OPC — One Person Company

A full team in a single Claude Code skill. You're the CEO — OPC is everyone else.

11 specialist agents (PM, Designer, Security, Tester, and more) that review, analyze, build, and brainstorm your code — so you don't have to context-switch between hats.

Why not just ask Claude directly?

You can — and for code-level bugs, a single Claude prompt is often more thorough. We tested both on this repo:

| | Single Claude | OPC (3 agents) | |---|:---:|:---:| | Code bugs (variable shadowing, DRY violations, exit codes) | 14 found | 9 found | | UX issues ("new user runs opc review in terminal expecting a CLI command") | 0 found | 5 found |

Single Claude found more code issues. OPC found different types of issues — things that require thinking from a specific persona's perspective. A security engineer looks for content exposure. A new user tries the install flow and gets confused. A DevOps engineer checks npm packaging. Claude won't switch to these mindsets unless you explicitly ask.

OPC's value isn't finding more bugs. It's finding bugs you wouldn't think to look for.

How It Works

One principle: the agent that does the work never evaluates it.

1. Task inference — OPC reads your request and picks the right pipeline: review, analysis, build, brainstorm, plan, or full pipeline. Every path ends with independent evaluation.

2. Parallel specialists — 2-5 role-specific agents run in parallel, each with domain expertise and anti-patterns (what NOT to flag). They don't see each other's output.

3. Verification gate — the orchestrator checks agent findings: verifies facts, challenges severity, dismisses false positives, and synthesizes a verdict (PASS / ITERATE / FAIL).

4. Iteration loop — if the verdict is FAIL or ITERATE, the implementer fixes or polishes, and the evaluator re-tests. Up to 10 rounds, with early exit on oscillation.

Quick Start

Install

npm install -g @touchskyer/opc

Skill files are automatically copied to ~/.claude/skills/opc/. If the postinstall fails, run opc install manually.

Manual install (no npm)

git clone https://github.com/iamtouchskyer/opc.git
ln -s $(pwd)/opc ~/.claude/skills/opc

Note: Symlink tracks the repo — git pull updates the skill immediately. Use cp -r instead if you want a stable snapshot.

Use it

# Review a PR
/opc review the changes in this PR

# Analyze an architecture problem
/opc analyze why the API is slow

# Execute with a plan
/opc implement the migration plan in PLAN.md

# Brainstorm approaches
/opc what are our options for auth?

# Interactive mode — agents ask you questions first
/opc -i review the payment flow

# Explicit roles
/opc security compliance

# Open the report viewer
/opc replay

Task Types

| Task type | When | Pipeline | |-----------|------|----------| | Review | PR review, audit, pre-launch | Context brief → multi-role evaluation → verification gate → report | | Analysis | Architecture, performance, diagnosis | Context brief → single deep-role evaluation → report | | Build | Direction is set, implement it | Plan → build → independent evaluation → iterate until PASS | | Brainstorm | Options, trade-offs, alternatives | Role perspectives → comparison table → evaluation → recommendation | | Plan | Scope, decompose, estimate | Task decomposition → evaluation → report | | Verification | QA, test, pre-release check | Context brief → multi-role evaluation (verification tags) → report | | Post-release | User test, onboarding check | Context brief → multi-role evaluation (post-release tags) → report | | Full pipeline | Complex or vague request | Design → Plan → Build → Evaluate → Deliver |

Built-in Roles

Product

| Role | Focus | |------|-------| | PM | Requirements, user value, scope, prioritization | | Designer | Interaction design, information architecture, visual system, accessibility |

User Lens

| Role | Focus | |------|-------| | New User | First impression, onboarding, setup friction, trust signals | | Active User | Workflow efficiency, power features, scale behavior, customization | | Churned User | Re-entry experience, change communication, win-back signals |

Engineering

| Role | Focus | |------|-------| | Frontend | Component architecture, framework patterns, performance, i18n, type safety | | Backend | API design, database, auth, input validation, data consistency | | DevOps | CI/CD, containers, deployment, secrets, monitoring, developer experience |

Quality

| Role | Focus | |------|-------| | Security | Vulnerabilities (OWASP), dependency audit, secrets, auth security, attack surface | | Tester | Boundary cases, state coverage, regression risk, integration points | | Compliance | GDPR/CCPA, WCAG accessibility, license compatibility, industry regulations |

Custom Roles

Add a .md file to roles/ following this format:

# Role Name

## Identity
One sentence: who you are and what you care about.

## Expertise
- **Area** — what you know about it
- **Area** — what you know about it
...

## When to Include
- Condition that triggers this role
- Condition that triggers this role
...

The coordinator reads When to Include to decide whether to dispatch your role. It's available immediately — no configuration needed.

If a task needs expertise not covered by any role file, the coordinator creates a temporary role on-the-fly.

How Review Works

You: /opc review this PR

1. Triage     → Task type: review
2. Roles      → Frontend, Backend, Security (auto-selected from changed files)
3. Brief      → Orchestrator builds context from git log, CLAUDE.md, specs
4. Dispatch   → 3 role evaluators run in parallel, each with role expertise + context
5. Verify     → Mechanical checks auto-reject incomplete outputs (no file:line, no VERDICT).
                 Orchestrator spot-checks facts, challenges severity, deduplicates.
6. Report     → Curated findings with severity, file:line references, and fix suggestions

Requirements

• Claude Code (CLI, desktop app, or IDE extension)
• Node.js ≥ 18 (for npm install only — not needed if you install manually)
• That's it. No runtime dependencies, no build step, no MCP server. Just markdown files.

Works better with memex (optional)

OPC works standalone — but pair it with memex and it learns across sessions. Memex remembers which roles were useful, which findings were false positives, and your project-specific context. OPC doesn't need to know how memex works — memex drives itself.

npm install -g @touchskyer/memex

Visualize reports (optional)

OPC saves structured reports to ~/.opc/reports/ after every run. Browse them in a web UI:

npx @touchskyer/opc-viewer

Or use /opc replay in Claude Code to open the viewer automatically.

The viewer shows a Slack-like replay of your review team's conversation, plus a filterable summary of findings by severity.

License

MIT