convergence

Convergence

A consolidated set of skills, agents, and workflows for Claude Code. Distilled from 6 open-source repos and validated against findings from the Coding Agents Summit 2026.

239+ original items → 15 focused tools. Every skill stays under 35 instructions.

Install

git clone <repo-url> convergence
cd convergence
./setup.sh

This symlinks skills and agents into ~/.claude/skills/ and ~/.claude/agents/. They'll be available in all your Claude Code sessions.

Alternate config directory (e.g. a work profile using CLAUDE_CONFIG_DIR):

CLAUDE_CONFIG_DIR=~/.claude-work ./setup.sh

Project-level install (instead of global):

cd your-project
mkdir -p .claude/skills
ln -s /path/to/convergence/skills/* .claude/skills/

What's Included

Workflow Skills

Use these sequentially for features. Each phase writes a static artifact file — you can resume from any point.

Skill	Instructions	Purpose
/convergence-research	18	Objective codebase exploration. Ticket-blind, fact-only
/convergence-design	22	Alignment discussion with the human. ~200 lines. Human decides
/convergence-outline	16	Vertical structure with phases, signatures, checkpoints
/convergence-implement	24	Execute outline phases with TDD and verification
/convergence-review	28	Code review on actual diff. Defaults to NEEDS WORK

Typical flows:

Small feature:    /convergence-design → /convergence-implement → /convergence-review
Large feature:    /convergence-research → /convergence-design → /convergence-outline → /convergence-implement → /convergence-review
Bug fix:          /convergence-debug → /convergence-tdd → /convergence-compound (if non-obvious)
Pre-ship:         /convergence-review + /convergence-security

Utility Skills

Use standalone, whenever the situation calls for them.

Skill	Instructions	Purpose
/convergence-debug	20	Systematic root cause investigation. No fixes without understanding
/convergence-tdd	14	Test-driven development. Red-green-refactor
/convergence-compound	12	Capture learnings after solving non-trivial problems
/convergence-security	30	Three-layer security audit: access, logging, scanning
/convergence-architecture	22	Layer analysis, callback scoring, god object detection, quality gates

Agents

Dispatched automatically by skills. You don't invoke these directly.

Agent	Instructions	Dispatched By
Research Agent	10	/convergence-research — runs ticket-blind in fresh context
Review Agent	12	/convergence-review — skeptical code reviewer
Security Agent	14	/convergence-security — vulnerability scanner
Learnings Researcher	10	/convergence-research, /convergence-debug — surfaces past learnings

Safety Hooks

The .claude/settings.json includes a hook that warns before destructive commands (rm -rf, DROP TABLE, git push --force, git reset --hard). Copy it to your project's .claude/settings.json to enable.

Design Principles

These aren't just guidelines — they're hard constraints that shaped every skill in this repo.

Instruction Budget. Every skill stays under 35 instructions. LLMs reliably follow ~150-200 instructions total. At 35 per skill plus system overhead, you can stack 2-3 skills without degradation.

Human In The Loop. Workflows force the agent to surface assumptions for human correction before writing code. No automated decision pipelines. The engineer makes design choices on a 200-line doc, not in 2,000 lines of code.

Static Artifacts. Every workflow phase writes output to a file. This survives context compaction, enables session resumption, and allows human review without context window dependency.

Verify The Code. Review actual code, not plans. "Confidence is not evidence." Run the command, read the output, then make the claim.

Vertical Slicing. Plans and implementations follow vertical slices (end-to-end with checkpoints), not horizontal layers (all DB, then all API, then all frontend). Each phase is independently testable.

Codebase First. Always scan existing patterns before proposing changes. Research is objective compression of truth — separate "what are we building" from "what exists."

Security By Default. Three layers: scoped credentials before starting, audit logging during, vulnerability scanning after.

Context Budget

Keep your active instruction total under 60 per session. CLAUDE.md, system prompt, and tool definitions use ~80-100 of the ~200 instruction budget. That leaves ~100-120 for skills.