Research first. Edit safely. Verify before done. Remember what worked. The 7 Laws of AI Agent Discipline — runtime hooks, enforcement skills, and project memory.
npx continuous-improvement install
Runtime discipline that stops agents from editing before understanding.
Investigate the codebase, read the spec, and list facts before touching a single file.
Each commit addresses exactly one thing. No drive-by refactors mixed with bug fixes.
Write the plan, get alignment, then code. Never code your way to understanding.
Run tests, typecheck, and lint before declaring completion. Proof beats promise.
What worked? What failed? Capture lessons so the next agent starts smarter.
One change → verify → next. Never carry a failure forward into the next task.
Store instincts, patterns, and decisions in project memory. Yesterday's lesson is today's shortcut.
Three layers of enforcement that physically block bad behavior.
A PreToolUse hook that blocks Edit / Write / Bash until the agent presents a fact-list investigation. No more editing before reading.
Six-phase verification ladder — typecheck, test, build, lint, e2e, healthcheck — before any task is marked done.
Instinct packs capture lessons across sessions. The same mistake never repeats twice.
Autonomous agent execution that converts PRDs to executable JSON and runs until all stories pass.
Fan-out to code-reviewer + security-auditor + test-engineer in parallel, then synthesize a go/no-go decision.
GitHub Action that lints agent transcripts for discipline violations before merge.
Install in 30 seconds. No Node setup required — just two slash commands inside Claude Code.