diff --git a/CHANGELOG.md b/CHANGELOG.md
index 614aab65f..2f0631805 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,6 +6,7 @@ This file is the release-notes source of truth for Marinara Engine. Reuse these
 
 ### Added
 
+- Added optional image prompting instructions to image connections, applying them inside existing selfie and Illustrator prompt-writing calls before provider review or generation.
 - Added a Gallery image-agent picker that can run any active custom image-producing agent alongside the base Illustrator (#4846).
 - Added an editable Storyboard Agent shot-planner stage that inspects each generated keyframe before video generation, persists its suitability classification, and falls back to the planned motion when image-aware refinement is unavailable or invalid. The Storyboard Agent page now explains the four-stage prompt workflow and orders its shared prompt editors from illustration through image-aware grounding to video generation (#4839, Pasta-Devs/Marinara-Agents#296).
 - Added batch selection to Character and Persona image galleries so selected images can be downloaded or deleted together after confirmation (#4832).
@@ -47,6 +48,7 @@ This file is the release-notes source of truth for Marinara Engine. Reuse these
 
 ### Fixed
 
+- Scoped image-connection prompt instructions to image-producing agents, bounded stored instructions to 20,000 characters, and kept each retry agent on its own configured image connection.
 - Kept chat settings profiles within their saved chat mode, migrated legacy Visual Novel profiles to Roleplay on import, and prevented reusable profiles from overwriting branch identity (#4849).
 - Served the Home shell directly for unknown routes so Termux mobile launches cannot recurse through the server's 404 handler (#4850).
 - Corrected legacy Google Gemini `/v1` connection URLs to `/v1beta` for connection checks, model discovery, chat generation, and embeddings (#4854).
diff --git a/packages/client/src/components/connections/ConnectionEditor.tsx b/packages/client/src/components/connections/ConnectionEditor.tsx
index 961cbb18d..cc89fe673 100644
--- a/packages/client/src/components/connections/ConnectionEditor.tsx
+++ b/packages/client/src/components/connections/ConnectionEditor.tsx
@@ -95,6 +95,8 @@ import {
   sanitizeImageGenerationProfile,
   sanitizeVideoGenerationProfile,
   suggestImageStyleProfileIdForModel,
+  MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH,
+  normalizeImagePromptInstructions,
   parseConnectionImageCaptioningDefaults,
   type APIProvider,
   type ComfyUiLoraSetting,
@@ -299,6 +301,7 @@ export function ConnectionEditor() {
   const [localComfyuiWorkflow, setLocalComfyuiWorkflow] = useState("");
   const [localImageService, setLocalImageService] = useState<string | null>(null);
   const [localImageEndpointId, setLocalImageEndpointId] = useState("");
+  const [localImagePromptInstructions, setLocalImagePromptInstructions] = useState("");
   const [localImageGenerationQuality, setLocalImageGenerationQuality] = useState<ImageGenerationQuality>("auto");
   const [localVideoGenerationSource, setLocalVideoGenerationSource] = useState("");
   const [localVideoService, setLocalVideoService] = useState<string | null>(null);
@@ -418,6 +421,7 @@ export function ConnectionEditor() {
     setLocalComfyuiWorkflow((c.comfyuiWorkflow as string) ?? "");
     setLocalImageService(imageService);
     setLocalImageEndpointId((c.imageEndpointId as string) ?? "");
+    setLocalImagePromptInstructions((c.imagePromptInstructions as string) ?? "");
     setLocalImageGenerationQuality(
       c.imageGenerationQuality === "low" || c.imageGenerationQuality === "medium" || c.imageGenerationQuality === "high"
         ? c.imageGenerationQuality
@@ -719,6 +723,7 @@ export function ConnectionEditor() {
       imageService: isImageProvider ? localImageGenerationSource || localImageService || null : null,
       imageEndpointId:
         isImageProvider && selectedImageService === "runpod_comfyui" ? localImageEndpointId || null : null,
+      imagePromptInstructions: isImageProvider ? normalizeImagePromptInstructions(localImagePromptInstructions) : null,
       imageGenerationQuality: isImageProvider ? localImageGenerationQuality : "auto",
       videoGenerationSource: isVideoProvider ? selectedVideoProvider || null : null,
       videoService: isVideoProvider ? selectedVideoDefaultsService : null,
@@ -818,6 +823,7 @@ export function ConnectionEditor() {
     localComfyuiWorkflow,
     localImageService,
     localImageEndpointId,
+    localImagePromptInstructions,
     localImageGenerationQuality,
     localMaxTokensOverride,
     localClaudeFastMode,
@@ -923,6 +929,7 @@ export function ConnectionEditor() {
       videoService,
       imageEndpointId:
         isImageProvider && selectedImageService === "runpod_comfyui" ? localImageEndpointId || null : null,
+      imagePromptInstructions: isImageProvider ? normalizeImagePromptInstructions(localImagePromptInstructions) : null,
       imageGenerationQuality: isImageProvider ? localImageGenerationQuality : "auto",
       comfyuiWorkflow:
         isImageProvider || (isVideoProvider && videoProvider === "comfyui") ? localComfyuiWorkflow || null : null,
@@ -961,6 +968,7 @@ export function ConnectionEditor() {
     selectedVideoProvider,
     selectedImageService,
     localImageEndpointId,
+    localImagePromptInstructions,
     localImageGenerationQuality,
     localComfyuiWorkflow,
     localClaudeFastMode,
@@ -2008,6 +2016,25 @@ export function ConnectionEditor() {
             </FieldGroup>
           )}
 
+          {localProvider === "image_generation" && (
+            <FieldGroup
+              label={localizeUi("ui.connections.connectioneditor.imagePromptingInstructions")}
+              icon={<Sparkles size="0.875rem" className="text-sky-400" />}
+              help={localizeUi("ui.connections.connectioneditor.imagePromptingInstructionsHelp")}
+            >
+              <textarea
+                value={localImagePromptInstructions}
+                maxLength={MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH}
+                onChange={(event) => {
+                  setLocalImagePromptInstructions(event.target.value);
+                  markDirty();
+                }}
+                placeholder={localizeUi("ui.connections.connectioneditor.imagePromptingInstructionsPlaceholder")}
+                className="w-full min-h-[96px] resize-y rounded-xl bg-[var(--secondary)] px-3 py-2.5 text-sm outline-none ring-1 ring-[var(--border)] transition-shadow placeholder:text-[var(--muted-foreground)]/50 focus:ring-sky-400/50"
+              />
+            </FieldGroup>
+          )}
+
           {supportsGptImageQuality && (
             <FieldGroup
               label={localizeUi("ui.connections.connectioneditor.gptImageQuality")}
diff --git a/packages/client/src/hooks/use-connections.ts b/packages/client/src/hooks/use-connections.ts
index f36f0293c..ec97b431b 100644
--- a/packages/client/src/hooks/use-connections.ts
+++ b/packages/client/src/hooks/use-connections.ts
@@ -54,6 +54,7 @@ export type CreateConnectionPayload = {
   comfyuiWorkflow?: string | null;
   imageService?: string | null;
   imageEndpointId?: string | null;
+  imagePromptInstructions?: string | null;
   imageGenerationQuality?: ImageGenerationQuality;
   videoGenerationSource?: string | null;
   videoService?: string | null;
diff --git a/packages/client/src/lib/connection-transfer.ts b/packages/client/src/lib/connection-transfer.ts
index 8ed5678f5..f34e7eac8 100644
--- a/packages/client/src/lib/connection-transfer.ts
+++ b/packages/client/src/lib/connection-transfer.ts
@@ -1,4 +1,9 @@
-import { PROVIDERS, type APIProvider, type ImageGenerationQuality } from "@marinara-engine/shared";
+import {
+  normalizeImagePromptInstructions,
+  PROVIDERS,
+  type APIProvider,
+  type ImageGenerationQuality,
+} from "@marinara-engine/shared";
 import type { CreateConnectionPayload } from "../hooks/use-connections";
 
 export type ConnectionTransferRow = {
@@ -29,6 +34,7 @@ export type ConnectionTransferRow = {
   videoService?: unknown;
   service?: unknown;
   imageEndpointId?: unknown;
+  imagePromptInstructions?: unknown;
   imageGenerationQuality?: unknown;
   comfyuiWorkflow?: unknown;
   treatAsLocalEndpoint?: unknown;
@@ -62,6 +68,7 @@ export type SafeConnectionExport = {
   videoGenerationSource: string | null;
   videoService: string | null;
   imageEndpointId: string | null;
+  imagePromptInstructions: string | null;
   imageGenerationQuality: ImageGenerationQuality;
   comfyuiWorkflow: string | null;
   treatAsLocalEndpoint: boolean;
@@ -133,6 +140,7 @@ export function normalizeImportedConnectionEntry(value: unknown): ConnectionImpo
       comfyuiWorkflow: asNullableString(value.comfyuiWorkflow),
       imageService,
       imageEndpointId: asNullableString(value.imageEndpointId),
+      imagePromptInstructions: normalizeImagePromptInstructions(value.imagePromptInstructions),
       imageGenerationQuality: asImageGenerationQuality(value.imageGenerationQuality),
       videoGenerationSource: provider === "video_generation" ? asNullableString(value.videoGenerationSource) : null,
       videoService,
@@ -177,6 +185,7 @@ function serializeConnectionForExport(connection: ConnectionTransferRow): SafeCo
     videoGenerationSource: isVideoProvider ? asNullableString(connection.videoGenerationSource) : null,
     videoService: isVideoProvider ? asNullableString(connection.videoService ?? connection.service) : null,
     imageEndpointId: asNullableString(connection.imageEndpointId),
+    imagePromptInstructions: normalizeImagePromptInstructions(connection.imagePromptInstructions),
     imageGenerationQuality: asImageGenerationQuality(connection.imageGenerationQuality),
     comfyuiWorkflow: asNullableString(connection.comfyuiWorkflow),
     treatAsLocalEndpoint: asBoolean(connection.treatAsLocalEndpoint),
diff --git a/packages/client/src/localization/locales/en.json b/packages/client/src/localization/locales/en.json
index 8a093232a..2cc612a4e 100644
--- a/packages/client/src/localization/locales/en.json
+++ b/packages/client/src/localization/locales/en.json
@@ -4082,6 +4082,9 @@
   "ui.connections.connectioneditor.httpsUsCentral1AiplatformGoogleapisComV1ProjectsMy": "https://us-central1-aiplatform.googleapis.com/v1/projects/my-project/locations/us-central1",
   "ui.connections.connectioneditor.ifYourProxyOrLocalServerIsnTDetected": "If your proxy or local server isn't detected, Windows Defender Firewall may be blocking the connection. Open",
   "ui.connections.connectioneditor.imageCaptioning": "Image Captioning",
+  "ui.connections.connectioneditor.imagePromptingInstructions": "Image Prompting Instructions",
+  "ui.connections.connectioneditor.imagePromptingInstructionsHelp": "When set, these instructions are added to the existing prompt-writer system message and processed by the configured Illustrator provider and model in selfie and Illustrator flows before the prompt is sent to this image backend. Useful for tags, quality tokens, negative-prompt conventions, or local/ComfyUI syntax. An invalid-JSON retry may trigger another call.",
+  "ui.connections.connectioneditor.imagePromptingInstructionsPlaceholder": "Example: use comma-separated tags, add quality tokens, and keep negative terms separate.",
   "ui.connections.connectioneditor.imageQualityAuto": "Auto",
   "ui.connections.connectioneditor.imageQualityHigh": "High",
   "ui.connections.connectioneditor.imageQualityLow": "Low",
diff --git a/packages/server/src/db/schema/connections.ts b/packages/server/src/db/schema/connections.ts
index 99a9cdee4..0b11aa616 100644
--- a/packages/server/src/db/schema/connections.ts
+++ b/packages/server/src/db/schema/connections.ts
@@ -63,6 +63,8 @@ export const apiConnections = fileTable("api_connections", {
   imageService: text("image_service"),
   /** For endpoint-based image services (e.g. RunPod Serverless ComfyUI): the endpoint ID. */
   imageEndpointId: text("image_endpoint_id"),
+  /** Instructions passed to the default language model before image generation. */
+  imagePromptInstructions: text("image_prompt_instructions"),
   /** OpenAI GPT Image quality for this connection. */
   imageGenerationQuality: text("image_generation_quality").notNull().default("auto"),
   /** Explicit video backend selection for video-generation connections. */
diff --git a/packages/server/src/routes/gallery.routes.ts b/packages/server/src/routes/gallery.routes.ts
index 501d26aa3..bc99e81cd 100644
--- a/packages/server/src/routes/gallery.routes.ts
+++ b/packages/server/src/routes/gallery.routes.ts
@@ -66,6 +66,7 @@ import {
 import { resolveIllustratorPromptRuntime } from "../services/generation/illustrator-prompt-runtime.js";
 import { resolveIllustratorImageConnectionId } from "../services/generation/illustrator-background-generation.js";
 import { resolveConversationSelfieSystemPrompt } from "../services/conversation/selfie-prompt.js";
+import { appendImagePromptInstructions } from "../services/generation/image-prompt-instructions.js";
 import {
   suppressesReferencePromptLine,
   resolveIllustratorCharacterReferences,
@@ -1255,6 +1256,10 @@ export async function galleryRoutes(app: FastifyInstance) {
     const selfieSystemPrompt = styleGuidance
       ? `${baseSelfieSystemPrompt}${formatImageStylePromptGuidance(styleGuidance)}`
       : baseSelfieSystemPrompt;
+    const selfieSystemPromptWithImageInstructions = appendImagePromptInstructions(
+      selfieSystemPrompt,
+      imageConn.imagePromptInstructions,
+    );
 
     const selfieAbortSignal = createResponseAbortSignal(reply, SCENE_VIDEO_GENERATION_TIMEOUT_MS, "Selfie generation");
     let promptRuntime;
@@ -1276,7 +1281,7 @@ export async function galleryRoutes(app: FastifyInstance) {
       : `Generate a casual selfie of ${characterName} based on the current conversation context.`;
 
     if (debugLogsEnabled) {
-      debugLog("[debug/gallery/selfie] prompt-builder system:\n%s", selfieSystemPrompt);
+      debugLog("[debug/gallery/selfie] prompt-builder system:\n%s", selfieSystemPromptWithImageInstructions);
       debugLog("[debug/gallery/selfie] prompt-builder user:\n%s", promptContext);
     }
 
@@ -1285,7 +1290,7 @@ export async function galleryRoutes(app: FastifyInstance) {
       try {
         const promptResult = await promptBuilder.chatComplete(
           [
-            { role: "system", content: selfieSystemPrompt },
+            { role: "system", content: selfieSystemPromptWithImageInstructions },
             { role: "user", content: promptContext },
           ],
           {
diff --git a/packages/server/src/routes/game.routes.ts b/packages/server/src/routes/game.routes.ts
index e27517871..e6e13157f 100644
--- a/packages/server/src/routes/game.routes.ts
+++ b/packages/server/src/routes/game.routes.ts
@@ -63,6 +63,7 @@ import { listPartySprites, readPreferredFullBodySpriteBase64 } from "../services
 import {
   buildSceneAnalyzerSystemPrompt,
   buildSceneAnalyzerUserPrompt,
+  compactImagePromptInstructions,
   type SceneAnalyzerContext,
 } from "../services/sidecar/scene-analyzer.js";
 import { postProcessSceneResult, type PostProcessContext } from "../services/sidecar/scene-postprocess.js";
@@ -10316,7 +10317,7 @@ export async function gameRoutes(app: FastifyInstance) {
       .catch(() => null);
     const imagePromptInstructions =
       typeof meta.gameImagePromptInstructions === "string"
-        ? meta.gameImagePromptInstructions.trim().slice(0, 5000)
+        ? compactImagePromptInstructions(meta.gameImagePromptInstructions)
         : "";
 
     // Compute approximate turn number: count user messages + 1 (current turn)
@@ -11199,7 +11200,7 @@ export async function gameRoutes(app: FastifyInstance) {
         null;
       const imagePromptInstructions =
         ownerMode === "game" && typeof meta.gameImagePromptInstructions === "string"
-          ? meta.gameImagePromptInstructions.trim().slice(0, 5000)
+          ? compactImagePromptInstructions(meta.gameImagePromptInstructions)
           : "";
       const useAvatarReferences = meta.storyboardAgentUseAvatarReferences !== false;
       const useStoryboardPromptTemplate =
@@ -12153,7 +12154,7 @@ export async function gameRoutes(app: FastifyInstance) {
       null;
     const imagePromptInstructions =
       typeof meta.gameImagePromptInstructions === "string"
-        ? meta.gameImagePromptInstructions.trim().slice(0, 5000)
+        ? compactImagePromptInstructions(meta.gameImagePromptInstructions)
         : "";
     const useAvatarReferences = input.useAvatarReferences ?? meta.gameImageUseAvatarReferences !== false;
     const includeCharacterAppearance =
@@ -12566,7 +12567,7 @@ export async function gameRoutes(app: FastifyInstance) {
         null;
       const imagePromptInstructions =
         typeof meta.gameImagePromptInstructions === "string"
-          ? meta.gameImagePromptInstructions.trim().slice(0, 5000)
+          ? compactImagePromptInstructions(meta.gameImagePromptInstructions)
           : "";
       const useAvatarReferences = input.useAvatarReferences ?? meta.gameImageUseAvatarReferences !== false;
       const includeCharacterAppearance =
diff --git a/packages/server/src/routes/generate.routes.ts b/packages/server/src/routes/generate.routes.ts
index fa8af2211..efcb96e9e 100644
--- a/packages/server/src/routes/generate.routes.ts
+++ b/packages/server/src/routes/generate.routes.ts
@@ -43,6 +43,7 @@ import {
   unwrapConversationInstructions,
   findKnownModel,
   LOCAL_SIDECAR_CONNECTION_ID,
+  normalizeImagePromptInstructions,
   normalizeTextForMatch,
   parseManagedGenerationParameterDefinitions,
   normalizeGameStoryboardKeyframeCount,
@@ -157,7 +158,12 @@ import {
 import { executeToolCalls, formatToolExecutionResultForModel } from "../services/tools/tool-executor.js";
 import { createAgentPipeline, type ResolvedAgent, type AgentInjection } from "../services/agents/agent-pipeline.js";
 import { DATA_DIR } from "../utils/data-dir.js";
-import { executeAgent, normalizeAgentContextSize, resolveAgentResultType } from "../services/agents/agent-executor.js";
+import {
+  executeAgent,
+  normalizeAgentContextSize,
+  resolveAgentResultType,
+  type AgentExecConfig,
+} from "../services/agents/agent-executor.js";
 import { matchCustomAgentActivation } from "./generate/agent-activation.js";
 import { listCharacterSprites } from "../services/game/sprite.service.js";
 import {
@@ -4336,7 +4342,32 @@ export async function generateRoutes(app: FastifyInstance) {
         // Pre-generation prompt-patch agents read the assembled prompt here; this is overwritten
         // with the fitted provider prompt before each main model call.
         agentContext.memory._mainPromptPreview = promptPreviewForAgents(finalMessages);
-        const pipeline = createAgentPipeline(pipelineAgents, agentContext, sendAgentEventAfterMainStream);
+        const resolveImagePromptAgentContext = async (agent: AgentExecConfig, context: AgentContext): Promise<AgentContext> => {
+          const isImagePromptAgent =
+            agent.type === "illustrator" ||
+            (agent.isCustomAgent === true && customAgentHasCapability(agent.settings, "trigger_image_generation"));
+          if (!isImagePromptAgent) return context;
+
+          const memory = { ...context.memory };
+          delete memory._imagePromptInstructions;
+          const imageConnectionId =
+            agent.type === "illustrator"
+              ? resolveIllustratorImageConnectionId(chatMode, chatMeta, agent.settings.imageConnectionId)
+              : typeof agent.settings.imageConnectionId === "string"
+                ? agent.settings.imageConnectionId.trim()
+                : "";
+          let imageConnection = imageConnectionId ? await connections.getWithKey(imageConnectionId) : null;
+          imageConnection ??= await connections.getDefaultForImageGeneration();
+          const imagePromptInstructions = normalizeImagePromptInstructions(imageConnection?.imagePromptInstructions);
+          if (imagePromptInstructions) memory._imagePromptInstructions = imagePromptInstructions;
+          return { ...context, memory };
+        };
+        const pipeline = createAgentPipeline(
+          pipelineAgents,
+          agentContext,
+          sendAgentEventAfterMainStream,
+          resolveImagePromptAgentContext,
+        );
         let directorSecretPlotResults: AgentResult[] = [];
         let directorSecretPlotArcForPrompt: unknown = directorSecretPlotMemory.overarchingArc;
 
@@ -7622,9 +7653,10 @@ export async function generateRoutes(app: FastifyInstance) {
                       historicalLorebookTarget.id,
                     ) ?? phaseRetryContext)
                   : phaseRetryContext;
+                const resolvedRetryContext = await resolveImagePromptAgentContext(agentCfg, retryCtx);
                 const retried = await executeAgent(
                   agentCfg,
-                  retryCtx,
+                  resolvedRetryContext,
                   agentCfg.provider,
                   agentCfg.model,
                   agentCfg.type === "spotify" ? undefined : agentCfg.toolContext,
diff --git a/packages/server/src/routes/generate/retry-agents-route.ts b/packages/server/src/routes/generate/retry-agents-route.ts
index 5ca95ec4c..50d0f1b60 100644
--- a/packages/server/src/routes/generate/retry-agents-route.ts
+++ b/packages/server/src/routes/generate/retry-agents-route.ts
@@ -21,6 +21,7 @@ import {
   normalizeWorldCustomFields,
   normalizeAgentPhaseValue,
   normalizeAgentPromptTemplateSelectionMap,
+  normalizeImagePromptInstructions,
   resolveMacros,
   resolveGameSetupArtStylePrompt,
   resolveAgentPromptTemplate,
@@ -547,10 +548,18 @@ async function executeManualIllustratorPromptRequest(args: {
             conns: args.conns,
             illustratorAgent: args.illustratorEntry.resolved,
           });
+    const imageConnectionId = resolveIllustratorImageConnectionId(
+      args.chat.mode,
+      args.chatMeta,
+      args.illustratorEntry.resolved.settings.imageConnectionId,
+    );
+    let imageConnection = imageConnectionId ? await args.conns.getById(imageConnectionId).catch(() => null) : null;
+    imageConnection ??= await args.conns.getDefaultForImageGeneration().catch(() => null);
     const generated = await writeManualIllustratorPromptPlan({
       illustratorAgent: args.illustratorEntry.resolved,
       context: args.agentContext,
       styleInstruction,
+      imagePromptInstructions: normalizeImagePromptInstructions(imageConnection?.imagePromptInstructions) ?? undefined,
       signal: args.agentContext.signal,
       debugLog: (message, ...values) => logDebugOverride(args.debugMode || isDebugAgentsEnabled(), message, ...values),
     });
@@ -2440,10 +2449,48 @@ async function validateSpotifyRetryPlayback(
   };
 }
 
+function isImagePromptRetryAgent(entry: ResolvedRetryAgent): boolean {
+  return (
+    entry.resolved.type === "illustrator" ||
+    (entry.resolved.isCustomAgent === true &&
+      customAgentHasCapability(entry.resolved.settings, "trigger_image_generation"))
+  );
+}
+
+async function resolveRetryImagePromptContext(args: {
+  entry: ResolvedRetryAgent;
+  context: AgentContext;
+  conns?: ReturnType<typeof createConnectionsStorage>;
+  chatMode?: ChatMode;
+  chatMeta?: Record<string, unknown>;
+}): Promise<AgentContext> {
+  const memory = { ...args.context.memory };
+  delete memory._imagePromptInstructions;
+
+  if (!args.conns || !args.chatMode || !args.chatMeta) {
+    return { ...args.context, memory };
+  }
+
+  const imageConnectionId =
+    args.entry.resolved.type === "illustrator"
+      ? resolveIllustratorImageConnectionId(args.chatMode, args.chatMeta, args.entry.resolved.settings.imageConnectionId)
+      : typeof args.entry.resolved.settings.imageConnectionId === "string"
+        ? args.entry.resolved.settings.imageConnectionId.trim()
+        : "";
+  let imageConnection = imageConnectionId ? await args.conns.getById(imageConnectionId).catch(() => null) : null;
+  imageConnection ??= await args.conns.getDefaultForImageGeneration().catch(() => null);
+  const imagePromptInstructions = normalizeImagePromptInstructions(imageConnection?.imagePromptInstructions);
+  if (imagePromptInstructions) memory._imagePromptInstructions = imagePromptInstructions;
+  return { ...args.context, memory };
+}
+
 async function executeRetryBatches(
   agentContext: AgentContext,
   resolvedAgents: ResolvedRetryAgent[],
   preGenerationContext?: AgentContext | null,
+  conns?: ReturnType<typeof createConnectionsStorage>,
+  chatMode?: ChatMode,
+  chatMeta?: Record<string, unknown>,
 ) {
   const retryAgents = mergeRetryPairedBuiltInRewriteAgents(resolvedAgents);
   const providerModelGroups = new Map<
@@ -2504,13 +2551,15 @@ async function executeRetryBatches(
     async (group) => {
       const toolAgents = group.agents.filter((agent) => shouldUseToolsDuringAgentExecution(agent.resolved));
       const batchAgents = group.agents.filter((agent) => !shouldUseToolsDuringAgentExecution(agent.resolved));
+      const imagePromptAgents = batchAgents.filter(isImagePromptRetryAgent);
+      const regularBatchAgents = batchAgents.filter((agent) => !isImagePromptRetryAgent(agent));
       const groupResults: AgentResult[] = [];
 
-      if (batchAgents.length > 0) {
-        const configs = batchAgents.map((agent) => agent.resolved);
+      if (regularBatchAgents.length > 0) {
+        const configs = regularBatchAgents.map((agent) => agent.resolved);
         const batchResults = await executeAgentBatch(configs, group.context, group.provider, group.model);
         for (const result of batchResults) {
-          const entry = batchAgents.find(
+          const entry = regularBatchAgents.find(
             (agent) => agent.resolved.id === result.agentId || agent.resolved.type === result.agentType,
           );
           groupResults.push(
@@ -2521,10 +2570,26 @@ async function executeRetryBatches(
         }
       }
 
+      for (const entry of imagePromptAgents) {
+        const imagePromptContext = await resolveRetryImagePromptContext({
+          entry,
+          context: group.context,
+          conns,
+          chatMode,
+          chatMeta,
+        });
+        groupResults.push(
+          await executeAgent(entry.resolved, imagePromptContext, group.provider, group.model, entry.resolved.toolContext),
+        );
+      }
+
       for (const entry of toolAgents) {
+        const toolContext = isImagePromptRetryAgent(entry)
+          ? await resolveRetryImagePromptContext({ entry, context: group.context, conns, chatMode, chatMeta })
+          : group.context;
         const result = await executeAgent(
           entry.resolved,
-          group.context,
+          toolContext,
           group.provider,
           group.model,
           entry.resolved.toolContext,
@@ -4285,7 +4350,14 @@ export async function registerRetryAgentsRoute(app: FastifyInstance) {
                 }),
               ]
             : nonLorebookAgents.length > 0
-              ? await executeRetryBatches(agentContext, nonLorebookAgents, preGenerationAgentContext)
+              ? await executeRetryBatches(
+                  agentContext,
+                  nonLorebookAgents,
+                  preGenerationAgentContext,
+                  conns,
+                  chatMode,
+                  chatMeta,
+                )
               : [];
       const results = rawResults
         .map(markInvalidJsonAgentResult)
diff --git a/packages/server/src/services/agents/agent-executor.ts b/packages/server/src/services/agents/agent-executor.ts
index 61ccff858..5dd35206f 100644
--- a/packages/server/src/services/agents/agent-executor.ts
+++ b/packages/server/src/services/agents/agent-executor.ts
@@ -18,6 +18,7 @@ import type {
 import {
   characterTrackerLockKey,
   compactQuestProgressForContext,
+  customAgentHasCapability,
   DEFAULT_AGENT_CONTEXT_SIZE,
   DEFAULT_AGENT_MAX_TOKENS,
   DEFAULT_CUSTOM_AGENT_CONTEXT_SOURCES,
@@ -1056,6 +1057,7 @@ export async function executeAgentBatch(
   context: AgentContext,
   provider: BaseLLMProvider,
   model: string,
+  resolveAgentContext?: (config: AgentExecConfig, context: AgentContext) => AgentContext | Promise<AgentContext>,
 ): Promise<AgentResult[]> {
   if (configs.length === 0) return [];
   const isolatedConfigs = configs.filter(shouldRunAgentIndividually);
@@ -1075,7 +1077,7 @@ export async function executeAgentBatch(
     const isolatedSettled = await settleAgentJobsWithConcurrencyLimit(
       isolatedConfigs,
       AGENT_BATCH_FALLBACK_MAX_CONCURRENT,
-      (config) => executeAgent(config, context, provider, model),
+      async (config) => executeAgent(config, resolveAgentContext ? await resolveAgentContext(config, context) : context, provider, model),
     );
     return isolatedSettled.map((entry, index) =>
       entry.status === "fulfilled"
@@ -1095,9 +1097,13 @@ export async function executeAgentBatch(
     );
     const batchedConfigs = configs.filter((config) => !shouldRunAgentIndividually(config));
     const [batchedResults, isolatedSettled] = await Promise.all([
-      executeAgentBatch(batchedConfigs, context, provider, model),
+      executeAgentBatch(batchedConfigs, context, provider, model, resolveAgentContext),
       settleAgentJobsWithConcurrencyLimit(isolatedConfigs, AGENT_BATCH_FALLBACK_MAX_CONCURRENT, (config) =>
-        executeAgent(config, context, provider, model),
+        resolveAgentContext
+          ? Promise.resolve(resolveAgentContext(config, context)).then((agentContext) =>
+              executeAgent(config, agentContext, provider, model),
+            )
+          : executeAgent(config, context, provider, model),
       ),
     ]);
     const isolatedResults = isolatedSettled.map((entry, index) =>
@@ -1113,7 +1119,8 @@ export async function executeAgentBatch(
   }
   if (configs.length === 1) {
     logger.info(`[agent-batch] Only 1 agent (${configs[0]!.type}), running individually`);
-    return [await executeAgent(configs[0]!, context, provider, model)];
+    const agentContext = resolveAgentContext ? await resolveAgentContext(configs[0]!, context) : context;
+    return [await executeAgent(configs[0]!, agentContext, provider, model)];
   }
 
   const requestOptionGroups = new Map<string, AgentExecConfig[]>();
@@ -1131,7 +1138,7 @@ export async function executeAgentBatch(
     );
     const groupedResults: AgentResult[] = [];
     for (const group of requestOptionGroups.values()) {
-      groupedResults.push(...(await executeAgentBatch(group, context, provider, model)));
+      groupedResults.push(...(await executeAgentBatch(group, context, provider, model, resolveAgentContext)));
     }
     return groupedResults;
   }
@@ -1282,7 +1289,7 @@ export async function executeAgentBatch(
       const retrySettled = await settleAgentJobsWithConcurrencyLimit(
         failed,
         AGENT_BATCH_FALLBACK_MAX_CONCURRENT,
-        (config) => executeAgent(config, context, provider, model),
+        async (config) => executeAgent(config, resolveAgentContext ? await resolveAgentContext(config, context) : context, provider, model),
       );
       const retries: AgentResult[] = [];
       for (let i = 0; i < retrySettled.length; i++) {
@@ -1627,6 +1634,7 @@ function shouldRunAgentIndividually(config: Pick<AgentExecConfig, "type" | "sett
   // must not be merged into unrelated batched agent requests.
   return (
     config.type === "illustrator" ||
+    customAgentHasCapability(config.settings, "trigger_image_generation") ||
     config.type === "lorebook-keeper" ||
     resolveAgentResultType(config) === "text_rewrite" ||
     musicDjUsesJsonOnlyProvider(config) ||
@@ -1811,6 +1819,8 @@ function buildStandardAgentMessages(config: AgentExecConfig, template: string, c
     includeTrackerData: contextSources.trackerData,
     preserveAssistantResponseMarkup: resultType === "text_rewrite",
     outputFormatBlock: buildAgentOutputFormatBlock([config], context, renderedTemplates),
+    includeImagePromptInstructions:
+      config.type === "illustrator" || customAgentHasCapability(config.settings, "trigger_image_generation"),
   });
 }
 
@@ -2280,6 +2290,7 @@ function buildAgentMessages(
     includeTrackerData?: boolean;
     preserveAssistantResponseMarkup?: boolean;
     outputFormatBlock?: string;
+    includeImagePromptInstructions?: boolean;
   } = {},
 ): ChatMessage[] {
   // ── 1. System message — already contains <role>, <lore>, <agents>, and extras ──
@@ -2366,7 +2377,21 @@ function buildAgentMessages(
   // some models, so add a terminal user instruction for that agent type too.
   const outputFormatBlock = options.outputFormatBlock?.trim() ?? "";
   const requiresTerminalUserInstruction =
-    finalParts.length > 0 || contextAgentTypes.includes("echo-chamber") || !!outputFormatBlock;
+    finalParts.length > 0 ||
+    contextAgentTypes.includes("echo-chamber") ||
+    !!outputFormatBlock ||
+    (options.includeImagePromptInstructions === true &&
+      typeof context.memory._imagePromptInstructions === "string" &&
+      context.memory._imagePromptInstructions.trim().length > 0);
+
+  const lateImagePromptInstructions =
+    typeof context.memory._imagePromptInstructions === "string" ? context.memory._imagePromptInstructions.trim() : "";
+  if (options.includeImagePromptInstructions === true && lateImagePromptInstructions) {
+    finalParts.push("\n<image_prompting_instructions>");
+    finalParts.push("Apply these image-backend instructions when writing the provider-ready image prompt. Do not copy the instructions as prompt content.");
+    finalParts.push(lateImagePromptInstructions);
+    finalParts.push("</image_prompting_instructions>");
+  }
 
   if (requiresTerminalUserInstruction) {
     const instruction = "Now return the requested format(s).";
diff --git a/packages/server/src/services/agents/agent-pipeline.ts b/packages/server/src/services/agents/agent-pipeline.ts
index c35181245..d3fa9d586 100644
--- a/packages/server/src/services/agents/agent-pipeline.ts
+++ b/packages/server/src/services/agents/agent-pipeline.ts
@@ -39,6 +39,11 @@ export interface AgentInjection {
   text: string;
 }
 
+export type AgentContextResolver = (
+  agent: AgentExecConfig,
+  context: AgentContext,
+) => AgentContext | Promise<AgentContext>;
+
 /** Callback fired whenever an agent produces a result. */
 export type AgentResultCallback = (result: AgentResult) => void;
 
@@ -165,8 +170,11 @@ async function executeGroup(
   group: AgentGroup,
   context: AgentContext,
   onResult?: AgentResultCallback,
+  resolveAgentContext?: AgentContextResolver,
 ): Promise<AgentResult[]> {
-  const groupContext = buildAgentContext(group.agents[0]!, context);
+  const groupContext = resolveAgentContext
+    ? await resolveAgentContext(group.agents[0]!, buildAgentContext(group.agents[0]!, context))
+    : buildAgentContext(group.agents[0]!, context);
   // Separate tool-using agents (can't be batched) from regular agents. Spotify always
   // returns one JSON intent; deterministic host-side playback runs after parsing.
   const toolAgents = group.agents.filter((a) => shouldUseToolsDuringAgentExecution(a));
@@ -189,7 +197,7 @@ async function executeGroup(
 
   const batchResultsPromise =
     batchAgents.length > 0
-      ? executeAgentBatch(batchAgents, groupContext, group.provider, group.model).then((results) => {
+      ? executeAgentBatch(batchAgents, groupContext, group.provider, group.model, resolveAgentContext).then((results) => {
           for (const result of results) {
             safeOnResult(result);
           }
@@ -207,7 +215,10 @@ async function executeGroup(
     toolAgents,
     AGENT_GROUP_MAX_CONCURRENT_TOOL_CALLS,
     (agent) =>
-      executeAgent(agent, buildAgentContext(agent, context), agent.provider, agent.model, agent.toolContext).then((result) => {
+      (resolveAgentContext
+        ? Promise.resolve(resolveAgentContext(agent, buildAgentContext(agent, context)))
+        : Promise.resolve(buildAgentContext(agent, context))
+      ).then((agentContext) => executeAgent(agent, agentContext, agent.provider, agent.model, agent.toolContext)).then((result) => {
         safeOnResult(result);
         return result;
       }),
@@ -249,6 +260,7 @@ async function executePhase(
   phase: string,
   context: AgentContext,
   onResult?: AgentResultCallback,
+  resolveAgentContext?: AgentContextResolver,
 ): Promise<AgentResult[]> {
   const phaseAgents = agents.filter((a) => a.phase === phase);
   if (phaseAgents.length === 0) return [];
@@ -275,7 +287,7 @@ async function executePhase(
   const settled = await settleAgentJobsWithConcurrencyLimit(
     groups,
     AGENT_PHASE_MAX_CONCURRENT_GROUPS,
-    (group) => executeGroup(group, context, onResult),
+    (group) => executeGroup(group, context, onResult, resolveAgentContext),
   );
 
   const results: AgentResult[] = [];
@@ -337,9 +349,10 @@ export async function runPreGenerationAgents(
   context: AgentContext,
   onResult?: AgentResultCallback,
   agentTypeFilter?: (agentType: string) => boolean,
+  resolveAgentContext?: AgentContextResolver,
 ): Promise<AgentInjection[]> {
   const filtered = agentTypeFilter ? agents.filter((a) => agentTypeFilter(a.type)) : agents;
-  const results = await executePhase(filtered, "pre_generation", context, onResult);
+  const results = await executePhase(filtered, "pre_generation", context, onResult, resolveAgentContext);
 
   const injections: AgentInjection[] = [];
   for (const result of results) {
@@ -378,8 +391,9 @@ export async function runPostProcessingAgents(
   agents: ResolvedAgent[],
   context: AgentContext,
   onResult?: AgentResultCallback,
+  resolveAgentContext?: AgentContextResolver,
 ): Promise<AgentResult[]> {
-  return executePhase(agents, "post_processing", context, onResult);
+  return executePhase(agents, "post_processing", context, onResult, resolveAgentContext);
 }
 
 /**
@@ -389,8 +403,9 @@ export async function runParallelAgents(
   agents: ResolvedAgent[],
   context: AgentContext,
   onResult?: AgentResultCallback,
+  resolveAgentContext?: AgentContextResolver,
 ): Promise<AgentResult[]> {
-  return executePhase(agents, "parallel", context, onResult);
+  return executePhase(agents, "parallel", context, onResult, resolveAgentContext);
 }
 
 // ──────────────────────────────────────────────
@@ -417,6 +432,7 @@ export function createAgentPipeline(
   agents: ResolvedAgent[],
   baseContext: AgentContext,
   onResult?: AgentResultCallback,
+  resolveAgentContext?: AgentContextResolver,
 ) {
   const allResults: AgentResult[] = [];
   const preGenerationInjections: AgentInjection[] = [];
@@ -433,7 +449,13 @@ export function createAgentPipeline(
      * Returns context injection strings to prepend to the prompt.
      */
     async preGenerate(agentTypeFilter?: (agentType: string) => boolean): Promise<AgentInjection[]> {
-      const injections = await runPreGenerationAgents(agents, baseContext, wrappedOnResult, agentTypeFilter);
+      const injections = await runPreGenerationAgents(
+        agents,
+        baseContext,
+        wrappedOnResult,
+        agentTypeFilter,
+        resolveAgentContext,
+      );
       preGenerationInjections.push(...injections);
       return injections;
     },
@@ -444,7 +466,7 @@ export function createAgentPipeline(
      * base context without mainResponse (since it doesn't exist yet).
      */
     async runParallel(): Promise<AgentResult[]> {
-      const results = await runParallelAgents(agents, baseContext, wrappedOnResult);
+      const results = await runParallelAgents(agents, baseContext, wrappedOnResult, resolveAgentContext);
       parallelPhaseResults.push(...results);
       return results;
     },
@@ -464,7 +486,7 @@ export function createAgentPipeline(
         parallelResults: options.parallelResults ?? parallelPhaseResults,
       };
 
-      return runPostProcessingAgents(agents, fullContext, wrappedOnResult);
+      return runPostProcessingAgents(agents, fullContext, wrappedOnResult, resolveAgentContext);
     },
 
     /** All results collected so far. */
diff --git a/packages/server/src/services/game/game-asset-generation.ts b/packages/server/src/services/game/game-asset-generation.ts
index 9fe95768c..edc6ea552 100644
--- a/packages/server/src/services/game/game-asset-generation.ts
+++ b/packages/server/src/services/game/game-asset-generation.ts
@@ -33,6 +33,7 @@ import type { ImageGenerationSize } from "../image/image-generation-settings.js"
 import { compileImagePrompt } from "../image/image-prompt-compiler.js";
 import { loadGameStoryboardImagePrompt } from "../image/game-storyboard-image-prompt.js";
 import { SPATIAL_LOCATION_REFERENCE_PROMPT_LINE } from "../image/spatial-location-reference.js";
+import { compactImagePromptInstructions } from "../sidecar/scene-analyzer.js";
 
 const NPC_AVATAR_DIR = join(DATA_DIR, "avatars", "npc");
 const CHAT_BACKGROUND_DIR = join(DATA_DIR, "backgrounds");
@@ -990,7 +991,7 @@ async function buildBackgroundRawPrompt(req: BackgroundGenRequest): Promise<stri
     ? await loadPrompt(req.promptOverridesStorage, GAME_BACKGROUND, backgroundVars)
     : GAME_BACKGROUND.defaultBuilder(backgroundVars);
   const imagePromptInstructionsLine = req.imagePromptInstructions?.trim()
-    ? `User image instructions: ${req.imagePromptInstructions.trim().replace(/\s+/g, " ").slice(0, 5000)}`
+    ? `User image instructions: ${compactImagePromptInstructions(req.imagePromptInstructions)}`
     : "";
   return imagePromptInstructionsLine && !rawPrompt.includes(imagePromptInstructionsLine)
     ? `${rawPrompt}\n${imagePromptInstructionsLine}`
@@ -1124,7 +1125,7 @@ async function buildSceneIllustrationRawPrompt(req: SceneIllustrationGenRequest)
   const referenceImages = sceneIllustrationReferenceImagesForProvider(req);
   const characterReferenceImagesAttached = referenceImages.length > (req.locationReferenceImageAttached ? 1 : 0);
   const imagePromptInstructionsLine = req.imagePromptInstructions?.trim()
-    ? `User image instructions: ${req.imagePromptInstructions.trim().replace(/\s+/g, " ").slice(0, 5000)}`
+    ? `User image instructions: ${compactImagePromptInstructions(req.imagePromptInstructions)}`
     : "";
   const useGamePromptTemplate = req.useGamePromptTemplate !== false;
   const scopedScenePrompt = req.prompt.trim();
diff --git a/packages/server/src/services/generation/conversation-selfie-command-runtime.ts b/packages/server/src/services/generation/conversation-selfie-command-runtime.ts
index f3b562748..b8fabf1c6 100644
--- a/packages/server/src/services/generation/conversation-selfie-command-runtime.ts
+++ b/packages/server/src/services/generation/conversation-selfie-command-runtime.ts
@@ -16,6 +16,7 @@ import { generateImage, saveImageToDisk } from "../image/image-generation.js";
 import { loadImageGenerationUserSettings } from "../image/image-generation-settings.js";
 import { generateIllustratorImageVariants } from "../image/illustrator-image-variants.js";
 import { resolveConversationSelfieSystemPrompt } from "../conversation/selfie-prompt.js";
+import { appendImagePromptInstructions } from "./image-prompt-instructions.js";
 import type { CharacterCommand, SelfieCommand } from "../conversation/character-commands.js";
 import { createGalleryStorage } from "../storage/gallery.storage.js";
 import { createCharacterGalleryStorage } from "../storage/character-gallery.storage.js";
@@ -201,6 +202,10 @@ async function generateSelfie(
   const selfieSystemPrompt = styleGuidance
     ? `${baseSelfieSystemPrompt}${formatImageStylePromptGuidance(styleGuidance)}`
     : baseSelfieSystemPrompt;
+  const selfieSystemPromptWithImageInstructions = appendImagePromptInstructions(
+    selfieSystemPrompt,
+    imgConnFull.imagePromptInstructions,
+  );
   const userPrompt = args.command.context
     ? `Context for the selfie: ${args.command.context}`
     : `Generate a casual selfie of ${args.charName} based on the current conversation context.`;
@@ -213,7 +218,7 @@ async function generateSelfie(
     [
       {
         role: "system",
-        content: selfieSystemPrompt,
+        content: selfieSystemPromptWithImageInstructions,
       },
       {
         role: "user",
diff --git a/packages/server/src/services/generation/illustrator-manual-prompt-generation.ts b/packages/server/src/services/generation/illustrator-manual-prompt-generation.ts
index 375172e6e..79ebbbf7a 100644
--- a/packages/server/src/services/generation/illustrator-manual-prompt-generation.ts
+++ b/packages/server/src/services/generation/illustrator-manual-prompt-generation.ts
@@ -238,6 +238,7 @@ export function buildManualIllustratorPromptMessages(args: {
   contextSize: unknown;
   selectedPromptTemplate?: string;
   styleInstruction?: string;
+  imagePromptInstructions?: string;
 }): ChatMessage[] {
   const promptModeInstruction = normalizeManualIllustratorPromptModeInstruction(args.selectedPromptTemplate ?? "");
   const systemPrompt = [
@@ -253,6 +254,9 @@ export function buildManualIllustratorPromptMessages(args: {
     args.styleInstruction
       ? `Additional Image Style instruction for the image prompt you write: ${args.styleInstruction}\nCombine it with the selected Illustrator prompt mode. It may refine rendering and visual treatment, but it must not replace or weaken the selected format, layout, framing, or text requirements.`
       : "No visual style profile is selected. Infer only the visual treatment supported by the scene context.",
+    args.imagePromptInstructions
+      ? `<image_prompting_instructions>\nApply these image-backend instructions when writing the provider-ready prompt. They are instructions, not text to copy into the prompt:\n${args.imagePromptInstructions}\n</image_prompting_instructions>`
+      : "",
     buildCharacterPersonaContext(args.context),
   ].join("\n\n");
   const messages: ChatMessage[] = [
@@ -295,6 +299,7 @@ export async function writeManualIllustratorPromptPlan(args: {
   illustratorAgent: ResolvedAgent;
   context: AgentContext;
   styleInstruction?: string;
+  imagePromptInstructions?: string;
   signal?: AbortSignal;
   debugLog?: (message: string, ...args: unknown[]) => void;
 }): Promise<ManualIllustratorPromptResult> {
@@ -309,6 +314,7 @@ export async function writeManualIllustratorPromptPlan(args: {
     contextSize: args.illustratorAgent.settings.contextSize,
     selectedPromptTemplate,
     styleInstruction: args.styleInstruction,
+    imagePromptInstructions: args.imagePromptInstructions,
   });
   args.debugLog?.(
     "[debug/illustrator/manual-illustration-prompt] messages:\n%s",
diff --git a/packages/server/src/services/generation/illustrator-prompt-runtime.ts b/packages/server/src/services/generation/illustrator-prompt-runtime.ts
index 59fd91a7d..3ed5278a2 100644
--- a/packages/server/src/services/generation/illustrator-prompt-runtime.ts
+++ b/packages/server/src/services/generation/illustrator-prompt-runtime.ts
@@ -13,6 +13,7 @@ export type IllustratorPromptConnection = FallbackConnection & {
   imageGenerationSource?: string | null;
   imageService?: string | null;
   imageEndpointId?: string | null;
+  imagePromptInstructions?: string | null;
   comfyuiWorkflow?: string | null;
 };
 
diff --git a/packages/server/src/services/generation/image-prompt-instructions.ts b/packages/server/src/services/generation/image-prompt-instructions.ts
new file mode 100644
index 000000000..1d13d3e83
--- /dev/null
+++ b/packages/server/src/services/generation/image-prompt-instructions.ts
@@ -0,0 +1,8 @@
+import { normalizeImagePromptInstructions } from "@marinara-engine/shared";
+
+export function appendImagePromptInstructions(prompt: string, instructions: unknown): string {
+  const normalizedInstructions = normalizeImagePromptInstructions(instructions);
+  return normalizedInstructions
+    ? `${prompt}\n\n<image_prompting_instructions>\nApply these image-backend instructions when writing the provider-ready prompt. They are instructions, not text to copy into the prompt:\n${normalizedInstructions}\n</image_prompting_instructions>`
+    : prompt;
+}
\ No newline at end of file
diff --git a/packages/server/src/services/sidecar/scene-analyzer.ts b/packages/server/src/services/sidecar/scene-analyzer.ts
index e8b0fd64c..04bb9ba0e 100644
--- a/packages/server/src/services/sidecar/scene-analyzer.ts
+++ b/packages/server/src/services/sidecar/scene-analyzer.ts
@@ -11,6 +11,7 @@ import {
   LOCATION_KINDS,
   MUSIC_GENRES,
   MUSIC_INTENSITIES,
+  MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH,
   type HudWidget,
   type GameNpc,
   type GameActiveState,
@@ -115,8 +116,8 @@ function buildBackgroundOptions(ctx?: SceneAnalyzerContext): string[] {
   return options;
 }
 
-function compactImagePromptInstructions(value: string | null | undefined): string {
-  return (value ?? "").trim().replace(/\s+/g, " ").slice(0, 5000);
+export function compactImagePromptInstructions(value: string | null | undefined): string {
+  return (value ?? "").trim().replace(/\s+/g, " ").slice(0, MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH);
 }
 
 function compactPromptLabel(value: string | null | undefined): string {
diff --git a/packages/server/src/services/storage/connections.storage.ts b/packages/server/src/services/storage/connections.storage.ts
index 2741fe7ef..4797f4a65 100644
--- a/packages/server/src/services/storage/connections.storage.ts
+++ b/packages/server/src/services/storage/connections.storage.ts
@@ -141,6 +141,7 @@ export function createConnectionsStorage(db: DB) {
         comfyuiWorkflow: input.comfyuiWorkflow ?? null,
         imageService: input.imageService ?? null,
         imageEndpointId: input.imageEndpointId ?? null,
+        imagePromptInstructions: input.imagePromptInstructions ?? null,
         imageGenerationQuality: input.imageGenerationQuality ?? "auto",
         videoGenerationSource: input.videoGenerationSource ?? null,
         videoService: input.videoService ?? null,
@@ -283,6 +284,9 @@ export function createConnectionsStorage(db: DB) {
       if (data.imageEndpointId !== undefined) {
         updateFields.imageEndpointId = data.imageEndpointId;
       }
+      if (data.imagePromptInstructions !== undefined) {
+        updateFields.imagePromptInstructions = data.imagePromptInstructions;
+      }
       if (data.imageGenerationQuality !== undefined) {
         updateFields.imageGenerationQuality = data.imageGenerationQuality;
       }
@@ -418,6 +422,7 @@ export function createConnectionsStorage(db: DB) {
         comfyuiWorkflow: source.comfyuiWorkflow,
         imageService: source.imageService,
         imageEndpointId: source.imageEndpointId,
+        imagePromptInstructions: source.imagePromptInstructions,
         imageGenerationQuality: source.imageGenerationQuality,
         videoGenerationSource: source.videoGenerationSource,
         videoService: source.videoService,
diff --git a/packages/shared/src/constants/defaults.ts b/packages/shared/src/constants/defaults.ts
index 6efa97892..edb9df408 100644
--- a/packages/shared/src/constants/defaults.ts
+++ b/packages/shared/src/constants/defaults.ts
@@ -18,6 +18,15 @@ export const DEFAULT_CONNECTION_ID = "__default_openrouter__";
 /** Shared width for cached background thumbnails used by compact client surfaces. */
 export const BACKGROUND_THUMBNAIL_WIDTH = 320;
 
+/** Maximum length for image-backend instructions stored on an image connection. */
+export const MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH = 20_000;
+
+export function normalizeImagePromptInstructions(value: unknown): string | null {
+  if (typeof value !== "string") return null;
+  const trimmed = value.trim().slice(0, MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH);
+  return trimmed || null;
+}
+
 /** Default system prompt for AI-backed translation. */
 export const DEFAULT_TRANSLATION_SYSTEM_PROMPT =
   "You are a translator. Translate the given text accurately into {{targetLanguage}}, preserving formatting, markdown, and any special characters like *asterisks* for actions. Output ONLY the translated text, nothing else -- no explanations, no extra commentary.";
diff --git a/packages/shared/src/schemas/connection.schema.ts b/packages/shared/src/schemas/connection.schema.ts
index 16890955c..80ab5a12c 100644
--- a/packages/shared/src/schemas/connection.schema.ts
+++ b/packages/shared/src/schemas/connection.schema.ts
@@ -3,6 +3,7 @@
 // ──────────────────────────────────────────────
 import { z } from "zod";
 import { IMAGE_GENERATION_QUALITIES } from "../types/connection.js";
+import { MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH } from "../constants/defaults.js";
 
 export const apiProviderSchema = z.enum([
   "openai",
@@ -69,6 +70,7 @@ export const createConnectionSchema = z.object({
   comfyuiWorkflow: z.string().nullable().default(null),
   imageService: z.string().nullable().default(null),
   imageEndpointId: z.string().nullable().default(null),
+  imagePromptInstructions: z.string().trim().max(MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH).nullable().default(null),
   imageGenerationQuality: imageGenerationQualitySchema.default("auto"),
   videoGenerationSource: z.string().nullable().default(null),
   videoService: z.string().nullable().default(null),
diff --git a/packages/shared/src/schemas/scene-analysis.schema.ts b/packages/shared/src/schemas/scene-analysis.schema.ts
index 852a557e3..5eaf9580b 100644
--- a/packages/shared/src/schemas/scene-analysis.schema.ts
+++ b/packages/shared/src/schemas/scene-analysis.schema.ts
@@ -1,4 +1,5 @@
 import { z } from "zod";
+import { MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH } from "../constants/defaults.js";
 import type { GameNpc, HudWidget } from "../types/game.js";
 import type { GameActiveState } from "../types/game.js";
 import type { SceneSpotifyTrackCandidate } from "../types/sidecar.js";
@@ -44,7 +45,7 @@ export const sceneAnalysisContextSchema = z.object({
   canGenerateBackgrounds: z.boolean().optional(),
   canGenerateIllustrations: z.boolean().optional(),
   artStylePrompt: z.string().nullable().optional(),
-  imagePromptInstructions: z.string().max(5_000).nullable().optional(),
+  imagePromptInstructions: z.string().max(MAX_IMAGE_PROMPT_INSTRUCTIONS_LENGTH).nullable().optional(),
 });
 
 export const sceneAnalysisRequestSchema = z.object({
diff --git a/packages/shared/src/types/connection.ts b/packages/shared/src/types/connection.ts
index eab82eb64..1d183630d 100644
--- a/packages/shared/src/types/connection.ts
+++ b/packages/shared/src/types/connection.ts
@@ -69,6 +69,8 @@ export interface APIConnection {
   imageService: string | null;
   /** For endpoint-based image services (e.g. RunPod Serverless): the endpoint ID sent alongside the base URL. */
   imageEndpointId: string | null;
+  /** Instructions applied by an extra default-language-model call before image generation. */
+  imagePromptInstructions: string | null;
   /** OpenAI GPT Image quality saved for this connection. */
   imageGenerationQuality: ImageGenerationQuality;
   /** Explicit video backend selection for video-generation connections (e.g. Gemini Omni). */