Last Tuesday, I watched two senior engineers argue for 45 minutes about whether a null value should return an empty array or a 404 error. Both were right. Both were wrong. The real problem? We never wrote down what "not found" actually means in our system. This wasn't a technical debate—it was a specification problem disguised as a code review. What Got Me Thinking About This I've been writing code professionally for years now, and I've seen teams work in all sorts of ways. Some dive straight into coding ("we'll figure it out as we go"), others spend weeks in meetings debating architecture before writing a single line. But the approach that consistently produces better results? Starting with a solid specification. Not the 200-page waterfall document that nobody reads. I'm talking about living, breathing specs that actually help you build better software. So What Exactly Is Spec-Driven Development? At its core, spec-driven development means you write down what you're building before you figure out how to build it. Think of it as having a conversation with your future self (and your team) about what success looks like. what how A good spec answers questions like: What problem are we solving?What should this thing do when everything works perfectly?What should happen when things go wrong?How do we know we're done? What problem are we solving? What should this thing do when everything works perfectly? What should happen when things go wrong? How do we know we're done? It's not about bureaucracy or process for the sake of process. It's about thinking through the hard questions before you've written thousands of lines of code that you'll need to rewrite. The $15,000 Assumption Let me tell you about a project from 2023 that still makes me wince. We were building an API for payment processing—nothing exotic, just standard refund functionality. The product brief said "support refunds" in bullet point #7. Clear enough, right? Week three, during integration testing, everything exploded. The backend team had built partial refunds as a separate `/refunds/partial` endpoint with its own transaction flow. Made perfect sense from a database perspective—cleaner state management. `/refunds/partial` The frontend team had built their UI around a single `/refunds` endpoint with an optional `amount` parameter. Made perfect sense from a UX perspective—one button, simpler flow. `/refunds` The product manager, when we finally asked, expected it to work exactly like Stripe does (which matched neither implementation). Three capable teams. Three different "obvious" interpretations. Zero documentation. We burned two weeks untangling the mess, rewrote about 3,000 lines of code, and had several uncomfortable conversations about "why didn't anyone ask?" The invoice from our contractors alone was $15,000. All because we couldn't be bothered to spend two hours writing down what "refund" meant in our system. My Approach to Spec-Driven Development Here's how I typically work now: 1. Start with the Problem, Not the Solution Before writing any spec, I spend time understanding why we're building something. What pain point exists? Who experiences it? How do they currently work around it? why This context shapes everything else. A spec for "let users export data" looks very different if users are trying to comply with GDPR versus trying to analyze trends in Excel. 2. Write Like You're Teaching a Friend Nobody ever reads specifications. You know why? They're boring as hell. "The system shall implement RESTful endpoints conforming to OpenAPI 3.0 specification with JSON payload serialization." My eyes glazed over just typing that. Compare it to: "We're building a REST API. Send JSON, get JSON back. Standard REST rules apply. Full details in the OpenAPI doc." Same information. One makes you want to code. The other makes you want to update your resume. I write specs the way I'd explain them to a new team member at lunch. If I wouldn't say it out loud to a person, I don't write it in a doc. The goal is communication, not impressing anyone with jargon. 3. Edge Cases Are Where Specs Earn Their Salary The happy path is trivial. Click "Save," data gets saved. A trained monkey could spec that out. But here's what actually happens in production: User's coffee shop WiFi drops mid-saveHammy and Sarah both edit the same customer record at 2:47 PMSomeone tries to upload a 500MB file to a field that expects 5KBNav's admin access gets revoked while he's halfway through a transaction User's coffee shop WiFi drops mid-save Hammy and Sarah both edit the same customer record at 2:47 PM Someone tries to upload a 500MB file to a field that expects 5KB Nav's admin access gets revoked while he's halfway through a transaction These aren't hypotheticals from a textbook. This is Tuesday. Last month, we had a bug where users could delete a resource, then immediately use it in another operation because our cache hadn't invalidated. Cost us a weekend and some very unhappy customers. Why? Because our spec said "DELETE removes the resource" and nobody asked "yeah, but when is it actually gone?" when Writing down what happens in these scenarios doesn't prevent the technical challenges—but it prevents three different engineers from solving the same problem three different ways, then finding out in production that none of them handle all the cases. 4. Include Examples Nothing clarifies like a concrete example. When specifying an API endpoint, I include: Example Request:
POST /api/payments/refund
{
  "transaction_id": "txn_12345",
  "amount": 50.00,
  "reason": "customer_request"
}

Example Success Response:
{
  "refund_id": "ref_67890",
  "status": "pending",
  "estimated_completion": "2024-03-15T10:30:00Z"
}

Example Error Response:
{
  "error": "insufficient_funds",
  "message": "Cannot refund more than the original transaction amount",
  "original_amount": 45.00
} Example Request:
POST /api/payments/refund
{
  "transaction_id": "txn_12345",
  "amount": 50.00,
  "reason": "customer_request"
}

Example Success Response:
{
  "refund_id": "ref_67890",
  "status": "pending",
  "estimated_completion": "2024-03-15T10:30:00Z"
}

Example Error Response:
{
  "error": "insufficient_funds",
  "message": "Cannot refund more than the original transaction amount",
  "original_amount": 45.00
} Suddenly everyone knows exactly what we're building. No ambiguity. 5. Make It a Living Document The biggest mistake? Treating specs like they're set in stone. Software evolves. Requirements change. New information surfaces. I keep specs in version control, right alongside the code. When we learn something new or change direction, we update the spec first, then update the code. The spec stays the source of truth. What Actually Changed When We Started Doing This I'm not going to pretend I have rigorous data, but here's what I noticed after our team adopted spec-first for major features: Estimates got scary accurate: Our sprint planning went from "probably two weeks, maybe four" to hitting estimates within a day or two. When you've thought through what you're building, you know what you're estimating. Estimates got scary accurate Code reviews got faster: Instead of "wait, why does this work this way?" we had "does this match the spec?" Much cleaner discussions. Code reviews got faster Onboarding stopped sucking: New engineers could read specs and understand system behavior without needing a dedicated guide. One junior dev told me she learned more from our specs in a week than from shadowing for a month. Onboarding stopped sucking Testing became obvious: The spec literally lists what should happen. Those become test cases. We went from "did we test enough?" to "did we test what the spec says?" Testing became obvious Fewer oh-shit moments: Production bugs didn't disappear, but the "nobody thought about this" bugs dropped noticeably. We measured it—went from about 40% of bugs being "unclear requirements" to under 15% over six months. Fewer oh-shit moments The async work thing: This one surprised me. With good specs, people across timezones could make decisions without waiting for me to wake up and answer questions. Our India team could make progress while I was asleep. Huge win. The async work thing When Specs Aren't the Answer I'm not dogmatic about this. There are times when spec-driven development doesn't make sense: Rapid prototyping: When you're exploring what's possible, coding is faster than speccing. Just be ready to throw it away.Trivial changes: Fixing a typo doesn't need a spec.Research spikes: Sometimes you need to code to understand the problem space. Rapid prototyping: When you're exploring what's possible, coding is faster than speccing. Just be ready to throw it away. Rapid prototyping Trivial changes: Fixing a typo doesn't need a spec. Trivial changes Research spikes: Sometimes you need to code to understand the problem space. Research spikes The key is knowing the difference between "I'm exploring" and "I'm building something that needs to last." The Tooling Has Gotten Ridiculously Good I'll be honest—five years ago, I would've written this same post but been less enthusiastic about it. Writing good specs took time. The kind of time that made project managers nervous and made you feel guilty when there was code to ship. time That's changed. We now have tools—AI assistants, better templates, collaborative platforms—that reduce the grunt work. I can sketch out "need an API for filtering transactions by date, amount, and category" and get back a proper spec with examples, validation rules, and error cases in minutes. It's not perfect, but it's a hell of a starting point. The important part: these tools don't replace thinking. They handle the tedious bits—consistent formatting, comprehensive examples, remembering to include all the standard HTTP status codes. You still need to know your domain, understand your users, and make the hard design decisions. Think of it like spell-check. It catches your typos, but it doesn't write your essay. Same principle. What this means practically: the time investment for spec-driven development has dropped from "ugh, this takes forever" to "eh, this is pretty quick." The ROI got a lot more attractive. Getting Started with Spec-Driven Development If you're sold on the idea but not sure where to start: 1. Pick one feature: Don't try to spec your entire system at once. Start with the next thing you're building. 1. Pick one feature 2. Use a template: Having a structure helps. Something like: 2. Use a template * Overview (what and why) * Requirements (functional and non-functional) * Interface (API, UI, whatever's relevant) * Data model (what gets stored/passed around) * Error handling (what can go wrong) * Testing approach (how we verify it works) 3. Review with your team: The value is in the discussion. Get feedback before you code. 3. Review with your team 4. Keep it updated: When reality diverges from the spec, update the spec. A wrong spec is worse than no spec. 4. Keep it updated 5. Measure the impact: Track how often specs catch issues before coding versus issues found later. You'll probably be surprised. 5. Measure the impact Look, I Get It Specs feel like extra work. Especially when you're staring at a gnarly bug at 4 PM on a Friday, or when your PM is asking why feature X isn't shipped yet. I've been that developer who thinks "I'll just code it, it's faster." Sometimes? Sure, go ahead. For a quick prototype or a simple bug fix, speccing it out is overkill. But for anything that matters—anything that'll stick around, anything that multiple people will touch, anything that'll scale—I've learned the hard way that thinking before typing saves time. Not occasionally. Consistently. The specs don't need to be perfect. They don't need to be exhaustive. They just need to answer the questions that would otherwise turn into assumptions, which turn into arguments, which turn into refactoring, which turn into regret. I still write bad code sometimes. I still ship bugs. But at least now when something goes wrong, it's because I made a technical mistake, not because I never bothered to figure out what I was actually building. That's a trade I'll take every time. _________________________________________________________________________________________________ Quick Reality Check: This isn't a manifesto. You don't need to spec everything tomorrow. Pick one thing—your next feature, your next API—and try it. See if it helps. If it doesn't, you wasted a few hours. If it does, well, you might just avoid your own $15,000 mistake. Quick Reality Check Got war stories about specs saving your bacon (or making things worse)? Drop a comment. I'm always curious how different teams handle this. Got war stories about specs saving your bacon (or making things worse)? Drop a comment. I'm always curious how different teams handle this. _________________________________________________________________________________________________

Spec-Driven Development: Why Writing the Manual Before Building the Machine Makes Sense

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Writing Code Is Easy Now. Reviewing It? That’s Where the Real Work Begins

The Limits of LLM-Generated Unit Tests

Machine Learning-based Vulnerability Protections For Android Open Source Project

Classifier-Based VP System Targets High-Risk Code Changes in Upstream Projects

How Lightweight ML Models Predict Vulnerable Code Changes

Inside the Data Pipeline Behind Classifying Android Security Flaws

Writing Code Is Easy Now. Reviewing It? That’s Where the Real Work Begins

The Limits of LLM-Generated Unit Tests

Machine Learning-based Vulnerability Protections For Android Open Source Project

Classifier-Based VP System Targets High-Risk Code Changes in Upstream Projects

How Lightweight ML Models Predict Vulnerable Code Changes

Inside the Data Pipeline Behind Classifying Android Security Flaws

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps