Spec-Driven Development Backlash

What is it

Spec-Driven Development (SDD) is the methodology where you write a detailed specification first — requirements, constraints, acceptance criteria, edge cases — and then let an AI coding agent (Claude Code, Cursor, Codex, Windsurf) generate the implementation against that spec. The pitch is seductive: specs become the source of truth, code becomes disposable, and "vibe coding" chaos gets replaced with engineering discipline.

The backlash is the counter-movement. Matt Pocock's video crystallized the frustration: developers are finding that writing exhaustive specs for an AI agent often takes as long as writing the code, that specs drift from reality the moment the agent improvises, and that the resulting ceremony feels like re-inventing waterfall with extra steps. The GitHub and V2EX threads that surfaced this term show practitioners pushing back on tool vendors who are selling "spec frameworks" as the answer to AI code quality.

The business significance is real: every AI coding tool vendor is currently deciding whether to ship a spec layer, a spec linter, or a spec-to-test pipeline. That's a live purchasing decision inside thousands of engineering teams right now. Whoever reduces the cost of spec-writing — rather than adding more ceremony — captures the budget. The backlash is the demand signal: developers don't want more process, they want specs that write themselves and stay honest.

Why now

Three forces converged in late 2025 and early 2026 to make this a live debate rather than an academic one.

First, agentic coding tools crossed the "good enough" threshold. Claude Code, Cursor's agent mode, Codex CLI, and Devin all now run multi-file, multi-hour tasks autonomously. That autonomy is exactly what makes spec quality matter — a bad spec now produces 2,000 lines of confidently wrong code instead of a single bad function. Vendors responded by shipping spec-driven workflows (GitHub's Spec Kit, Amazon's Kiro with its spec files, Cursor's rules and plan mode), which turned SDD from a blog-post idea into a product category.

Second, the tooling fragmented fast. Every vendor invented its own spec format — Kiro's .kiro/specs, Spec Kit's markdown templates, Cursor rules, AGENTS.md. Teams running two tools now maintain two spec dialects. That fragmentation is the classic precondition for a neutral tooling layer.

Third, the cost curve inverted. Token prices dropped and agent runs got cheap, so the bottleneck moved from "can the model write the code" to "did we tell it the right thing." Pocock's video landed precisely because thousands of developers had just spent a quarter discovering that their spec files were stale, bloated, and unread by anyone but the agent.

This is a 6-12 month window. Once the major IDEs converge on a spec standard (or one vendor wins the format war), the neutral-tooling opportunity narrows sharply.

Market Evidence

The evidence base is thin but directionally clear. The term has 2 independent sources (GitHub and V2EX), 2 total mentions, and a 100% growth rate from a near-zero base, first seen 2026-10-21. Trend score sits at 65/100 — meaningfully above noise, but far from a breakout. Stage is explicitly "nascent."

Read this honestly: 2 mentions is not a market. It's a leading indicator. What makes it worth tracking is who is mentioning it and where. A GitHub discussion plus a V2EX thread means the conversation spans both Western open-source maintainers and Chinese-speaking developers — two communities that rarely surface the same methodology complaint simultaneously unless the friction is structural rather than local.

The 100% growth rate is a mathematical artifact of going from 1 to 2 sources. Ignore it as a magnitude; respect it as a direction. The real signal is qualitative: Matt Pocock — a high-reach TypeScript educator with a large developer following — made a video arguing against the dominant vendor narrative. When respected educators publicly push back on a methodology vendors are actively selling, that's the moment a counter-market forms.

Verdict: real friction, not yet a real market. Treat this as a 90-day watchlist item with a cheap validation experiment, not a build-now signal. The opportunity score of 0/100 is honest — the scoring model has nothing to work with yet. Your job is to generate the data that moves it.

Who's Behind It

The "whales" here are the AI coding tool vendors, and they are the reason this debate exists at all. GitHub (Spec Kit), AWS (Kiro), Anysphere (Cursor), and OpenAI (Codex CLI) are all shipping spec-oriented workflows because specs are how they justify premium pricing and enterprise trust. Their competitive dynamic is a format war: whoever's spec file becomes the default gets lock-in on the whole workflow.

On the backlash side, the loudest voices are practitioner-educators, not companies. Matt Pocock is the focal point — his video gave the diffuse frustration a name. Around him sit the open-source maintainers in GitHub discussions and the V2EX developer community, both of which skew toward senior engineers who resent process imposed by tooling.

The absence of a company on the backlash side is the key structural fact. There is no vendor selling "anti-SDD." That's the gap: the counter-position is currently a complaint, not a product. The moment someone packages the complaint into a tool that makes specs cheaper to maintain, they own a category the whales have no incentive to build — because the whales profit from more ceremony, not less.

Watch Pocock's follow-up content and the GitHub issue velocity on Spec Kit. If maintainers start filing bugs about spec drift, the tooling opportunity is confirmed.

TAM & Market Size

Be realistic about the buyer. The people feeling this pain are professional developers using agentic coding tools daily. That's a subset of the ~28 million developers worldwide, but a fast-growing one. Cursor alone reported over a million users by early 2025; Claude Code and Codex CLI add millions more installs. Call the addressable population 2-4 million developers who run agents on real codebases and have hit spec-quality friction.

Willingness to pay is the crux. Individual developers already pay $20/month for Cursor and $20-200/month for Claude. They will pay for something that saves agent rework — but only if it's cheap and demonstrably reduces tokens or review time. A credible price band for an individual tool is $9-19/month; a team seat is $25-40/month.

Budget-holders inside companies are engineering managers who already buy AI coding seats. They will pay for governance — spec versioning, drift detection, audit trails — at $500-2,000/month per team of 10-20. That's where the real revenue sits.

The demand score of 0/100 reflects zero validated willingness-to-pay data, not zero demand. Your first job is to convert the 2-mention signal into 20-50 conversations that establish a price anchor. Do not build a TAM slide before you have five people saying "yes, I'd pay $15 for that."

Competitive Landscape

The competitive field splits into three layers, and the gap sits between them.

Layer one: the IDEs and agent vendors themselves — Cursor, Kiro, Claude Code, Codex, GitHub Copilot. They ship spec features but are structurally biased toward more spec ceremony, because that's their moat. They will never build a tool that says "your spec is over-engineered, delete half of it."

Layer two: spec-framework and template projects — GitHub Spec Kit, BMAD-METHOD, various AGENTS.md conventions. These are free, open-source, and proliferating. They solve the "what format" question but not the "keeping specs honest and cheap" question. They're your distribution channel, not your competitor.

Layer three: documentation and knowledge tools (Notion, Confluence, Linear) that teams already use to store specs. They have zero AI-agent awareness.

The gap: nobody owns spec maintenance — drift detection, staleness scoring, spec-to-test generation, and cross-tool format translation. That's a genuine whitespace. Competition score of 0/100 reflects that no incumbent is targeting this specific job.

Time pressure: if Cursor or GitHub ships native spec-drift detection, your wedge shrinks. You have roughly 6-9 months before the platform players notice. Build as a cross-tool layer they can't easily replicate without cannibalizing their own ceremony.

Business Model

Recommend a hybrid: freemium CLI/linter for individuals, paid team SaaS for organizations. Here's why. The individual developer is your distribution engine — a free spec-lint that flags stale specs and over-long specs spreads through GitHub repos and HN. The team plan monetizes the governance need: shared spec registry, drift alerts in CI, and an audit view for managers.

Pricing:

• Free: single repo, up to 3 specs, drift warnings in CLI.
• Pro ($15/month): unlimited specs, CI integration, spec-to-test generation, cross-format conversion (Kiro ↔ Spec Kit ↔ AGENTS.md).
• Team ($29/seat/month, min 5 seats): shared registry, drift dashboard, Slack alerts, SSO.
• Enterprise (custom, $1,200+/month): audit logs, on-prem, policy enforcement.

The $15 Pro price sits deliberately below a Cursor seat so it reads as an add-on, not a replacement. The $29 team seat is priced against a fraction of one engineer-hour of rework per month — an easy manager justification.

12-month forecast: conservative $4k MRR (≈250 Pro, 2 teams); base $18k MRR (≈800 Pro, 15 teams); optimistic $55k MRR (≈2,000 Pro, 45 teams, 2 enterprise). CAC via developer content and open-source is low — estimate $25-60 for Pro, $400-900 for team deals. Payback under 3 months for Pro, under 6 for teams. The model works only if the free tier genuinely spreads; if it doesn't, the funnel collapses and you should pivot to a pure content play.

MVP Blueprint

Ship in 5 days. Core job: make a stale or bloated spec visible and fixable in one command.

Day 1-2: CLI that parses spec files (markdown, YAML frontmatter) from a repo, extracts linked code paths via simple heuristics (file mentions, function names), and computes a drift score — how much the referenced code has changed since the spec was last touched. Output: a ranked list of specs by staleness.

Day 3: bloat score — flag specs over N words or with low testable-assertion density (count "must/should" statements vs. prose). This is the anti-ceremony feature and the whole thesis in one number.

Day 4: spec-fix command that calls an LLM (Claude or GPT API) to rewrite the worst spec into a tighter, testable version, with a diff for human approval. This is the "cheap specs" promise made concrete.

Day 5: GitHub Action wrapper so it runs in CI and posts a PR comment. That's the distribution mechanism.

Stack: Node/TypeScript (developer-native, easy npx distribution), simple git-based diffing (no database), one LLM API call per fix. No dashboard, no auth, no billing on day 5. Ship to npm, post to HN and the GitHub thread that spawned the term.

Cut: web UI, team features, format conversion, analytics. Those are week 3+. The MVP's only job is to prove people run it twice.

Commercial Opportunities

1. Spec-drift CI service. Target: platform teams at 20-200-engineer companies running Cursor or Claude Code. Product: a CI bot that fails PRs when specs go stale and posts a suggested rewrite. Expected $3k-15k MRR within 6 months from 5-20 team subscriptions at $500-1,500/month. Beats alternatives because it plugs into existing CI with zero workflow change — no new editor, no migration.

2. Cross-format spec translator. Target: teams running two agent tools (common: Cursor + Claude Code). Product: convert Kiro specs ↔ Spec Kit ↔ AGENTS.md so one spec serves all agents. Expected $2k-8k MRR. This wins because format fragmentation is a headache vendors won't fix for each other — a neutral translator has no incentive conflict.

3. Spec-quality content and course. Target: developers burned by SDD ceremony. Product: a paid guide and template pack on writing minimal, testable specs, priced $49-99. Expected $1k-5k/month, low effort, and it doubles as top-of-funnel for the tool. This beats building a course on generic AI coding because it's contrarian — you're selling against the vendor narrative, which is inherently more shareable.

Prioritize #1. It has the clearest budget-holder and the highest price tolerance.

Product Ideas

🥇 SpecDrift — one-line: "Catch stale specs before your AI agent codes against them." Target: developers using agentic tools on existing codebases. Why now: spec files are proliferating across repos and nothing tracks whether they still match the code. A CLI plus GitHub Action, free tier for individuals, $15/month Pro. This is the wedge — narrow, painful, unowned.

🥈 SpecLint — one-line: "A linter for specs that flags over-engineering before the agent does." Target: teams that adopted SDD and are drowning in ceremony. Why now: the backlash is specifically about spec bloat; a tool that scores testability and brevity is the counter-position made mechanical. Free CLI, $29/seat team plan for policy enforcement. Pairs naturally with SpecDrift as a second module.

🥉 SpecBridge — one-line: "Write your spec once, run it in Cursor, Claude Code, and Kiro." Target: teams running multiple agent tools. Why now: format fragmentation is peaking and no vendor will build a neutral translator. $19/month Pro, $99/month team. Higher ceiling but later — wait until SpecDrift proves the audience.

Ship in that order. Each one validates the next. Do not attempt SpecBridge first; it's the most technically fiddly and the least proven demand.

SEO Opportunity

Search volume for "spec-driven development" is climbing but still small — likely low thousands monthly, spiking around vendor launches. The contrarian angles are where the long-tail sits and where competition is near zero (SEO difficulty 0/100).

Target these long-tail keywords: "spec-driven development is a waste of time," "AI coding spec drift," "how to write specs for Claude Code," "spec kit alternative," "AGENTS.md best practices." The first two are pure backlash intent — low volume, high conversion, almost no competing content.

Content strategy: publish one contrarian, data-backed post ("We analyzed 500 spec files — 60% were stale within 30 days") and let it rank against the vendor marketing. Backlash content outperforms neutral tutorials because it's opinionated and shareable. Link every post to the free CLI.

Risk Assessment

When this thesis is wrong: if the backlash is a personality-driven blip (one educator's hot take) rather than structural friction, demand evaporates. Two mentions is dangerously thin evidence.

Risk 1 — Market: vendors fix it themselves. If Cursor ships native drift detection inside the editor, a standalone CLI loses its reason to exist. Mitigation: stay cross-tool and CI-first, where IDEs are weakest.

Risk 2 — Tech: spec drift is hard to detect reliably. Heuristic file-matching will produce false positives that erode trust fast. Mitigation: start with conservative detection and let users tune thresholds; never cry wolf.

Risk 3 — Execution: the audience is developers, the most price-sensitive and churn-prone buyers alive. A $15 tool must prove value in the first session or it's uninstalled. Mitigation: optimize the first-run experience relentlessly.

Cheap validation: post the drift-score idea in the exact GitHub and V2EX threads that spawned the term; offer to run it on five volunteers' repos manually. If fewer than three say "yes, run it again next month," walk away. Set a hard 30-day kill date.

Action Plan

Today: write a one-paragraph pitch of the drift-score CLI and post it as a comment in the GitHub discussion and V2EX thread that surfaced this term. Ask directly: "Would you run a tool that flags specs your agent has drifted from?" Count replies.

Week 1: manually run a crude drift check on 5 volunteer repos (git log on spec-referenced files). Deliver a written report to each. This costs nothing and produces the first real demand data. Target: 3 of 5 asking for a repeat.

Month 1: if validated, ship the 5-day MVP (CLI + GitHub Action), publish to npm, write the contrarian data post, and submit to HN. Goal: 200 installs, 20 repeat users, first 5 Pro signups.

Month 3: if 50+ weekly active users and 10+ paying, add team features (shared registry, Slack alerts) and start outbound to platform teams. Goal: $2k MRR. If the numbers aren't there, kill it and publish the learnings as content — the audience is still valuable.

Set the kill criteria in writing before you start. Discipline beats optimism here.

Related Terms

AI Code Review Fatigue — the parallel backlash against AI-generated PR review noise. Same buyer, same "tooling adds work instead of removing it" complaint; a natural second product line.

Agentic Coding Governance — the enterprise push to audit and constrain AI agents. Spec drift detection is a governance primitive, so this trend is the enterprise upsell path.

Context Engineering — the practice of curating what goes into an agent's context window. Specs are a form of context; tightening specs is context engineering by another name.