<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Norsica Blog</title>
    <link>https://www.norsica.jp/blog</link>
    <atom:link href="https://www.norsica.jp/feed.xml" rel="self" type="application/rss+xml" />
    <description>Notes on AI-Native development and building with LLMs by Shinsuke KAGAWA.</description>
    <language>en</language>
    <lastBuildDate>Fri, 20 Nov 2026 01:30:00 GMT</lastBuildDate>
    <item>
      <title>Strict Boundaries, Free Path: How I Design a Software Factory</title>
      <link>https://www.norsica.jp/blog/how-i-design-a-software-factory</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/how-i-design-a-software-factory</guid>
      <pubDate>Fri, 20 Nov 2026 01:30:00 GMT</pubDate>
      <description>Replace each human check with a mechanism that fixes boundaries and evidence, leave the path to agents, and a software factory merges without human code review.</description>
      <category>Software Factory</category>
      <category>Code Review</category>
      <category>AGENTS.md</category>
      <category>Merge Gate</category>
      <category>Retrospective</category>
    </item>
    <item>
      <title>I Didn&apos;t Give My AI Team Its Epics. I Gave It a Definition of One.</title>
      <link>https://www.norsica.jp/blog/i-gave-my-ai-team-a-definition-of-an-epic</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/i-gave-my-ai-team-a-definition-of-an-epic</guid>
      <pubDate>Sat, 14 Nov 2026 07:30:00 GMT</pubDate>
      <description>Software factories start from scope that people cut. My AI team cuts its own. I gave it a definition of an epic, and it came back with twelve.</description>
      <category>Software Factory</category>
      <category>Epics</category>
      <category>User Journey</category>
      <category>MVP</category>
      <category>AI Team</category>
    </item>
    <item>
      <title>I Never Asked My AI Team to Build All of It</title>
      <link>https://www.norsica.jp/blog/i-never-asked-my-ai-team-to-build-all-of-it</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/i-never-asked-my-ai-team-to-build-all-of-it</guid>
      <pubDate>Fri, 30 Oct 2026 12:00:00 GMT</pubDate>
      <description>My AI team kept failing acceptance criteria it had written for itself, and every miss became more work. The criteria were the one thing nobody was checking.</description>
      <category>Acceptance Criteria</category>
      <category>AI Team</category>
      <category>Orchestrator</category>
      <category>Goals</category>
      <category>Scope</category>
    </item>
    <item>
      <title>Read the Logs Before You Rewrite the Prompt</title>
      <link>https://www.norsica.jp/blog/read-the-logs-before-you-rewrite-the-prompt</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/read-the-logs-before-you-rewrite-the-prompt</guid>
      <pubDate>Thu, 22 Oct 2026 12:00:00 GMT</pubDate>
      <description>I halved my AI orchestrator&apos;s hourly cost by moving routine steps into scripts. When the scripts broke, the logs showed exactly where.</description>
      <category>Agents</category>
      <category>Orchestrator</category>
      <category>Prompt</category>
      <category>Logs</category>
      <category>Cost</category>
    </item>
    <item>
      <title>I Took the Wheel Back From My AI Team. Then I Gave It Back.</title>
      <link>https://www.norsica.jp/blog/i-took-the-wheel-back-from-my-ai-team</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/i-took-the-wheel-back-from-my-ai-team</guid>
      <pubDate>Wed, 21 Oct 2026 13:07:00 GMT</pubDate>
      <description>See how six corrections to an AI team led to a search-first planning rule, then a night when the team found and fixed three problems on its own.</description>
      <category>Agents</category>
      <category>Orchestrator</category>
      <category>Goal</category>
      <category>Planning</category>
      <category>Research</category>
    </item>
    <item>
      <title>My AI Team Followed Three Rules. It Kept Two.</title>
      <link>https://www.norsica.jp/blog/my-ai-team-followed-three-rules-it-kept-two</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/my-ai-team-followed-three-rules-it-kept-two</guid>
      <pubDate>Sun, 04 Oct 2026 05:14:00 GMT</pubDate>
      <description>Watch an AI team judge three rules it followed, keep two, and delete one with no proven benefit as goals and retrospectives help it manage its work.</description>
      <category>Agents</category>
      <category>Agile</category>
      <category>Retrospective</category>
      <category>Orchestrator</category>
      <category>Goals</category>
    </item>
    <item>
      <title>The Day I Left the Team</title>
      <link>https://www.norsica.jp/blog/the-day-i-left-the-team</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/the-day-i-left-the-team</guid>
      <pubDate>Wed, 30 Sep 2026 09:01:00 GMT</pubDate>
      <description>Handed a product to six AI agents and left the team. Anchoring broke their decisions until I reordered the steps; 11 agreements have passed review since.</description>
      <category>AI Agents</category>
      <category>Anchoring</category>
      <category>Orchestrator</category>
      <category>Claude Code</category>
      <category>GPT-6 Astra</category>
    </item>
    <item>
      <title>What Retrieval Still Hasn&apos;t Decided</title>
      <link>https://www.norsica.jp/blog/what-retrieval-still-hasnt-decided</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/what-retrieval-still-hasnt-decided</guid>
      <pubDate>Sun, 20 Sep 2026 11:24:00 GMT</pubDate>
      <description>Split post-retrieval into rerank, filter, and compress, three different judgments measured on live runs. Dedup cut 0.2% of characters, so I didn&apos;t ship it.</description>
      <category>Reranking</category>
      <category>RAG</category>
      <category>Evidence Filtering</category>
      <category>Semantic Deduplication</category>
      <category>CLI</category>
    </item>
    <item>
      <title>Switching from GPT-5.6 Sol to GPT-6 Astra: Start with Medium Effort</title>
      <link>https://www.norsica.jp/blog/switching-from-sol-to-astra</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/switching-from-sol-to-astra</guid>
      <pubDate>Sat, 05 Sep 2026 04:10:00 GMT</pubDate>
      <description>Start with GPT-6 Astra medium: it cost 19% less than Sol high in my Codex comparison. High built better code, but medium found a bug its review missed.</description>
      <category>GPT-6 Astra</category>
      <category>GPT-5.6 Sol</category>
      <category>Codex CLI</category>
      <category>Agent Workflows</category>
      <category>Galley</category>
    </item>
    <item>
      <title>Was the Monorepo Really Best for Coding Agents?</title>
      <link>https://www.norsica.jp/blog/was-the-monorepo-really-best-for-coding-agents</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/was-the-monorepo-really-best-for-coding-agents</guid>
      <pubDate>Thu, 20 Aug 2026 07:14:00 GMT</pubDate>
      <description>Coding agents now reach across repositories, so the monorepo advantage moved from retrieval to governance: who owns shared rules, and which one wins a conflict.</description>
      <category>Monorepo</category>
      <category>Coding Agent</category>
      <category>Agent Harness</category>
      <category>Multi-Repository</category>
      <category>AI-Native Development</category>
    </item>
    <item>
      <title>Reasoning Effort Is Not a Quality Setting</title>
      <link>https://www.norsica.jp/blog/reasoning-effort-is-not-a-quality-setting</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/reasoning-effort-is-not-a-quality-setting</guid>
      <pubDate>Wed, 05 Aug 2026 12:52:00 GMT</pubDate>
      <description>Across 28 runs, higher reasoning effort expanded what coding agents explored, not what they delivered. Its value depended on the phase and the model.</description>
      <category>Coding Agent</category>
      <category>Reasoning Effort</category>
      <category>Agent Workflow</category>
      <category>Code Review</category>
      <category>Overengineering</category>
    </item>
    <item>
      <title>When Better Models Make Old Agent Workflows Worse</title>
      <link>https://www.norsica.jp/blog/when-better-models-make-old-agent-workflows-worse</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/when-better-models-make-old-agent-workflows-worse</guid>
      <pubDate>Sun, 02 Aug 2026 14:10:00 GMT</pubDate>
      <description>Workflow rules written for older model failures make capable coding agents do extra work. Be strict about boundaries and evidence, flexible about the path.</description>
      <category>Agent Workflow</category>
      <category>Coding Agent</category>
      <category>Boundary Constraint</category>
      <category>Prompt Engineering</category>
      <category>Technical Debt</category>
    </item>
    <item>
      <title>What Makes Development AI-Native?</title>
      <link>https://www.norsica.jp/blog/what-makes-development-ai-native</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/what-makes-development-ai-native</guid>
      <pubDate>Wed, 11 Feb 2026 08:34:00 GMT</pubDate>
      <description>AI-Native development requires four structural pillars most organizations lack. Assess your readiness with a 100-point diagnostic criteria.</description>
      <category>AI-Native</category>
      <category>Context Closure</category>
      <category>LLM</category>
      <category>Development Process</category>
      <category>Guardrails</category>
    </item>
    <item>
      <title>Planning Is the Real Superpower of Agentic Coding</title>
      <link>https://www.norsica.jp/blog/planning-superpower-agentic-coding</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/planning-superpower-agentic-coding</guid>
      <pubDate>Mon, 26 Jan 2026 06:05:00 GMT</pubDate>
      <description>Turn LLM execution into verification with explicit planning. Research shows this approach improves task success by 33% in complex agentic coding scenarios.</description>
      <category>Agentic Coding</category>
      <category>Planning</category>
      <category>LLM</category>
      <category>Task Decomposition</category>
      <category>LLM Agents</category>
    </item>
    <item>
      <title>The Leadership Bottleneck in AI-Native Development</title>
      <link>https://www.norsica.jp/blog/leadership-bottleneck-ai-native-development</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/leadership-bottleneck-ai-native-development</guid>
      <pubDate>Wed, 21 Jan 2026 09:20:00 GMT</pubDate>
      <description>AI coding tools raise output while outcomes stay flat. The bottleneck has moved to context and decisions, and AI-native leadership means leaders change first.</description>
      <category>AI-Native</category>
      <category>Leadership</category>
      <category>AI-Native Transformation</category>
      <category>Organizational Design</category>
      <category>Context Engineering</category>
    </item>
    <item>
      <title>Stop Putting Everything in AGENTS.md</title>
      <link>https://www.norsica.jp/blog/stop-putting-everything-in-agents-md</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/stop-putting-everything-in-agents-md</guid>
      <pubDate>Mon, 19 Jan 2026 05:15:00 GMT</pubDate>
      <description>Learn where to write LLM rules, what to write, and how to verify they work. Keep AGENTS.md lean while ensuring your rules actually stick.</description>
      <category>AGENTS.md</category>
      <category>Agentic Coding</category>
      <category>Rule Externalization</category>
      <category>LLM</category>
      <category>Prompt Engineering</category>
    </item>
    <item>
      <title>Why LLMs Are Bad at &apos;First Try&apos; and Great at Verification</title>
      <link>https://www.norsica.jp/blog/llm-verification-over-generation</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/llm-verification-over-generation</guid>
      <pubDate>Mon, 12 Jan 2026 09:59:00 GMT</pubDate>
      <description>Stop optimizing for first-shot perfection. LLMs perform better at verifying and improving existing artifacts than at controlled first-time generation.</description>
      <category>LLM</category>
      <category>Verification</category>
      <category>Workflow Design</category>
      <category>Context Engineering</category>
      <category>AGENTS.md</category>
    </item>
    <item>
      <title>Building a Local RAG for Agentic Coding: From Fixed Chunks to Semantic Search with Keyword Boost</title>
      <link>https://www.norsica.jp/blog/local-rag-agentic-coding</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/local-rag-agentic-coding</guid>
      <pubDate>Mon, 12 Jan 2026 09:35:00 GMT</pubDate>
      <description>Started with a simple RAG for MCP, then ended up implementing semantic chunking and rethinking hybrid search. For people who&apos;ve built RAG systems and started hitting quality limits.</description>
      <category>RAG</category>
      <category>Semantic Chunking</category>
      <category>MCP</category>
      <category>Agentic Coding</category>
      <category>Hybrid Search</category>
    </item>
    <item>
      <title>Rethinking Team Development in the Age of LLMs</title>
      <link>https://www.norsica.jp/blog/rethinking-team-development-llm-era</link>
      <guid isPermaLink="true">https://www.norsica.jp/blog/rethinking-team-development-llm-era</guid>
      <pubDate>Mon, 12 Jan 2026 09:35:00 GMT</pubDate>
      <description>Code review moves to specs, plans, and rules. Design the whole Spec → Plan → Test → Implementation cycle, and scale by adding small teams, not people.</description>
      <category>Code Review</category>
      <category>Small Teams</category>
      <category>Process Design</category>
      <category>Vertical Slicing</category>
      <category>LLM</category>
    </item>
  </channel>
</rss>
