| name | yellowteam |
| description | Run a yellow team AI-smell review against a target codebase or set of documents. Deterministic 12-rule audit against the harness writing style guide (No Not-X-but-Y, no em dashes, no rhetorical tetracolons, no cinematic short sentences, no rhetorical anchors, banned vocabulary, no rule-of-three flourishes, no participial tails, no ensure-hedge, no vague intensifiers, no decorations/emojis, no AI smell). Produces an AI Smells Report (JSON + MD + HTML) per target. Usage: /yellowteam [--target /path/to/repo] [--scope all|prose|code] |
| user-invocable | true |
Yellow Team AI-Smell Review
Run the full deterministic style audit against a target. Yellowteam mirrors
blueteam (defensive), redteam (offensive), and greenteam (code quality)
with a fourth dimension: writing-style integrity.
The framework runs 12 scanners, one per hard rule from the harness
writing style guide, and produces an AI Smells Report with per-file
findings, severity, the offending quote, and a concrete rewrite.
Usage
/yellowteam # Audit ./ with full scope
/yellowteam --target /path/to/repo # Audit a different target
/yellowteam --scope prose # Only README / .md / .txt / docs (skip code)
/yellowteam --scope code # Only source-file comments + docstrings
/yellowteam --skip rule07,rule10 # Skip noisier rules
Prerequisites
cd .claude/security/yellowteam && npm install
Pipeline
Execute these steps in order. Read .claude/security/yellowteam/CLAUDE.md
for the full protocol.
| Step | Phase | Skill File |
|---|
| 1 | Deterministic per-rule scanners (12 rules) | skills/01-deterministic-scan-agent.md |
| 2 | AI judgement pass (filter over-fires, clear noise) | skills/02-judgement-pass-agent.md |
| 3 | Report generation (MD + HTML AI Smells Report) | skills/03-report-generation-agent.md |
End-to-end orchestrator
cd .claude/security/yellowteam
node pipeline/run_all.js --target /path/to/repo --verbose
node scripts/report_generator.js
node scripts/validate_report.js
Output drops:
deliverables/yellowteam_findings.json: canonical machine-readable
deliverables/yellowteam_findings.md: human-readable per-file findings
deliverables/yellowteam_findings.html: browsable AI Smells Report
deliverables/per-scanner/rule*.json: raw per-rule output
Severity scale
HIGH | MEDIUM | LOW | INFO. The four rules with default-HIGH severity
(Not-X-but-Y, tetracolons, rhetorical anchors, decorations/emoji) are
near-certain AI tells. MEDIUM rules are strong signals with occasional
legitimate uses. LOW rules over-fire, the judgement pass must filter.
When to invoke
Auto-invoke before producing any:
- README, CHANGELOG, RUNBOOK
- ADR or architecture decision
- Release notes
- Training material
- External-facing communications
- Customer-facing documentation
- Sales / marketing copy
- LinkedIn posts
- Any prose intended for a non-internal audience
Manually invoke during periodic codebase audits or before a major release.
Cross-framework integration
After yellowteam completes, findings can inform:
- Blueteam findings in user-facing security communications (those need
HIGH style cleanliness ,
/yellowteam should clear them)
- Greenteam findings in CLAUDE.md / README updates (avoid recap loops
when documenting harness rule changes)
- Documentation for new skills, the style guide applies to skill
description: and SKILL.md prose
Safety rules
- Do not modify the target's source files; yellowteam is read-only
- Do not delete cleared findings, keep them with
status: cleared so the
audit reader sees what was reviewed-and-dismissed vs. not-reviewed
- Do not skip the judgement pass on documents that will be published.
rules 7 and 10 over-fire and need filtering
- All HTML reports must be script-generated, never hand-crafted