Skip to main content

Why SpecWeave?

AI changed how we write code. SpecWeave changes how we ship products.

Every AI coding tool generates code fast. But generating code was never the hard part. The hard part is: what happens after the chat ends?


The Problem Everyone Has

You open Claude Code (or ChatGPT, or Copilot). You describe a feature. AI generates code. It works. You move on.

Two weeks later:

  • "Why did we use JWT instead of sessions?" - Nobody remembers.
  • "What were the acceptance criteria?" - Lost in chat history.
  • "How does this auth system work?" - Ask Claude again and hope you get the same answer.

This is "vibe coding" - and it's how 90%+ of developers use AI today.

The 5 Failure Modes of Vibe Coding

Failure ModeWhat HappensReal Cost
Context EvaporationAI conversations end, decisions vanishTeams repeat conversations. New hires have zero context.
Scattered CodeEach request produces isolated code with no shared patternsTechnical debt from day 1. Constant refactoring.
No Quality GatesCode ships without tests, reviews, or documentationBugs in production. Security vulnerabilities.
No TraceabilityCan't trace code back to requirements"Why does this exist?" becomes unanswerable.
Team ChaosMultiple developers generate conflicting implementationsIntegration hell. Onboarding takes weeks.
Context Exhaustion"Prompt is too long" kills your session mid-work10-15 min re-explaining per restart. Same decisions repeated.

The Hidden Math

Vibe Coding a Feature:
2 hours building
4 hours fixing bugs (no upfront design)
3 hours explaining to teammates (no docs)
─────────────────────
9 hours total

SpecWeave:
30 min planning (spec + plan + tasks)
1.5 hours building (guided by spec)
0 hours explaining (docs already exist)
─────────────────────
2 hours total

Industry Validation: "No Vibes Allowed"

SpecWeave's approach isn't theoretical — it reflects an emerging industry consensus. In "No Vibes Allowed: Solving Hard Problems in Complex Codebases", Dex Horthy (HumanLayer, YC F24) presents the Research-Plan-Implement (RPI) methodology — a structured workflow that independently validates SpecWeave's core design:

RPI ConceptSpecWeave Equivalent
Research — investigate the codebase before touching codeBuilt-in: PM scans existing docs, ADRs, and project structure before writing anything. Deep Interview mode for complex features. sw:brainstorm for multi-approach exploration
Plan — detailed markdown specs with file locations and test strategyspec.md + plan.md + tasks.md — the three-file structure
Implement — execute the validated plansw:do or sw:auto — autonomous execution from the plan
One step at a time — incremental, focused work unitsIncrements — each feature is a self-contained unit of work
Human reviews the plan, not the codeYou review spec + plan at ExitPlanMode, AI handles implementation
"Overkill sometimes" — scale workflow to match problem complexityComplexity ceiling: simple features skip brainstorm, complex ones use full RPI with agent teams

The key insight: bad plans generate orders of magnitude more problems than bad code. Errors amplify through each phase — a small mistake in research becomes a bigger mistake in the plan, which becomes an even bigger mistake in code:

Errors compound at each stage. A small flaw in research produces a proportionally larger flaw in the plan, which produces an even larger proportion of bad code. This is why human review should concentrate on specifications — the highest-leverage checkpoint where misunderstandings are cheapest to correct.

As Sean Grove puts it: "Specs are the new code."

Where Human Review Matters Most

SpecWeave places human review at the plan level, not the code level. You review spec.md and plan.md — the artifacts that matter most — while AI handles implementation. For complex or uncertain features, say "let's brainstorm this first" (or sw:brainstorm) to add a research phase before planning, giving you a second review checkpoint:

SpecWeave takes this further: as long as the spec and plan are solid, and progress is tracked through external tools (GitHub, JIRA, ADO), the implementation follows naturally — one increment at a time.


What SpecWeave Does Differently

SpecWeave wraps every AI coding session in a spec-driven workflow that turns conversations into permanent, searchable documentation.

Skills Are Structured Expertise

Unlike prompts that you copy-paste, SpecWeave skills are reusable, structured instructions — each one defines how AI thinks and acts for a specific domain. ~48 built-in skills ship out of the box. Customize any skill without forking via skill-memories/. Deep dive →

Start Building in Minutes

Skills encode expertise so you don't need to learn it yourself. Install SpecWeave, describe your feature, and skills handle planning, testing, and documentation. Learn more →

Enterprise Ready from Day One

Compliance audit trails in git. Brownfield codebase analysis. Bidirectional sync with GitHub, JIRA, Azure DevOps. Multi-repo coordination. Enterprise overview →

The Three Files

Every feature produces three files that persist across sessions:

FilePurposeOwner
spec.mdWHAT: User stories, acceptance criteriaPM Agent
plan.mdHOW: Architecture decisions, tech choicesArchitect Agent
tasks.mdDO: Implementation tasks with test plansDeveloper
.specweave/increments/0001-user-auth/
├── spec.md ← "What are we building and why?"
├── plan.md ← "How should we build it?"
└── tasks.md ← "What's the step-by-step plan?"

6 months later: Search "OAuth" and find exact decisions, who approved them, and why they were made.

Autonomous Execution

SpecWeave doesn't just plan - it executes autonomously for hours:

"I want to add user authentication. Then ship it while I sleep."

What happens during autonomous execution:

  1. Reads the spec and tasks
  2. Implements each task in order
  3. Runs tests after each task
  4. Fixes failures automatically
  5. Syncs progress to GitHub/JIRA
  6. Moves to next task
  7. Repeats until done

You review finished work, not work-in-progress.

68+ Specialized AI Agents

SpecWeave isn't one AI assistant - it's a team of 68+ specialized agents powered by Claude Opus 4.6:

AgentRoleWhen It Activates
PMRequirements, user stories, acceptance criteria"Let's build [feature]" or sw:increment
ArchitectSystem design, ADRs, tech stack decisionsComplex features
QA LeadTest strategy, quality gates, validation"Check quality" or sw:validate
SecurityOWASP review, vulnerability detectionMention "security"
DevOpsCI/CD, infrastructure, deploymentInfrastructure work
Tech LeadCode quality, patterns, refactoring"Review the code" or sw:grill
FrontendReact/Vue/Angular architectureFrontend work
Backend.NET/Node.js/Python patternsBackend work

Agents auto-activate based on context. Just describe what you need -- mention "security" and security expertise loads, say "let's build X" and the PM activates. No configuration needed.


SpecWeave vs Alternatives

vs Raw AI Coding (ChatGPT / Claude / Copilot)

DimensionRaw AI CodingSpecWeave
MemorySession-onlyPermanent specs
PlanningNone (you decide what to ask)Structured spec + plan + tasks
QualityHope it works3-gate validation
DocumentationNone (maybe comments)Auto-generated living docs
Team collaborationShare chat links?Shared specs + JIRA/GitHub sync
Autonomous workCopy-paste one response at a timeHours of unattended execution

vs GitHub SpecKit

SpecKit formalizes spec-driven development. SpecWeave extends it into a full lifecycle:

DimensionSpecKitSpecWeave
ScopeOne feature at a timeMulti-increment lifecycle
After spec creationManual implementationAutonomous execution
External toolsNoneGitHub, JIRA, Azure DevOps
DocumentationStatic snapshotsLiving docs (auto-update)
Codebase supportGreenfieldGreenfield + Brownfield
Quality gatesDeveloper disciplineAutomated 3-gate enforcement

In set theory: SpecKit ⊂ SpecWeave. One SpecKit run = one SpecWeave increment.

Full comparison

vs Manual Spec-Driven Development

You could write specs by hand. SpecWeave automates what's tedious:

  • Auto-generates spec.md, plan.md, tasks.md from a one-line description
  • Auto-links tasks to acceptance criteria (bidirectional AC-IDs)
  • Auto-syncs progress to GitHub Issues, JIRA, Azure DevOps
  • Auto-validates before closing (tests pass, docs updated, ACs met)
  • Auto-updates living documentation after every completed task

When SpecWeave Is Overkill

Be honest: SpecWeave isn't for everything.

ScenarioUse SpecWeave?Why
Quick script / one-off toolNoJust vibe code it
Learning exercise / tutorialNoOverhead not worth it
Throwaway prototypeNoIt's throwaway
Personal weekend projectMaybeIf you want to learn the workflow
Feature for production appYesSpecs prevent bugs and tech debt
Team projectYesSpecs enable collaboration
Enterprise systemYesCompliance, audit trails, traceability

Rule of thumb: If someone else will read your code, or if you'll maintain it for more than a month, use SpecWeave.


Getting Started

npm install -g specweave
cd your-project
specweave init .

Then describe what you want:

"I want to add user authentication"

SpecWeave handles the rest.

Quick Start Guide | Academy (16 lessons) | All 100+ Skills