Works with Claude Code · Cursor · OpenCode · Copilot · Codex · Gemini CLI
Spec-driven planning. Parallel TDD execution. A wiki that never forgets.
The Problem
Every AI session starts from zero.
The agent hallucinates architecture it didn't build. Re-solves problems solved three sessions ago. Contradicts decisions made last week. The context window resets. The knowledge vanishes.
Hallucinated architecture
The agent confidently describes code it never wrote, decisions it never made.
Repeated problem-solving
The same bug gets debugged. The same trade-off gets evaluated. Again and again.
Contradictory decisions
Tuesday's decision conflicts with Friday's. Nobody notices until production.
How It Works
Five phases. One discipline.
Define the stack once. Get skill recommendations, library choices, and a production checklist. Every session that follows is grounded in this profile.
Intent before implementation. Requirements expressed as testable SHALL statements with GIVEN/WHEN/THEN scenarios. The spec fits a fresh 200k context window. A grilling session blocks execution until the spec is READY.
One branch per spec. One subagent per task. Each gets a clean context — the spec and nothing else. RED before GREEN, always. Critical review findings block all progress.
Full suite. Every spec scenario has a passing test. Then distill everything learned into the wiki — so the next session inherits the knowledge.
PR. Changelog. Archive the phase. Mark the spec complete. Start the next cycle. The wiki grows. The codebase remembers.
Features
Everything a discipline layer needs.
Spec that fits a context window
Every spec is written to fit a fresh 200k-token window. No executor needs prior chat history. If the plan is too big, it gets decomposed.
TDD without exceptions
Code written before a failing test gets deleted. RED before GREEN. The test suite is the ground truth, not the agent's confidence.
Wiki as persistent memory
After every shipped feature, knowledge is distilled into structured, interlinked pages. Future sessions open the wiki — not raw specs.
Parallel subagent execution
Five agents running in parallel, each with a fresh context, each working from the same spec, each contributing to the same wiki.
Grill before you build
A relentless interview stress-tests every spec against the wiki and techstack. Nothing proceeds until verdict is READY.
Works with every agent
Claude Code, Cursor, OpenCode, Copilot, Codex, Gemini CLI. One install. Skills symlinked to all your agents automatically.
Works with every AI coding agent
One install. Skills symlinked automatically.
Get Started
Stop starting from zero.
Install SuperSpecs once. Every AI session that follows starts informed, not blind. Knowledge compounds. Problems stay solved.