Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/JoseDiazRohena/soft-factory itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 22, 2026soft-factory
Prompts, slash commands, agents, skills and MCP configs for working with LLMs.
| Folder | What goes there |
|---|---|
prompts/ | Standalone prompts and prompt snippets (any harness) |
mcp/ | MCP server configs and notes |
claude/commands/ | Claude Code slash commands (~/.claude/commands/*.md) |
claude/agents/ | Claude Code subagent definitions (~/.claude/agents/*.md) |
claude/skills/ | Skills (one folder per skill, each with a SKILL.md) |
Everything under claude/ is written for Claude Code. next.md and the agent
instructions also read AGENTS.md, so their text can be reused in other
harnesses, but the frontmatter is Claude-specific.
Keep secrets out of this repo: put API keys in .env files (ignored) or
reference environment variables in MCP configs.
What exactly are we doing here?
Orchestrator
This is a command meant to enable loop engineering workflows. The insight behind loop engineering is to keep the context window small by creating a new agent every time a change is approved and committed. The agent starts fresh and uses the contents of a TODO.md and a SPEC.md at the root of what you're building, along with the state of the code, to figure out what's going on.
One downside of this strategy is that it doesn't let us take advantage of the magic of caching. To improve the cache performance, we don't throw out subagents after a task. Rather, we wait until their context reaches a certain size before refreshing them.
Note: Right now that threshold is purely vibes. Properly done, we would build some instrumentation and testing to find the best balance of cache performance and quality. But that's for later!
How this one works: Invoking the command when Claude Code is in your working repo turns your session into an orchestrator that creates worker and validator agents. Whenever a worker returns from a task, a validator checks the work and raises any issues for your review. It's not fully automated: you have the final say before committing.
Install
Clone the repo and run the install script. It symlinks every command, agent
and skill into your Claude Code config folder, so a git pull updates them
in place.
git clone https://github.com/JoseDiazRohena/soft-factory.git
cd soft-factory
./install.sh
By default it installs into $CLAUDE_CONFIG_DIR, or ~/.claude if that isn't
set. Pass a folder to install somewhere else, for example a single project:
./install.sh ~/code/my-project/.claude
- Files that already exist at the destination (and aren't symlinks) are left
alone and reported as
skip. - Running it again is safe. It refreshes existing links and picks up new files.
prompts/andmcp/aren't installed. Claude Code has no standard folder for them, so use them straight from the repo.
Windows: run the script from Git Bash (comes with Git for Windows) or WSL. Windows only allows symlinks if Developer Mode is on (Settings → System → For developers) or the terminal is running as Administrator. Without either, the script stops with an error instead of quietly copying files.
From WSL, ~ is WSL's own home folder, not your Windows one. If you run
Claude Code on Windows itself, pass the Windows path:
./install.sh /mnt/c/Users/<you>/.claude
To uninstall, delete the symlinks from ~/.claude/commands,
~/.claude/agents and ~/.claude/skills.