Spec-Driven Development with Kiro, Cursor, and Claude Code in 2026
Cursor went from `$1B` to `$2B` ARR in three months. AWS Kiro launched broadly at re:Invent 2025 with EARS-notation specs. The fastest freelance devs in 2026 are picking a spec workflow, not just an editor.
The Delivvo team· May 30, 2026 8 min read
Cursor went from $100M ARR in 2024 to $1B ARR in November 2025 to $2B ARR by February 2026, the fastest SaaS run from $1M to $1B in any category (DigiDai, 2026). The same year Cursor reported it was used by 64% of the Fortune 500 and was shipping more than 100M enterprise lines of code per day (GetPanto, 2026). GitHub Copilot hit 4.7M paid subscribers in January 2026, up 75% year on year, and reached roughly 90% of the Fortune 100 (GetPanto, 2026).
Keep reading
The Stack Overflow 2025 Developer Survey found 84% of developers use or plan to use AI tools, 51% use them daily, and trust in AI-generated code sits at just 33% (Stack Overflow via Uvik, 2026). JetBrains' January 2026 ecosystem survey reports 74% of developers have adopted AI dev tools, with Copilot at 29% work usage, Cursor at 18%, and Claude Code at 18% (JetBrains via Uvik, 2026).
Then AWS shipped Kiro broadly at re:Invent 2025 with a different bet entirely: EARS-notation specs, agent hooks, and checkpointing as first-class primitives (GeekWire, 2025). The three leading editors are now openly competing on the same idea, which is that the spec should come first and the code should come second.
For freelance engineers, the question is no longer "which AI editor is faster." It is "which spec workflow do I sell."
What spec-driven development actually means
Spec-driven development is the discipline of writing a structured specification before generating code, then using that specification as the source of truth the AI works against. Three things change.
The unit of work is the spec, not the diff. The freelancer's first deliverable is a written specification of acceptance criteria, edge cases, and constraints. Code comes after.
The AI does not freelance. Agents work against the spec, not against vibes. Hooks fire when specs change. Checkpoints let you roll back.
Reviews happen on the spec, not on the PR. The buyer reviews and signs off on the spec. The PR review is mechanical from there.
This is not a brand-new idea. Behaviour-driven development and contract-first API design both made the same argument a decade ago. What changed in 2026 is that AI editors made the practice cheap enough to actually do on every project.
The three workflows in 2026
Kiro: specs as first-class, EARS-notation native
AWS Kiro is the strongest opinion in the market. Specs are written in EARS notation (Easy Approach to Requirements Syntax), agent hooks trigger on spec edits or file events, and the editor maintains a checkpoint history so you can roll back any AI-driven change. Pricing: Free (50 credits), Pro $20/mo (1,000 credits), Pro+ $40/mo (2,000), Power $200/mo (10,000) (Kiro.dev).
Kiro is the cleanest fit if you build greenfield projects from a brief, or if your client has a structured product team that already writes acceptance criteria. It struggles in messy legacy codebases where the spec layer does not match reality.
Cursor: chat-first with optional rules and AGENTS.md
Cursor's growth has been propelled by chat-first ergonomics and the Composer agent. The spec discipline in Cursor is opt-in: you write rules in .cursorrules or AGENTS.md files that tell the agent what your project's conventions are, and the agent is supposed to obey them. Real-world adherence depends on prompt design.
The 2026 freelance pattern: a small, hand-written AGENTS.md per project that captures the spec at a coarse level, then a fine-grained spec written in the chat or a markdown file for each feature. Cursor's strength is iteration speed. Its weakness is that the spec discipline is on the freelancer to enforce.
Claude Code: terminal-native with the strongest model behaviour
Claude Code took the opposite design bet from Cursor: a terminal-native CLI rather than a wrapped editor. The strength is depth of context handling and obedience to long, structured prompts. The Anthropic model behind it consistently respects multi-step instructions and stays inside the spec more reliably than the alternatives in head-to-head tests.
The 2026 freelance pattern: spec files committed to the repo, often as markdown under /specs, and a Claude Code session driven from those specs with hooks for type-checking and test runs. Combines well with MCP servers for project-specific tool access.
How freelance engineers should pick
The pick is less about the editor and more about the engagement shape.
| Engagement type | Best workflow | Why | |---|---|---| | Greenfield product build | Kiro | Spec-first matches a clean brief | | Feature work in a mature codebase | Claude Code | Better context handling, stronger spec adherence | | Pair programming, exploratory work | Cursor | Fastest iteration loop | | Long-running agent jobs | Claude Code or Cursor background agents | Both have improved this dramatically in 2026 | | Enterprise with audit trail requirements | Kiro | Checkpoints and structured specs satisfy audit |
A freelancer who works across all four engagement types will end up using all three tools. A freelancer who specialises in one engagement type should pick the matching workflow and go deep.
The packaging that wins in 2026
Three offers we see freelance engineers selling against the spec-driven shift, with current price bands.
Spec sprint, then build
A one-week paid sprint to convert a client's brief into a structured specification, with acceptance criteria, edge cases, and a phased build plan. Fixed fee $2,500 to $6,000. The build phase that follows is then priced against the agreed scope.
This is the highest-conversion opening offer in 2026 because the spec sprint produces a concrete deliverable even if the buyer does not move forward with the build. It also gives the freelancer the chance to spot scope problems before pricing the bigger phase.
AI-augmented full-stack delivery
A fixed-price product or feature delivery using a spec-driven workflow as the freelancer's internal practice. Pricing is the same as a traditional engagement, often $15,000 to $80,000 per feature scope, but the delivery is faster, the audit trail is cleaner, and the client gets the same outcome.
This offer wins because the buyer does not care which editor you use. They care about the delivery. The freelancer captures the productivity gain as margin.
Codebase modernisation engagement
A multi-week project to refactor a legacy codebase against a written modern specification, using an AI editor to drive the heavy lifting and a senior engineer to review. Common pricing: $25,000 to $100,000 depending on the surface area. This work was almost unaffordable before AI editors. Spec-driven workflows now make it economic.
What freelance engineers actually charge in 2026
Rates have moved decisively. The signal is the gap between the median Upwork developer rate ($39/hr blended) and the senior AI-native freelance band ($150 to $300/hr on specialist platforms). The same skill in the same person, the same hour of work, but two different rates depending on whether the engagement is sold as "writes code" or as "ships features using an AI workflow."
The shift is not about tools mastery. It is about positioning. Freelancers who quote the same work the same way they did in 2023 anchor to 2023 rates. Freelancers who quote on outcome, specification quality, and delivery speed anchor to a different rate altogether.
Common mistakes the 2025 cohort made
Three traps to avoid.
The first is selling tools instead of outcomes. Telling a client "I use Cursor" is a tools pitch. Telling a client "we go from brief to shipped feature in four weeks with an audited spec" is an outcome pitch. The same work, but the second commands a meaningfully higher rate.
The second is letting the spec live in a chat window. Specs that exist only in an editor session evaporate when the session closes. The clients who pay top rates expect specs to be checked into the repo, version-controlled, and reviewable.
The third is skipping evaluation. The Stack Overflow data is unambiguous: trust in AI-generated code sits at just 33%. Buyers who get burned once by hallucinated logic pay extra for projects that ship with a documented test pass and a regression check. Evaluation is part of the deliverable in 2026, not a separate phase.
Run the engagement like an engineering project, not a vibe
This is the most-overlooked discipline. AI editors made it possible to ship faster than the surrounding process can keep up with. Use a real contract, a signed scope based on the spec, a structured change-order rate, and an acceptance checklist tied to the spec's criteria. If you bill through a workspace tool like Delivvo, put the spec deliverable and the build deliverable in as separate signed milestones with their own files. A clean trail of "spec signed, code shipped, tests passed" is what closes the next engagement.
FAQ
What is EARS notation?
Easy Approach to Requirements Syntax. A structured format for writing requirements as short, testable statements (e.g. "When [trigger], the [system] shall [response]"). AWS adopted it for Kiro because it converts cleanly into evaluation prompts and test cases. The notation itself is from the early 2010s, originally developed for safety-critical software.
Is Claude Code better than Cursor in 2026?
They are different products. Claude Code is terminal-native, stronger on long-context spec adherence, and tends to win on tasks that require careful, multi-step instruction following. Cursor is editor-native, faster on interactive iteration, and wins on tasks that feel like pair programming. Most freelance engineers we see use both, picking based on the engagement.
Does spec-driven development slow you down?
It feels slower for the first hour and is meaningfully faster across a full engagement. The spec sprint typically pays itself back in the first round of changes the client requests, because the freelancer can point to the signed spec to define what is in scope versus out.
Do I need to learn Kiro if I already use Cursor or Claude Code?
Not necessarily. The spec discipline matters more than the tool. A freelancer who writes structured specs in markdown and drives Cursor or Claude Code against them captures most of the benefit. Kiro is a strong opinion for greenfield work but is not required to sell spec-driven engagements.
What about GitHub Copilot in 2026?
Copilot is still the most-installed AI dev tool at work (29% of developers per JetBrains). It is a good baseline for inline completion but does not lead on the spec-driven workflow front the way Kiro, Claude Code, and Cursor do. Freelancers tend to use it as a complement rather than as the primary driver.
The 2026 takeaway
The leading AI editors are converging on spec-driven development from three different angles. Kiro made it the product. Claude Code made it the workflow. Cursor made it the convention. Freelance engineers who pick one of the three workflows, package the spec as a deliverable, and quote on outcome rather than tool are charging multiples of what generalist developers can on the same work. The shift in 2026 is positioning, not productivity.