Supervisor is a quality-control layer for Claude Code built around one simple principle:
The first answer should not automatically become the final answer.
It researches externally verifiable claims, delegates work to specialized agents, challenges the result with independent reviewers, searches for hallucinations, runs applicable tests, fixes defects, and submits the result to a final release judge before delivery.
How it works, in 3 steps
- You work normally — give Claude Code a task, project, feature, bug, research question, or frontend request.
- Supervisor investigates and reviews the work — it selects specialists, searches the web when verification is needed, tests applicable claims, challenges assumptions, and sends defects back for correction.
- A final judge decides whether the result can be released — critical defects, unsupported claims, missing requirements, or failed checks return the work to the review loop.
Why
AI-generated work can look correct while still containing:
- outdated documentation
- invented APIs or packages
- unsupported factual claims
- broken code
- incomplete requirements
- hidden edge cases
- security issues
- weak architecture
- poor frontend decisions
- untested assumptions
- false claims of verification
Supervisor is designed to find those problems before the answer reaches you.
Without Supervisor:
USER
|
v
CLAUDE
|
v
FIRST ANSWER
|
v
USER
With Supervisor:
USER
|
v
UNDERSTAND
|
v
RESEARCH
|
v
FAN-OUT SPECIALISTS
|
v
BUILD
|
v
TEST
|
v
INDEPENDENT REVIEW
|
v
ANTI-HALLUCINATION
|
v
RED TEAM
|
v
JUDGES
|
v
FIX
|
v
RE-TEST
|
v
FINAL RELEASE JUDGE
|
v
USER
Web verification
Supervisor treats web research as part of the quality system rather than an optional afterthought.
For substantive answers containing externally verifiable claims, the Supervisor skill is configured to use real web research unless browsing is explicitly prohibited or the task is exclusively a transformation of user-provided material.
The process is:
SEARCH
|
v
FIND AUTHORITATIVE SOURCES
|
v
CHECK FRESHNESS
|
v
CROSS-VERIFY
|
v
COMPARE WITH THE DRAFT
|
v
REJECT UNSUPPORTED CLAIMS
Supervisor prioritizes:
- official documentation
- primary sources
- current information
- repository evidence
- executed tests
- observable runtime results
A failed search does not become a fabricated conclusion.
NOT FOUND != DOES NOT EXIST
If something cannot be verified, Supervisor should mark it as:
UNVERIFIED
instead of inventing an answer.
Anti-hallucination
Supervisor includes an independent truth-verification layer.
It checks potentially fabricated or unsupported:
APIs
PACKAGES
FUNCTIONS
METHODS
FLAGS
VERSIONS
URLs
DATES
NUMBERS
BENCHMARKS
CITATIONS
TEST RESULTS
RUNTIME CLAIMS
TECHNICAL CLAIMS
Important factual and technical work can be challenged by several independent roles:
EVIDENCE RESEARCHER
|
v
SOURCE AUTHORITY JUDGE
|
v
HALLUCINATION HUNTER
|
v
CITATION ENTAILMENT JUDGE
|
v
NUMERICAL AUDITOR
|
v
SKEPTIC JUDGE
The objective is not to make an answer sound more confident.
The objective is to make confidence depend on evidence.
Coding
Supervisor includes a dedicated code implementation specialist:
SUPERVISOR CODE EXPERT
The Code Expert can inspect the repository, understand the existing architecture, implement changes, investigate current documentation, identify root causes, and run available validation.
But the implementer is not allowed to be the only reviewer.
CODE EXPERT
|
v
CODE REVIEWER
|
v
TEST JUDGE
|
v
CORRECTNESS JUDGE
|
v
ARCHITECTURE JUDGE
|
v
SECURITY JUDGE
|
v
PERFORMANCE JUDGE
|
v
INTEGRATION JUDGE
Depending on the project, Supervisor can validate with:
BUILD
TYPECHECK
LINT
UNIT TESTS
INTEGRATION TESTS
E2E
RUNTIME
LOGS
STATIC ANALYSIS
The following statements are not treated as proof:
"should work"
"probably works"
"looks correct"
"likely fixed"
If stronger evidence is available, Supervisor is expected to obtain it.
Frontend
Supervisor also contains a dedicated frontend implementation specialist:
SUPERVISOR FRONTEND EXPERT
Frontend work can be evaluated across:
- visual hierarchy
- typography
- spacing
- composition
- responsive behavior
- accessibility
- interaction states
- loading states
- empty states
- error states
- animation
- motion
- usability
- performance
- design-system consistency
- visual polish
The frontend implementer does not approve its own result.
When rendering or browser inspection is available:
IMPLEMENT
|
v
RENDER
|
v
INSPECT
|
v
VISUAL JUDGE
|
v
ACCESSIBILITY JUDGE
|
v
FIND DEFECTS
|
v
FIX
|
v
RENDER AGAIN
|
v
RE-JUDGE
Source code looking correct is not considered proof that the actual interface looks correct.
Agents
Supervisor currently ships with 27 specialized agents.
| Area | Specialists |
|---|---|
| Implementation | Code Expert, Frontend Expert |
| Research | Researcher, Evidence |
| Requirements | Requirements, Completeness, User Intent |
| Engineering | Code Reviewer, Architecture, Test, Integration |
| Quality | Correctness, Quality, Robustness |
| Risk | Security, Performance |
| Frontend | Visual, Frontend Visual Judge, Accessibility Judge |
| Verification | Source Authority, Hallucination Hunter, Citation Entailment, Numerical Auditor |
| Adversarial | Skeptic, Red Team, Devil's Advocate |
| Release | Final Judge |
Supervisor does not blindly run every agent for every request.
It uses adaptive fan-out.
SIMPLE TASK
|
v
SMALL SPECIALIST GROUP
MEDIUM TASK
|
v
MULTIPLE INDEPENDENT REVIEWERS
COMPLEX TASK
|
v
BROADER SPECIALIST PANEL
HIGH-RISK CODE / FRONTEND TASK
|
v
IMPLEMENTATION
+ TESTING
+ REVIEW
+ ADVERSARIAL CHECK
+ FINAL RELEASE GATE
The important rule is:
IMPLEMENTER != REVIEWER
IMPLEMENTER != FINAL JUDGE
Red Team
Supervisor contains an adversarial reviewer whose purpose is not to justify the work.
Its purpose is to try to break it.
The Red Team searches for:
HIDDEN BUGS
UNSUPPORTED ASSUMPTIONS
MISSING REQUIREMENTS
EDGE CASES
INCORRECT LOGIC
FRAGILE IMPLEMENTATIONS
FAKE FUNCTIONALITY
INTEGRATION FAILURES
SECURITY RISKS
BAD UX
HARDCODED SHORTCUTS
UNVERIFIED CLAIMS
Its central question is:
What would have to go wrong to prove that this result should not be released yet?
Quality gates
Supervisor uses multiple gates before release.
GATE 1 REQUIREMENTS
GATE 2 CORRECTNESS
GATE 3 INTEGRATION
GATE 4 ROBUSTNESS
GATE 5 QUALITY
GATE 6 USER INTENT
GATE 7 EVIDENCE
GATE 8 FINAL RELEASE
Any material failure can return the work to the correction loop.
Judge scoring
Specialist judges can score their dimensions from:
0 - 49 Unacceptable
50 - 69 Weak
70 - 79 Acceptable
80 - 89 Good
90 - 94 Excellent
95 - 100 Exceptional
Default critical threshold:
>= 90
Exceptional-quality target:
>= 95
Defect priority
Supervisor classifies problems by severity:
P0 BLOCKING / CRITICAL
P1 MAJOR
P2 MODERATE
P3 MINOR
P4 COSMETIC
Before release:
ALL P0 MUST BE FIXED
ALL P1 MUST BE FIXED
Material P2 issues should also be corrected when they significantly affect quality, reliability, correctness, or user experience.
Root-cause fixes
Supervisor discourages patch stacking.
Instead of:
BUG
|
v
PATCH
|
v
PATCH
|
v
PATCH
it prefers:
BUG
|
v
ROOT CAUSE
|
v
FIX ARCHITECTURE / LOGIC
|
v
REMOVE WORKAROUNDS
|
v
RETEST
After significant corrections it should also ask:
Why did this happen?
Could the same class of bug exist elsewhere?
Which assumption failed?
Should the tests be expanded?
Did the fix break previously working behavior?
Final release gate
The plugin includes a Stop hook with an independent final release judge.
When Claude is about to finish:
CLAUDE WANTS TO RESPOND
|
v
FINAL RELEASE JUDGE
|
v
CHECK USER INTENT
CHECK REQUIREMENTS
CHECK WEB EVIDENCE
CHECK TEST EVIDENCE
CHECK REVIEW STATUS
CHECK HALLUCINATIONS
CHECK P0 / P1 DEFECTS
|
v
READY?
/ \
NO YES
| |
v v
BLOCK RELEASE
|
v
CONTINUE WORKING
A response should not be released while a known critical blocker remains.
Verification Receipt
Substantive Supervisor responses are expected to include a concise verification summary.
Example:
Verification Receipt
Web
- Official documentation checked
- Current API behavior verified
Local
- Build passed
- Typecheck passed
- Tests passed
Independent review
- Code review completed
- Correctness review completed
- Hallucination check completed
- Final release gate passed
Unverified
- None known
Supervisor must not claim that a search, test, build, render, screenshot, benchmark, subagent, or verification occurred if it did not actually happen.
Installation
Clone or download the repository.
Validate the plugin:
claude plugin validate ./supervisor
Load Supervisor locally:
claude --plugin-dir ./supervisor
Then start Claude Code normally.
You can verify the loaded components with:
/help
/agents
Plugin skills use a namespace.
Usage
Run Supervisor explicitly:
/supervisor:supervisor <your task>
Example:
/supervisor:supervisor Build this React dashboard and do not release it until the implementation, responsive design, accessibility, tests and visual quality have been independently reviewed.
Research example:
/supervisor:supervisor Research the current API implementation using authoritative sources, verify the claims, implement the solution, test it and challenge the final result before answering.
Code example:
/supervisor:supervisor Find the root cause of this bug, verify the framework behavior against current documentation, fix it, run the relevant tests and submit the result to independent review.
Plugin architecture
supervisor/
|
|-- .claude-plugin/
| `-- plugin.json
|
|-- skills/
| `-- supervisor/
| |-- SKILL.md
| `-- references/
|
|-- agents/
| |-- supervisor-code-expert.md
| |-- supervisor-frontend-expert.md
| |-- supervisor-code-reviewer.md
| |-- supervisor-correctness.md
| |-- supervisor-security.md
| |-- supervisor-red-team.md
| |-- supervisor-hallucination-hunter.md
| |-- supervisor-final-judge.md
| `-- ...
|
|-- hooks/
| |-- hooks.json
| `-- hooks.prompt-only.example.json
|
|-- scripts/
| |-- inject-supervision.mjs
| `-- validate.mjs
|
|-- CHANGELOG.md
`-- README.md
Core principle
Supervisor is not designed to find reasons to approve the first result.
It is designed to find reasons why the result should not be approved yet.
Only after the important reasons for rejection have been eliminated should the result reach the user.
DON'T TRUST
|
v
VERIFY
DON'T ASSUME
|
v
TEST
DON'T SELF-APPROVE
|
v
USE INDEPENDENT REVIEW
DON'T SHIP THE FIRST DRAFT
|
v
SUPERVISE IT
Documentation
Claude Code plugin documentation:
https://code.claude.com/docs/en/plugins
Claude Code plugin reference:
https://code.claude.com/docs/en/plugins-reference
Claude Code hooks:
https://code.claude.com/docs/en/hooks
Claude Code skills:
No comments yet
Be the first to share your take.