From 90a9e34c7e17b83118576596d890f232f2e240cf Mon Sep 17 00:00:00 2001
From: Marc Mintel <marc@mintel.me>
Date: Sun, 22 Feb 2026 18:07:50 +0100
Subject: [PATCH] fix(journaling): enforce stricter LLM evaluation rules for
 YouTube video selection

---
 packages/journaling/src/agent.ts | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/packages/journaling/src/agent.ts b/packages/journaling/src/agent.ts
index dcf4eaa..bcbba86 100644
--- a/packages/journaling/src/agent.ts
+++ b/packages/journaling/src/agent.ts
@@ -195,11 +195,11 @@ Return JSON: { "facts": [ { "statement": "...", "source": "Organization Name Onl
 RULES:
 1. Extract only the 2-4 most important technical or business keywords from the provided text.
 2. Ignore all markdown syntax, frontmatter (---), titles, and descriptions.
-3. Keep the query generic enough to find popular educational tech videos.
+3. Keep the query generic enough to find popular educational tech videos, BUT ensure it specifically targets the core technical subject. Append "tutorial" or "b2b explanation" if necessary to find high-quality content.
 4. DO NOT append specific channel names (e.g., "Fireship", "Vercel") to the query.
 5. DO NOT USE QUOTES IN THE QUERY.
 
-Return a JSON object with a single string field "query". Example: {"query": "core web vitals performance"}`,
+Return a JSON object with a single string field "query". Example: {"query": "core web vitals performance tutorial"}`,
         },
         {
           role: "user",
@@ -243,15 +243,15 @@ Return a JSON object with a single string field "query". Example: {"query": "cor
       }
 
       // Step 3: Ask the LLM to evaluate the relevance of the found videos
-      const evalPrompt = `You are a strict technical evaluator. You must select the MOST RELEVANT educational tech video from the list below based on this context: "${topic.slice(0, 500)}..."
+      const evalPrompt = `You are a strict technical evaluator. You must select the MOST RELEVANT educational tech video from the list below based on this core article context: "${topic.slice(0, 800)}..."
 
 Videos:
-${ytVideos.map((v, i) => `[ID: ${i}] Title: "${v.title}" | Channel: "${v.channel}" | Snippet: "${v.snippet || "none"}"`).join("\n")}
+${ytVideos.map((v, i) => `[ID: ${i}] Title: "${v.title}" | Channel: "${v.channel}" | Snippet: "${v.snippet || 'none'}"`).join("\n")}
 
 RULES:
-1. The video MUST be highly relevant to the context.
-2. The channel SHOULD be a high-quality tech, development, or professional B2B channel (e.g., Google Developers, Vercel, Theo - t3.gg, Fireship, Syntax, ByteByteGo, IBM Technology, McKinsey, Gartner, Deloitte). AVOID gaming, generic vlogs, clickbait, or unrelated topics.
-3. If none are truly relevant, return -1.
+1. The video MUST be highly relevant to the EXACT technical topic of the context.
+2. The channel SHOULD be a high-quality tech, development, or professional B2B channel (e.g., Google Developers, Vercel, Theo - t3.gg, Fireship, Syntax, ByteByteGo, IBM Technology, McKinsey, Gartner, Deloitte). AVOID gaming, generic vlogs, clickbait, off-topic podcasts, or unrelated topics.
+3. If none of the videos are strictly relevant to the core technical or business subject (e.g. they are just casually mentioning the word), YOU MUST RETURN -1. Be extremely critical. Do not just pick the "best of the worst".
 4. If one is highly relevant, return its ID number.
 
 Return ONLY a JSON object: {"bestVideoId": number}`;
@@ -343,7 +343,7 @@ CRITICAL: Do NOT provide more than 2 trendsKeywords. Keep it extremely focused.`
     try {
       let parsed = JSON.parse(
         response.choices[0].message.content ||
-          '{"trendsKeywords": [], "dcVariables": []}',
+        '{"trendsKeywords": [], "dcVariables": []}',
       );
       if (Array.isArray(parsed)) {
         parsed = parsed[0] || { trendsKeywords: [], dcVariables: [] };