Spec-Driven Development in 2026: What Actually Changed Transcript, AI Summary & Key Points

00:00 One framework grew 863% over the last six months. Another one in the same category grew 18. Last October I made a video comparing three spec-driven development frameworks — BMAD, Spec Kit, and OpenSpec. Later I covered Taskmaster and Agent OS in separate videos. Eighty-six thousand of you watched the first one. The combined GitHub stars of those five jumped from around eighty-seven thousand to two hundred and two thousand.

00:30 Plus one hundred and thirty-one percent in six months. That part isn't the interesting number. The interesting number is the spread. Inside that same average, one framework ran away with it while another stalled. One rewrote itself from scratch, and one quietly shrunk its own scope. Spec-driven development gained momentum over the last six months — partly because the tools got better, partly because of marketing.

00:56 This video is the six-month update: What changed, who won, who stalled, and what the rest of the community has been saying. No hands-on retest this time — that's a separate one, coming later. OpenSpec grew from four thousand one hundred stars in October to thirty-nine thousand five hundred now. Plus eight hundred and sixty-three percent in six months.

01:23 The runaway story of the last half year. Honestly, back when I first compared the three frameworks, I recommended Spec Kit. What actually changed my mind wasn't the v1 release — it was what happened after I published the video. I kept using both OpenSpec and Spec Kit on real work, and over a few weeks it became obvious OpenSpec was producing the same quality, just much faster.

01:47 Spec Kit was simply too slow to plan. So I dropped Spec Kit and stuck with OpenSpec. Then — and this is where it gets interesting — I eventually dropped even OpenSpec. Claude Code's built-in plan mode started covering most of what I needed for day-to-day work, and Superpowers handled the bigger features. OpenSpec didn't do anything wrong. The native tooling just caught up.

02:08 The growth numbers suggest a lot of other people went through some version of that same funnel — which is part of why OpenSpec, the lightest-weight framework in the set, came out on top. A lot changed on their side. The v-one release is a significant rewrite — smaller output, clearer structure, less markdown sprawl. The team also joined Y Combinator's Winter 26 batch and expanded tool compatibility to more than 25 agents and IDEs.

02:36 The community coverage matches the numbers. Augment Code's ranking has it at number four out of six, but Augment Code's own product is ranked number one in the same list, so I wouldn't weight that placement too heavily. Rick Hightower's piece consistently calls out OpenSpec as the lightest-ceremony option of the four he compared. My own take: lightweight beats heavy when the frontier model is already doing most of the implementation work.

03:03 A small disclosure. A lot of people at Toggl — where I work — use OpenSpec as their framework of choice. A couple of enthusiasts even built an internal UI wrapper around it, essentially a Kanban-style board for OpenSpec workflows. So factor that in. I have a biased sample of one engineering organization. Stars aren't a perfect metric either — YC cohorts and Hacker News spikes can juice them.

03:29 But plus eight hundred sixty-three percent is a lot, and that one's public, not mine to manipulate. My honest take. If you're coming to spec-driven development for the first time, OpenSpec is where I'd start. It's fast. It won't bury you in markdown before you write any code, and the community around it is growing faster than any of the original five I'm covering today.

03:51 That doesn't mean it's the right fit for every team. BMAD's fans will tell you otherwise, and they have a point too. But as a default starting position for a new team exploring the space — OpenSpec is the one I'd hand somebody first. BMAD went from nineteen thousand one hundred stars to forty-four thousand four hundred. Plus one hundred and thirty-three percent, which tracks with the rest of the market growth but doesn't lead it.

04:16 The interesting thing here isn't the raw number — it's what they shipped along the way. Version six shipped back in February. v6.0.0 came out on February 17th, and they've been pushing point releases ever since — v6.3 is from early April. The stable six release was essentially a ground-up rewrite. They consolidated several older agents into a single persona called Amelia, introduced what they call a Skills Architecture, and shipped a Cross-Platform Agent Team that works across more than twenty agent runtimes.

04:53 Twelve-plus specialized agent personas are available. On top of that, the team spun out BMad Builder as its own project for building custom agents and workflows. It's at v1.4 already. Dev Loop Automation is another piece of the ecosystem. There's also a small marketplace of community-contributed skills. That's a lot of surface area. Community reception is mostly positive on the technical side.

05:19 On the critical side, the most common complaint is the ceremony itself: the number of markdown files per spec, the full PM-to-developer handoff rituals, and a PRD phase that asks you to define business success metrics for a weekend project. That's the tradeoff BMAD has always made. And by the way — no wonder so many BMAD fans showed up in the comments of my last video to defend it.

05:43 The community around this one runs hot. BMAD has one of the most active communities on this leaderboard; their Discord usually has close to two thousand people online at any given time, which is genuinely a lot for a framework this specialized. I'm not going to claim it's literally the biggest without doing a head-count across every project, but it's among the most active, easily.

06:05 Quick disclosure on my own side. I did re-test BMAD v6 fairly recently — the full four-phase flow, which is the actual ceremony: Planning, Solutioning, Implementation, all of it. The results weren't great. Seven hours to build a cat game I could build in ten minutes with plan mode. I have a separate video on that whole experience, linked below. A fair number of viewers in the comments made a point worth repeating: BMAD is designed for bigger project scopes, and a small test project doesn't let it show what it can

06:35 actually do. That's a legitimate critique. I'm not going to rerun the full four-phase ceremony on a bigger project though — that would cost me another eight to twenty hours, and I'd rather keep my weekends. What I do want to retest is Party Mode and Quick Flow, on a larger scope, and see whether the lighter-touch modes hold up better than the heavy one did.

06:56 That'll be its own video — hit subscribe if you don't want to miss it. My honest take. BMAD isn't really a team tool — I don't think "for teams" is even the right framing. What kind of team is this, where each member orchestrates their own nested squad of virtual PM, architect, and developer agents? BMAD is for individual builders who genuinely want the full agile ritual layered on top of their AI coding setup.

07:21 Solo devs, serious vibe coders, agent orchestrators — whatever we're calling them this month. If you like structured multi-agent workflows and don't mind reading spec documents, this is the one. Spec Kit grew from thirty-nine thousand three hundred stars to eighty-seven thousand five hundred. Plus one hundred and twenty-three percent. That's a huge raw number — almost fifty thousand new stars — and it puts Spec Kit at the top of the leaderboard in absolute stars by a comfortable margin.

07:54 What GitHub shipped. One hundred and thirty-six releases since October. Support for more than fourteen AI agent platforms — Claude Code, Cursor, Windsurf, Copilot, and a long tail of smaller agents. Microsoft also published an official Spec Kit training module on Microsoft Learn. Other SDD frameworks have courses on platforms like Udemy, sure, but the Microsoft Learn module signals something different in an enterprise context.

08:20 The creator, Den Delimarsky, also left Microsoft, where he built Spec Kit as a GitHub project — for Anthropic, along the way. Could be read as a founder leaving the project; could be read as a vote of confidence in where AI-assisted development is going more broadly. Probably a bit of both. Community reception is mixed. Martin Fowler's analysis of Spec Kit was the sharpest critique in the space.

08:44 His observation: Spec Kit generates eight-plus markdown files per spec, and in his words, those files were "repetitive, verbose and tedious to review." His closer — and this is verbatim — was "I'd rather review code than all these markdown files." He also noted that agents frequently ignored the instructions in those specs. Augment Code's ranking puts Spec Kit at number three out of six.

09:08 Rick Hightower's comparison describes it as middle-ground execution — it executes through the connected AI, but it doesn't manage parallelism the way frameworks like GSD do. My honest take. Spec Kit is faster than BMAD. It's also noticeably slower than OpenSpec. The difference isn't really about structure. When I ran Spec Kit back in October, the output quality felt about the same as OpenSpec's, just arriving inside a much larger pile of markdown.

09:36 And that plan phase — I remember it took around five minutes just to complete planning, which felt incredibly slow compared to OpenSpec finishing the same step in a fraction of the time. My instinct — and I'll be upfront this is instinct, not a proven claim — is that a meaningful chunk of Spec Kit's growth comes from GitHub's brand and Microsoft's Learn module, not from the product being the best on merit.

09:58 I'd love to see some real-world comparisons that confirm or deny that. Remember the thesis from the hook: momentum from actual tools. Momentum from marketing. Spec Kit is probably where the marketing side of that shows up most clearly. If you're considering Spec Kit, test it yourself against OpenSpec on a real task before committing a team to it. Go in with eyes open about what you're buying.

10:22 Agent OS went from roughly 2500 stars to 4300, about plus 72%, which sounds healthy in isolation, but in a market that more than doubled. It's the slowest growth story on the leaderboard. And the repository activity tells you why. The last release — version three — shipped on January 20th, which was three months ago, and there's been nothing since. For comparison, Spec Kit has shipped around 40 releases in the same three months.

10:52 BMAD ships weekly to biweekly. OpenSpec ships weekly, and even Taskmaster manages a release every few weeks. Agent OS has gone completely quiet. In practical terms, the project looks abandoned. What did ship was version three, and this is the interesting part. V3 was an explicit, public downsizing of the framework. The creator, Brian Casel, was direct about the reason in his changelog — frontier models, which is the industry's term for the large models like Claude and GPT that sit at the top of the capability curve,

11:26 now handle a lot of what earlier Agent OS versions were built to handle — spec writing, task breakdown, implementation orchestration, all of it. So V3 intentionally retired those phases and refocused the framework on what Casel calls its core strength: establishing and injecting standards. The V3 commands reflect that — /discover-standards, /inject-standards, /index-standards, /shape-spec, and /plan-product.

11:52 Five commands, narrow scope, clear purpose. Community coverage is thin. If you go looking for Agent OS in the comparison pieces — Rick Hightower, Augment Code, Vishal Mysore's map of thirty-plus frameworks — it mostly gets a brief mention. It's no longer one of the frameworks people are actively comparing. Now, my honest take, and this is where I want to be careful.

12:14 The headline question I had going in was: what does Agent OS actually give me in twenty twenty-six that CLAUDE.md files and Claude Code's built-in skills and plan mode don't already provide? Standards enforcement used to be something you needed a dedicated framework for. Now the coding tool itself handles a lot of it. But reading Casel's V3 rationale, the steel-man is actually reasonable — Agent OS isn't trying to replace plan mode, it's trying to layer on top of it by making standards discovery and injection more

12:45 automatic. /inject-standards can bake your rules into Claude Skills, subagents, or any prompt, which is genuinely more sophisticated than pasting them into a CLAUDE.md file by hand. So there's a real value proposition if standards automation is specifically what you need. Whether that's worth adopting a whole framework for, given that the project hasn't pushed a release in three months — that's the question I'd weigh carefully.

13:11 If you're using V3 and it's working for you, drop a comment. I'd genuinely like to hear from people it's clicking for. And one thing I'll say for Casel: shipping a smaller V3, instead of pretending the tool still does what it used to do, is respectable. Not every maintainer would make that call. Taskmaster went from 22,500 stars to 26,500. Plus 18%.

13:36 In a market where the average framework grew more than 100%, plus 18 is effectively going backwards. And this is the one that surprised me most on the leaderboard, because six months ago, Taskmaster was arguably the most convenient framework of the set. What happened. A few things, in combination. The project moved from the original repository into a dedicated one.

13:59 The team launched a commercial layer called Hamster over at tryhamster.com — a paid offering built around the same task-decomposition engine. The license on the open-source side is MIT plus a Commons Clause rider, which restricts commercial resale. That has actually been the case for a while — Taskmaster wasn't pure MIT to begin with — but it's worth noting that among all the frameworks on this leaderboard, Taskmaster is the only one that isn't pure MIT.

14:27 Around this same period, the community also raised concerns about Sentry telemetry calls getting shipped on by default with an opt-out rather than opt-in. Release cadence also slowed — Taskmaster ships roughly every few weeks. Now, closer to monthly than weekly. Community coverage is still positive on the technical side. Rick Hightower's comparison explicitly says Taskmaster "shines at the decomposition layer" — turning product requirements documents into dependency-aware task graphs.

14:57 That part of the tool really does work well. The results Taskmaster produces are, in my experience back when I tested it, genuinely good. The engine is solid. My honest take. When I used Taskmaster six months ago, I thought it was the most convenient of the frameworks I tried. And the code it produced was actually good — I want to be fair to the tool, because the output was not the problem.

15:20 The developer experience was fine overall, honestly. But over time I noticed two specific things I wish worked better. First, the way Taskmaster scopes tasks — it uses tags, where you have separate buckets like a master tag and whatever custom ones you create on top. Switching between them never felt natural to me. It could be smoother. Second, the tasks themselves are stored in JSON files.

15:43 You can absolutely read them in the terminal when you run the CLI commands — that part is fine. But opening the raw JSON in an editor and trying to skim through it is not a pleasant experience. Both of those are fixable, and neither is a dealbreaker. The engine is solid. So the core tool still works. But you combine the commercial pivot, the license situation, the telemetry concern, and the slower release cadence — and you can see why the star growth stalled.

16:11 It's an interesting case study in monetization timing. Taskmaster was growing well when the commercial pivot happened. Twelve months from now there'll probably be a blog post from somebody — a proper teardown — arguing either that the pivot saved the project or that it stalled it. Right now, based on the numbers, the stall side of that argument is winning.

16:32 While the original five were busy rewriting themselves or going commercial, the space around them kept expanding. A couple of new projects have grown large enough that you can't really talk about spec-driven development in 2026 without mentioning them. Superpowers, by Jesse Vincent. One hundred and fifty-six thousand stars — more than any single one of the five I just covered.

16:55 Superpowers isn't strictly a spec-driven framework. It's an agentic skills system for Claude Code that enforces test-driven development and a structured brainstorm-plan-execute workflow. It's adjacent to SDD rather than squarely inside it. But it's the largest structured-AI-coding project on GitHub right now, and I use it personally, daily, alongside plan mode.

17:18 When I have a big feature to plan, Superpowers asks many more questions than plan mode does. I already have a full video on Superpowers — link in the description if you want the deeper walk-through. GSD — Get Shit Done, by TACHES. Fifty-one thousand six hundred stars. This one is execution-first. Wave-based parallel orchestration, fresh context windows for researchers and planners and executors, all coordinated by a top-level runtime.

17:45 Forty-seven releases since December. A different model from most of the leaderboard — GSD is more about running specs than writing them, with parallel agent contexts doing the actual work. Beads, by Steve Yegge. Twenty thousand eight hundred stars and the maintainer is still pushing commits — last push was a couple of days ago. Beads is a distributed graph issue tracker for AI agents, backed by Dolt, which is a version-controlled SQL database.

18:12 Their pitch is literally "replaces messy markdown plans with a dependency-aware graph." It sits in Taskmaster's category — task tracking and agent memory — but without the Commons Clause license and with a sharper thesis. A few of you have asked me to look at Beads. This is my brief flag: interesting, I want to try it properly, it'll be its own video.

18:33 And then there's the native-features side of this story, which is the thing that changed most quietly. Claude Code itself now ships with skills, plan mode, ultra plan mode — that last one is a newer addition — and CLAUDE.md files for project-level standards. A lot of what the smaller spec-driven frameworks were built to provide — planning structure, standards enforcement, multi-agent coordination — has been absorbed into the coding tool you're already running.

18:59 That's part of why some of these frameworks are growing slower than you'd expect. The competition isn't other frameworks anymore. It's the tools they sit on top of. Vishal Mysore mapped more than thirty agentic coding frameworks in a March Medium piece. The space is crowded. A lot of these projects won't make it through another six months. The ones that do will be the ones with strong communities or a clear reason to exist alongside the native features.

19:26 Spec-driven development now has a Wikipedia page. It has four papers on arXiv — that's the open-access preprint server academics use to publish research before formal peer review. Four of them in the first quarter of this year alone. Thoughtworks put SDD on their Technology Radar in the "Assess" category. Six months ago, the idea was still something you argued about on X.

19:50 Today it's a topic for conference keynotes and enterprise rollouts. Martin Fowler wrote a long, cautiously skeptical analysis — he uses a German word for "making things worse through attempted improvement," which I'm not going to try to pronounce on camera. His bigger worry is that SDD reminds him of model-driven development. Quick context: model-driven development was a nineties movement where you'd draw diagrams and generate code from them, and it mostly didn't work out.

20:19 His concern is that SDD makes the same promise — write the spec, generate the code — and might hit the same walls. His piece is the most coherent skeptical case I've read on SDD so far. On the critical side, François Zaninotto from Marmelab wrote a piece back in November called "Spec-Driven Development: The Waterfall Strikes Back" — arguing SDD brings waterfall-era problems back in markdown form.

20:43 It hit Hacker News the same week. Marc Brooker from AWS wrote a direct rebuttal arguing that SDD gets you designing earlier and more iteratively, rather than locking you into a big upfront design phase the way old-school waterfall did. The debate has shifted from "is SDD a good idea" to "how do you do SDD well." On the enthusiast side, Prezi Engineering ran four teams through SDD workflows.

21:09 One engineer migrated an application to Material UI in two minutes, which they contrasted with another company where seven frontend teams took twelve months to do a similar migration the old way. Those are two different apps and two different scopes, so the numbers don't translate directly — but the Prezi team treated it as proof SDD actually works at scale.

21:31 That's the kind of anecdote that gets people to give SDD a try. Six months of spec-driven development, compressed. OpenSpec ran away with it — plus eight hundred and sixty-three percent. The fastest workflow of the five, and the one I personally kept using the longest. Spec Kit grew on GitHub's brand as much as on its own merits. BMAD rewrote itself and kept its loyal community.

21:55 Taskmaster stalled right as it tried to go commercial. Agent OS went quiet and narrowed its scope to the one job it still does well. Around the edges, Superpowers, GSD, and Beads showed up. And Claude Code itself quietly absorbed a bunch of what these frameworks used to do. The headline: spec-driven development is real now, it's not all winning, and some of it is being replaced by features shipping inside the tools you already use.

22:21 New players keep appearing. This video is going to age badly — which is, honestly, the point of making it at all. A hands-on retest of the original five is on the list for a future video. If that's something you want, let me know in the comments. And one last thing. Every framework in this video is pitched as "the right way" to plan software with AI. My cat never writes a spec. She walks up at five in the morning, sings a song, and food appears. Fastest feedback loop I have ever seen.