Propose LCARS threshold adjustments based on accumulated scoring evidence
Manually trigger LCARS pattern consolidation from session summaries
Show LCARS scoring stats, drift history, correction fitness rate, and active patterns
On-demand LLM-as-judge evaluation of response quality against a structured rubric
Scan environment for non-standard CLI tools and show registry status
Matches all tools
Hooks run on every tool call, not just specific ones
Admin access level
Server config contains admin-level keywords
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
Uses power tools
Uses Bash, Write, or Edit tools
Uses power tools
Uses Bash, Write, or Edit tools
A Claude Code plugin that detects and corrects filler, preambles, and low-density responses automatically across conversations.
| Without LCARS | With LCARS |
|---|---|
| "Great question! I'd be happy to help you with that. The capital of France is Paris. Let me know if you need anything else!" | "Paris." |
LLMs default to social interaction patterns: "Great question!", "I'd be happy to help!", sign-offs, hedging, and verbose padding. These patterns are distracting and even manipulative. System prompts help, but they decay over long conversations and the model drifts back toward trained defaults.
LCARS scores every response and injects targeted corrections when drift is detected. It runs in the background with zero perceived latency.
claude plugins marketplace add melek/lcars
claude plugins install lcars@melek-lcars
Verify: /lcars:setup
User query ──→ classify (async) ──→ query type saved
│
Claude responds ──→ score (async) ──→ scores.jsonl
│
drift detection
(query-type-aware)
│ (if drift)
correction selected
from decision table
│
Next user message ──→ correction injected via additionalContext
| Metric | What it measures | Example trigger |
|---|---|---|
| Filler count | 24 filler patterns ("Great question", "Happy to help", "Let me know if") | Any filler phrase in response |
| Preamble position | Words before the actual answer begins | "Sure! I'd be glad to help with that. The answer is..." (11 words of preamble) |
| Info density | Content words / total words | Responses padded with function words score lower |
| Query type | Density threshold | Why |
|---|---|---|
| Default | 0.60 | Research-validated baseline |
| Code | 0.50 | Variable names and syntax naturally lower density |
| Diagnostic | 0.55 | Step-by-step explanation is appropriate |
| Directive | 0.50 | Task commands vary in density |
| Conversational | 0.40 | Follow-ups and acknowledgments are inherently terse |
Drift is classified as low or high severity. The decision table (data/corrections.json) maps drift type × severity × query type to correction templates:
| Layer | When | ~Tokens |
|---|---|---|
| Anchor | Always | 50 |
| Correction | After drift | 0–60 |
| Stats | On resume (>4h gap) | ~30 |
Typical cost: ~50 tokens/session.
A self-contained strategy crystallization system. Instead of requiring manual tuning, LCARS observes its own correction effectiveness and proposes improvements.
observe → validate → crystallize → stage → approve
/lcars:foundryLCARS discovers CLI tools in your environment (jq, rg, gh, etc.) and tracks their usage. Discovered tools can be promoted into session context so the model knows what's available. The registry initializes automatically on first session.
npx claudepluginhub melek/lcars --plugin lcarsForce Claude to re-validate when you have doubts (!doubt)
An agile retrospective skill for your Claude collaboration sessions. Capture what you learned. Apply the fix. Make the next session better.
Learn from conversations with auto-generated entities
Meta-cognition: refine input through brainstorming, refine output through challenge and condensed communication mode.
Ultra-compressed communication mode. Cuts ~75% of tokens while keeping full technical accuracy by speaking like a caveman.
Memory compression system for Claude Code - persist context across sessions