PRISM

██████╗ ██████╗ ██╗███████╗███╗   ███╗      ╱│
██╔══██╗██╔══██╗██║██╔════╝████╗ ████║     ╱ │  ━━━━━ Plan
██████╔╝██████╔╝██║███████╗██╔████╔██║    ╱  │ ━━━━━━ Research
██╔═══╝ ██╔══██╗██║╚════██║██║╚██╔╝██║   ╱   │━━━━━━━ Implement
██║     ██║  ██║██║███████║██║ ╚═╝ ██║  ╱    │ ━━━━━━ Synthesize
╚═╝     ╚═╝  ╚═╝╚═╝╚══════╝╚═╝     ╚═╝ ╱_____│  ━━━━━ Maintain

A toolkit of commands, sub-agents, and skills for AI coding workflows.

PRISM installs a curated set of slash commands, sub-agents, and a skill into your AI coding tools. It is currently supported in Claude Code (commands + agents + skills) and Windsurf (commands installed as workflows).

The acronym tracks the five phases of a structured coding workflow:

Phase	What it covers
Plan	Produce a detailed, file-grounded implementation plan and iterate on it
Research	Document the codebase as-is — locate, analyze, find patterns
Implement	Execute approved plans with small, well-defined coder/architect units
Synthesize	Hand off context between sessions, resume work, capture decisions
Maintain	Validate that work matches the plan; review colleagues' branches locally

---

Install

./install.sh                  # interactive prompt
./install.sh --claude         # install Claude Code only
./install.sh --windsurf       # install Windsurf workflows (to CWD)
./install.sh --all            # install everywhere
./install.sh --help

Destinations:

• Claude Code → $CLAUDE_HOME (defaults to ~/.claude)
  • commands/ → ~/.claude/commands/
  • agents/ → ~/.claude/agents/
  • skills/ → ~/.claude/skills/
• Windsurf → $WINDSURF_PROJECT_DIR (defaults to ./.windsurf)
  • commands/ → ./.windsurf/workflows/
  • Windsurf does not natively support sub-agents or skills — only commands install as workflows.

Existing files at the destination are renamed to <name>.bak.<timestamp> before being overwritten, so re-running the installer is non-destructive.

Environment overrides

Variable	Purpose	Default
CLAUDE_HOME	Claude Code install root	$HOME/.claude
WINDSURF_PROJECT_DIR	Windsurf .windsurf directory	$PWD/.windsurf

---

What gets installed

Commands (commands/)

Slash commands in Claude Code (/<name>), or workflows in Windsurf.

Command	Description
/create_plan	Interactively build a detailed implementation plan grounded in the codebase (uses opus)
/iterate_plan	Update an existing plan based on feedback, keeping it grounded in reality (uses opus)
/implement_plan	Execute an approved plan phase-by-phase with verification
/validate_plan	Verify an implementation against its plan and surface deviations
/research_codebase	Document the codebase as-is via parallel sub-agents (uses opus)
/create_handoff	Write a concise handoff document to transfer work to another session
/resume_handoff	Resume work from a handoff document with context recovery
/local_review	Set up a worktree to review a colleague's branch (gh_username:branchName)

Sub-agents (agents/) — Claude Code only

Specialized agents invoked via the Task tool with subagent_type.

Agent	Role	Model
architecture-agent	Makes pattern/design judgment calls; decomposes broad work into coder-agent-sized sub-tasks. Read-only.	opus
coder-agent	Implements small, well-defined coding tasks with precise specs and acceptance criteria.	sonnet
codebase-locator	Finds where code lives — files, directories, components.	sonnet
codebase-analyzer	Explains how code works with file:line references. Documentarian, not critic.	sonnet
codebase-pattern-finder	Surfaces concrete examples of existing patterns to model new work after.	sonnet
thoughts-locator	Discovers relevant documents in the ~/thoughts/ directory.	sonnet
thoughts-analyzer	Deep-dives into specific thoughts documents to extract key insights.	sonnet
web-search-researcher	External research via WebSearch / WebFetch.	sonnet

The research/analyzer agents are strict documentarians — they describe what exists rather than recommending changes. architecture-agent and coder-agent are the action-oriented pair: architect decides and decomposes, coder executes.

Skills (skills/) — Claude Code only

Skill	Description
rlm-recursive-context	Recursive Language Models-style handling for inputs that don't fit in one pass: treat long inputs as external objects, slice them, and stitch results from sub-call variables.

---

How the pieces fit together

A typical PRISM session walks through the acronym:

1. Research — /research_codebase (or codebase-locator + codebase-analyzer directly) to map out the relevant area.
2. Plan — /create_plan to produce a phased plan; /iterate_plan to refine it.
3. Implement — /implement_plan drives execution, delegating to architecture-agent and coder-agent for the actual edits.
4. Synthesize — /create_handoff at the end of a session; /resume_handoff at the start of the next.
5. Maintain — /validate_plan to confirm the implementation matches the plan; /local_review to inspect a colleague's branch in an isolated worktree.

The ~/thoughts/ workspace

PRISM commands persist artifacts under ~/thoughts/{project_name}/, where {project_name} is basename $(pwd) of the working directory. Subdirectories used by the commands:

• ~/thoughts/{project_name}/research/ — research documents
• ~/thoughts/{project_name}/plans/ — implementation plans
• ~/thoughts/{project_name}/tickets/ — ticket files
• ~/thoughts/{project_name}/handoffs/ — handoff documents

This keeps long-lived planning artifacts outside the project repo while still scoping them per-project.

---

License

Apache License 2.0 — see LICENSE.