obra/superpowers

Why It's Trending

obra/superpowers [1] is trending as the highest-profile entry in a three-way agent-skills framework surge this week — alongside affaan-m/ECC and mattpocock/skills. Claude Code's rapid mainstream adoption has created intense demand for structured best practices, shared skill libraries, and proven methodologies. obra/superpowers is the most comprehensive publicly available answer: a full SDLC methodology built around the skills/instincts/memory vocabulary.

The 258K stars in 9 months is exceptional velocity for a methodology repo. The 'subagent-driven-development' topic label signals this is positioned as a new category of software engineering practice — not just a prompt collection, but a named development methodology with a defined vocabulary and workflow. This kind of naming and positioning tends to accelerate adoption by giving teams a shared language.

Technical Overview

The framework organizes AI coding agent configuration into four orthogonal dimensions: skills (task-specific instructions and context), instincts (default behaviors that apply across tasks), memory (session and persistent context injection), and security (constraints and guardrails). Each dimension is managed as composable files in a defined directory structure rather than monolithic prompt files.

The 'subagent-driven development' methodology extends this into a workflow: research-first planning using agent subagents, skill-scoped task execution, and structured verification. The Shell language (not a TypeScript or Python library) means these are prompt and script files — intentionally language-agnostic and portable across providers. The 22,990 forks and 322 open issues suggest active community adoption and ongoing development.

For Understory Labs

field-notes already has the skeleton of a skills architecture: enrichment-guide.md (the scoring rubric skill), block-schema.md (the format constraint instinct), and projects.json (the memory layer). obra/superpowers' vocabulary maps cleanly onto this existing structure and suggests it is already well-designed — but could be more explicitly organized.

The most actionable insight is the skills/instincts/memory separation. Currently our context files are flat in the context/ directory without explicit roles. Restructuring them with explicit designations (enrichment-guide.md is a skill, block-schema.md is a constraint instinct, projects.json is memory) would make the architecture more legible and easier to extend when new briefing types are added. For bud, the n8n+CCR pipeline could adopt structured agent skills for the email classification and enrichment steps, making each pipeline stage more auditable.

High-signal methodology framework directly relevant to field-notes' Claude Code skill architecture. Worth a full read of the directory structure to audit our existing context/ files against the skills/instincts/memory decomposition.