unknown" confidence: { type: "number", minimum: 0, maximum: 1 } slots: type: "object" additionalProperties: false properties: senior_ref: { type: ["string", "null"], maxLength: 128 } free_text: { type: ["string", "null"], maxLength: 2000 } date_time: { type: ["string", "null"], format: "date-time" } mapping_hint: odoo_targets: - model: "x_synergia_intent_event" mode: "create" fields: trace_id: "$.meta.trace_id" channel: "$.meta.channel" actor_id: "$.meta.actor.id" intent_type: "$.result.intent_type" confidence: "$.result.confidence" - id: "synergia.vision_feature_extraction.v1" title: "Description visuelle neutre (photo)" description: "Produit une description factuelle d'une image sans diagnostic." channels: ["webapp_nurse", "portal_family"] permissions_required: ["vision.describe", "media.read"] permissions_denied: ["odoo.write:*", "notify.send"] policy: require_human_approval: false medical_inference: "forbidden" io: input_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "payload"] properties: meta: $ref: "https://synergia.example/schemas/SkillInputEnvelope.json#/properties/meta" payload: type: "object" additionalProperties: false required: ["image_ref"] properties: image_ref: type: "object" additionalProperties: false required: ["kind", "value"] properties: kind: { type: "string", enum: ["url", "odoo_attachment_id", "blob_id"] } value: { type: "string", minLength: 1, maxLength: 512 } prompt_guardrails: type: "object" additionalProperties: false properties: forbid: type: "array" items: { type: "string" } default: ["diagnosis", "treatment", "cause_of_injury"] output_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "result"] properties: meta: $ref: "https://synergia.example/schemas/SkillOutputEnvelope.json#/properties/meta" result: type: "object" additionalProperties: false required: ["caption", "observations"] properties: caption: { type: "string", minLength: 1, maxLength: 800 } observations: type: "array" minItems: 0 maxItems: 20 items: type: "object" additionalProperties: false required: ["label", "evidence"] properties: label: { type: "string", maxLength: 80 } evidence: { type: "string", maxLength: 240 } safety_flags: type: "array" items: { type: "string" } mapping_hint: odoo_targets: - model: "x_synergia_evidence" mode: "create" fields: trace_id: "$.meta.trace_id" kind: "image_caption" content: "$.result.caption" ref: "$.payload.image_ref.value" - id: "synergia.risk_flag_parser.v1" title: "Structuration 'personne à risque' (UI)" description: "Convertit la sélection UI + justification en risk_flag structuré (sans décider)." channels: ["webapp_nurse"] permissions_required: ["classify.intent"] permissions_denied: ["odoo.write:*", "notify.send"] policy: require_human_approval: false requires_reason: true io: input_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "payload"] properties: meta: $ref: "https://synergia.example/schemas/SkillInputEnvelope.json#/properties/meta" payload: type: "object" additionalProperties: false required: ["reason_code"] properties: reason_code: type: "string" enum: - "fall_risk" - "med_non_adherence" - "confusion_acute" - "dehydration_risk" - "social_isolation" - "unsafe_home" - "other" free_text: { type: ["string", "null"], maxLength: 800 } evidence_refs: type: "array" items: { type: "string", maxLength: 256 } output_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "result"] properties: meta: $ref: "https://synergia.example/schemas/SkillOutputEnvelope.json#/properties/meta" result: type: "object" additionalProperties: false required: ["risk_flag"] properties: risk_flag: type: "object" additionalProperties: false required: ["reason_code", "statement", "confidence"] properties: reason_code: { type: "string" } statement: { type: "string", minLength: 1, maxLength: 900 } confidence: { type: "number", minimum: 0, maximum: 1 } mapping_hint: odoo_targets: - model: "x_synergia_risk_flag" mode: "create" fields: trace_id: "$.meta.trace_id" senior_id: "$.meta.context.senior_id" reason_code: "$.result.risk_flag.reason_code" statement: "$.result.risk_flag.statement" confidence: "$.result.risk_flag.confidence" - id: "synergia.binary_reply_parser.v1" title: "Parsing OUI/NON (RCS/SMS/WhatsApp)" description: "Désambiguïse une réponse courte en booléen + fiabilité." channels: ["rcs_sms_whatsapp"] permissions_required: ["classify.intent"] permissions_denied: ["odoo.write:*", "notify.send"] io: input_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "payload"] properties: meta: $ref: "https://synergia.example/schemas/SkillInputEnvelope.json#/properties/meta" payload: type: "object" additionalProperties: false required: ["raw_reply"] properties: raw_reply: { type: "string", minLength: 1, maxLength: 80 } locale: { type: "string", default: "fr-FR" } output_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "result"] properties: meta: $ref: "https://synergia.example/schemas/SkillOutputEnvelope.json#/properties/meta" result: type: "object" additionalProperties: false required: ["value", "confidence", "normalized"] properties: value: { type: ["boolean", "null"] } confidence: { type: "number", minimum: 0, maximum: 1 } normalized: { type: "string", enum: ["yes", "no", "unknown"] } - id: "synergia.call_transcript_builder.v1" title: "Transcription & segments horodatés (WebRTC)" description: "Construit une transcription horodatée segmentée, sans interprétation." channels: ["webrtc"] permissions_required: ["transcript.write"] permissions_denied: ["odoo.write:*", "notify.send"] policy: require_human_approval: false store_audio: "optional" io: input_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "payload"] properties: meta: $ref: "https://synergia.example/schemas/SkillInputEnvelope.json#/properties/meta" payload: type: "object" additionalProperties: false required: ["audio_ref"] properties: audio_ref: type: "object" additionalProperties: false required: ["kind", "value"] properties: kind: { type: "string", enum: ["url", "blob_id"] } value: { type: "string", minLength: 1, maxLength: 512 } diarization: { type: "boolean", default: true } output_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "result"] properties: meta: $ref: "https://synergia.example/schemas/SkillOutputEnvelope.json#/properties/meta" result: type: "object" additionalProperties: false required: ["segments", "language"] properties: language: { type: "string", default: "fr" } segments: type: "array" minItems: 1 items: type: "object" additionalProperties: false required: ["t0_ms", "t1_ms", "speaker", "text"] properties: t0_ms: { type: "integer", minimum: 0 } t1_ms: { type: "integer", minimum: 0 } speaker: { type: "string", maxLength: 64 } text: { type: "string", minLength: 1, maxLength: 600 } mapping_hint: odoo_targets: - model: "x_synergia_call_transcript" mode: "create" fields: trace_id: "$.meta.trace_id" visit_id: "$.meta.context.visit_id" segments_json: "$.result.segments" - id: "synergia.sensor_event_normalizer.v1" title: "Normalisation événements IoT" description: "Unifie le format des événements capteurs (mmWave, pilulier, wearables)." channels: ["iot"] permissions_required: ["sensor.normalize"] permissions_denied: ["odoo.write:*", "notify.send"] io: input_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "payload"] properties: meta: $ref: "https://synergia.example/schemas/SkillInputEnvelope.json#/properties/meta" payload: type: "object" additionalProperties: true required: ["source", "event_type", "ts"] properties: source: { type: "string", maxLength: 64 } event_type: { type: "string", maxLength: 64 } ts: { type: "string", format: "date-time" } output_schema: $schema: "https://json-schema.org/draft/2020-12/schema" type: "object" additionalProperties: false required: ["meta", "result"] properties: meta: $ref: "https://synergia.example/schemas/SkillOutputEnvelope.json#/properties/meta" result: type: "object" additionalProperties: false required: ["sensor_event"] properties: sensor_event: type: "object" additionalProperties: false required: ["type", "timestamp", "attributes"] properties: type: type: "string" enum: ["fall_suspect", "adherence", "presence", "vitals", "unknown"] timestamp: { type: "string", format: "date-time" } attributes: { type: "object", additionalProperties: true } confidence: { type: ["number", "null"], minimum: 0, maximum: 1 }

3) “Policy Set” minimal par canal (recommandation)

Vous pouvez définir une politique simple, réutilisable :

policy_sets: layer1_channels_default: allow_permissions: - "classify.intent" - "vision.describe" - "media.read" - "sensor.normalize" - "transcript.write" deny_permissions: - "odoo.write:*" - "notify.send" - "fs.read" - "fs.write" - "exec.run" require_audit: true require_redaction: true

4) Exemple d’appel (input) et réponse (output)

4.1 WebApp Nurse — “personne à risque”

Input

{ "meta": { "trace_id": "trc_20251219_134401_9K3D", "timestamp": "2025-12-19T13:44:01Z", "tenant_id": "smartco-ch", "channel": "webapp_nurse", "actor": { "type": "human", "id": "usr_idel_1029", "role": "idel" }, "context": { "senior_id": "rp_5561", "visit_id": "vst_8891", "locale": "fr-FR" } }, "payload": { "reason_code": "fall_risk", "free_text": "Instabilité observée au lever, appui mural.", "evidence_refs": ["photo:att_772", "note:obs_19"] } }

Output


{
  "meta": {
    "trace_id": "trc_20251219_134401_9K3D",
    "skill": "synergia.risk_flag_parser.v1",
    "version": "1.0.0",
    "status": "success",
    "duration_ms": 83,
    "warnings": []
  },
  "result": {
    "risk_flag": {
      "reason_code": "fall_risk",
      "statement": "Risque de chute signalé : instabilité au lever (appui mural).",
      "confidence": 0.86
    }
  }
}

Préconisation utilisation des Anthropic Agent Skills (au sens Agent Skills / MCP-compatible) pour la couche 1 “Canaux usager” de Synergia,

1. Principe général d’utilisation des Agent Skills Anthropic dans Synergia

Règle d’or

Objectif

Skills Anthropic recommandés

Skills explicitement interdits

B. WebApp Caregiver (aide à domicile) @antonio> Capteur de signaux faibles non médicaux.

Objectif

Skills Anthropic recommandés

Point clé

C. App / Portail Famille @Armel> Canal émotionnel + consentement, sans médicalisation.

Objectif

Skills Anthropic recommandés

Important

D. RCS / SMS / WhatsApp / WhatsUs @antonio > réponses ultra-courtes en événements fiables.

Objectif

Skills Anthropic recommandés

Note critique

E. WebRTC (voix / visio) > @Finary + @Antonio Transformer une interaction riche en preuves exploitables.

Objectif

Skills Anthropic recommandés

Règle stricte

F. IoT / Capteurs (mmWave, pilulier, wearables) @gabriel 2026

Objectif

Skills Anthropic recommandés

Important

Skills Anthropic dans Synergia ≠ “plugins”

0) Conventions Synergia (obligatoires)

0.1 Enveloppe d’entrée standard (SkillInputEnvelope)

0.2 Enveloppe de sortie standard (SkillOutputEnvelope)

1) Permissions déclaratives Synergia (Policy Layer)

1.1 Vocabulaire de permissions (exemple minimal)

2) Skill Manifest (YAML) — “Anthropic ↔ Synergia”

3) “Policy Set” minimal par canal (recommandation)

4) Exemple d’appel (input) et réponse (output)

4.1 WebApp Nurse — “personne à risque”

5) Point d’attention clé (architecture)

Suivez-nous