Claude Commands (oddballteam)

A small, principled collection of 13 Claude Code slash commands for engineering workflows maintained by the Oddball team for VA.gov and general software development. The repo is notable for its explicit "AI as assistant, not actor" philosophy: every command provides analysis and recommendations but explicitly forbids auto-commits, auto-pushes, and automated production changes. Commands are organized into two folders — general/ (broadly applicable) and veterans-affairs/ (VA.gov-specific). Accuracy guidelines require stating confidence percentages after analysis; attribution guidelines mandate "Co-Authored-By: Claude" in commits. The collection is small (13 commands) but methodologically opinionated.

differs_from_seeds: Most similar to spec-kit's "mirror commands" archetype, but Oddball ships zero skills, zero hooks, and zero MCP. The defining distinction from all seeds is the explicit no-automation mandate as a design constraint (not an accident of size) — comparable in spirit to agent-os's "human-first" ethos but expressed as command-level policies.

Overview — Claude Commands (oddballteam)

Origin

Created by the Oddball team, a government contractor that works on VA.gov (US Department of Veterans Affairs). The commands encode the team's AI usage policies while keeping them useful for any engineering team.

Philosophy (verbatim from README)

"These commands provide AI-assisted guidance for common engineering tasks. They are designed to:

• Guide, not execute - AI provides recommendations and analysis; engineers make decisions and take actions
• Enhance productivity - Streamline research, reviews, and documentation tasks
• Maintain security - No credentials, secrets, or PII handling; all sensitive operations require human action
• Follow best practices - Align with industry standards and compliance requirements"

The "AI as Assistant, Not Actor" principle (verbatim from README)

"These commands are designed with a clear principle: AI assists, humans decide and act.

• Commands analyze code, tickets, and documentation
• Commands suggest approaches and identify issues
• Commands guide through processes with recommendations
• Engineers review AI suggestions before acting
• Engineers execute all commits, pushes, and deployments
• Engineers approve all changes to production systems"

Security principles (verbatim from README)

"- No credential storage - Commands never store or transmit credentials

• No auto-commit/push - All git operations require explicit user action
• No PII handling - Commands follow strict data handling guidelines
• Human review required - All AI suggestions must be reviewed before implementation"

Accuracy guidelines

All commands must state confidence percentage after analysis (90%+ ideal; if below 90%, dig deeper before proceeding).

Attribution guidelines

• Commits: include Co-Authored-By: Claude <noreply@anthropic.com>
• PRs, tickets, reviews, reports: include Generated with Claude Code footer
• Terminal-only output: no attribution needed

Architecture — Claude Commands (oddballteam)

Distribution

Standalone GitHub repo. Install by cloning and copying to ~/.claude/commands/.

Install

git clone https://github.com/oddballteam/claude-commands
cp -r claude-commands/commands ~/.claude/commands/

(No install script; manual copy implied by the repo structure.)

Required runtime

• Claude Code CLI

Directory tree

.
├── README.md
└── commands/
    ├── general/
    │   ├── accuracy-guidelines.md  (reference file)
    │   ├── attribution.md           (reference file)
    │   ├── git-pr.md
    │   ├── pr-review.md
    │   ├── review-docs.md
    │   └── ticket-create.md
    └── veterans-affairs/
        ├── ci-check.md
        ├── collab-review.md
        ├── collab-review-checklist.md  (helper file)
        ├── collab-triage.md
        ├── on-call.md
        ├── sprint-monitor.md
        └── ticket-pointing.md

Target AI tools

Claude Code CLI exclusively.

Config files

None beyond the command files themselves.

Components — Claude Commands (oddballteam)

Commands (13 total)

General commands (4 functional + 2 reference files)

Veterans Affairs commands (6 functional + 1 helper)

Skills

None.

Subagents

None.

Hooks

None.

Scripts

None.

MCP servers

None.

Uniqueness — Claude Commands (oddballteam)

differs_from_seeds

Most similar to spec-kit (command format with YAML frontmatter and numbered steps), but Oddball ships zero skills, zero hooks, and explicitly prohibits automation that spec-kit enables. Against agent-os (also human-first, no automation), Oddball's commands are more narrowly scoped (VA.gov workflows) and even more conservative — agent-os at least auto-writes CLAUDE.md files, which Oddball would not do.

Positioning

A government-contractor's responsible-AI policy encoded as slash commands. The "AI as assistant, not actor" constraint is not just documentation — it's baked into every command's design (no gh pr create auto-invocation, no git push, confidence percentage requirement). Uniquely in this batch, the "no automation" constraint is the primary engineering decision.

Observable failure modes

• No skill equivalents: users who want the commands to auto-trigger without typing /cmd have no path forward.
• VA.gov-specific commands limit portability; collab-review, sprint-monitor, ticket-pointing assume VA.gov tooling (Jira, specific CI/CD).
• No install script: discoverability and deployment are manual.
• 2 stars / dormant-ish: low adoption signals the conservative AI policy may not resonate widely.

Cross-references

No explicit references to other frameworks in the README.