# Your Best Prompts Are Living in Slack. That Is Costing You More Than You Think > Source: > Published: 2026-07-08 16:01:02+00:00 Part 1 of “The Prompt Library in the AI Development Lifecycle” why this is quietly becoming the most underrated asset on your data team, and what it actually looks like inside Snowflake. A few weeks ago I watched a data engineer spend twenty to thirty minutes reconstructing a prompt. Not writing a new one. Reconstructing one she had already written, three weeks earlier, for the exact same task. An incremental merge pattern into a Snowflake target table. She knew it existed. She remembered getting it almost right. She just could not find it. It was buried somewhere in a Slack thread, under forty other messages about an unrelated release. She eventually gave up and rewrote it from scratch One thing I want to call out early, because people mix this up a lot, a prompt library is not the same thing as a skill, and both are not the same thing as an SDD. Three different layers, each doing a different job. If you have followed my SDD writing before [[Article Link]](https://medium.com/towards-artificial-intelligence/the-missing-layer-in-ai-data-pipelines-why-spec-driven-development-matters-244b48aeb7a6?sharedUserId=rahulsahay123), you will probably see where this is going already — a prompt library is not some separate idea, it is basically the missing half of spec-driven development. I will come back to this properly a bit later in the article, but keep it in mind as you read — the two are joined at the hip, not competing concepts. Every data team I talk to right now is deep into AI-assisted development. Cortex Code, Copilot, ChatGPT, whatever the tool of choice is — prompts are flying everywhere. Slack messages, sticky notes, personal notebooks, half-remembered phrasing that worked once and was never written down again What almost nobody has is a system for it. Think about how software engineering got here. Nobody writes a sort function from scratch anymore. We have libraries, patterns, reusable components — decades of collective effort turned into something you just import. Prompts are heading the exact same direction, except most organisations are still at the “everyone writes their own sort function” stage. A **Prompt Library** is simply the fix for that. A curated, versioned, organised collection of prompts your team actually reuses for code generation, data engineering, analytics, governance, whatever the task calls for. I put this list together after enough conversations that the pattern stopped feeling like coincidence. If you have read anything I have written about Spec-Driven Development (SDD) , you already know my position on this. The SDD is the **what** — business keys, relationships, grain, governance tags, acceptance criteria. The Prompt Library is the **how** — the reusable, versioned way your team actually asks the AI to act on that spec. A good library prompt does not restate business logic ``` Read the SDD file at {sdd_path}.Using section {sdd_section}, generate the {artifact_type} for {entity}.Verify against acceptance criteria {ac_reference} when done. ``` The temptation is to say “let’s just make a shared Google Doc.” That gets you to level two of five, and it stalls there fast — nobody updates a doc nobody owns. The version that actually holds up lives closer to your data platform itself. A simple table, honestly (though git repo is also an option): CREATE TABLE PROMPT_LIBRARY.PUBLIC.PROMPTS ( PROMPT_ID VARCHAR PRIMARY KEY, CATEGORY VARCHAR, PROMPT_NAME VARCHAR, PROMPT_TEMPLATE VARCHAR, PARAMETERS VARIANT, VERSION INTEGER, CREATED_BY VARCHAR, USAGE_COUNT INTEGER DEFAULT 0, STATUS VARCHAR DEFAULT ‘ACTIVE’ ); Most teams sit somewhere on this ladder without realising it: Do not try to build the level-5 version on day one. Start smaller than feels satisfying: The intelligence layer is the fun part to talk about. The cataloguing is the part that actually pays back first. This one was Part 1 ie the **why**. Three more still to come, same sequence as those four questions I mentioned earlier: Part 2 is the **how → **actual architecture, tooling, the YAML format, CI/CD pipelines, and the governance model that keeps all this trustworthy at real scale. Part 3 is the **what** → the actual reusable patterns, task by task, including how they reference an SDD instead of hardcoding business logic inside the prompt itself. Part 4 is the **prove** → instrumentation, A/B testing, executive dashboards, the full business case laid out properly with numbers. *Next up — Part 2. We go from strategy into the actual build.* A Prompt Library is not a nice-to-have side project. It is becoming table stakes for teams serious about AI-driven development — and for Snowflake users specifically, the tight fit between Cortex AI, semantic views, and native governance makes it a natural place to host one. Every prompt someone refines, every pattern someone captures, makes the next person’s AI interaction a little faster and a little more consistent. That compounds. Quietly, but it compounds. The real question is not whether your team needs a prompt library. It is how much you are already losing every single day without one. I spend most of my time exploring how AI is transforming data engineering, software delivery, and consulting. As a Snowflake SME, Snowflake CoCo Champion, and practitioner working across modern data platforms and GenAI solutions,I enjoy sharing practical perspectives on AI-assisted engineering, architecture, delivery frameworks, and the business shifts shaping the future of data teams. My focus is on real-world implementation, lessons learned, and helping organizations move from experimentation to production with confidence. If these topics resonate with you, I’d be happy to connect and exchange ideas. Feel free to reach out and connect on LinkedIn :https://www.linkedin.com/in/rahul-sahay-8573923/ [Your Best Prompts Are Living in Slack. That Is Costing You More Than You Think](https://pub.towardsai.net/your-best-prompts-are-living-in-slack-that-is-costing-you-more-than-you-think-1dddef75a48f) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.