lcars

LCARS

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."

Why

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.

Install

claude plugins marketplace add melek/lcars
claude plugins install lcars@melek-lcars

Verify: /lcars:setup

How it works

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

1. Anchor — Every session starts with a ~50-token behavioral anchor: answer first, no filler, no affect simulation.
2. Classify — Each user query is classified by type (factual, code, diagnostic, directive, conversational, etc.) so thresholds adjust appropriately.
3. Score — After each response: filler phrase count, preamble word position, information density.
4. Detect drift — Scores are checked against query-type-aware thresholds. Code responses have a lower density bar; factual responses are held to a higher standard.
5. Correct — If drift is detected, a correction is injected on the user's next message within the same session.

What gets scored

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 thresholds

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

Corrections

Drift is classified as low or high severity. The decision table (data/corrections.json) maps drift type × severity × query type to correction templates:

• Filler detected → "Prior response contained filler. Omit."
• Low density on a code query → no correction (expected)
• Low density on an ambiguous query → no correction (too heterogeneous to correct reliably)
• Compound high drift → full behavioral reset

Context budget

Layer	When	~Tokens
Anchor	Always	50
Correction	After drift	0–60
Stats	On resume (>4h gap)	~30

Typical cost: ~50 tokens/session.

Foundry

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

1. Scoring and drift data accumulate over sessions
2. The PreCompact hook consolidates summaries and identifies recurring patterns
3. Overfit gates prevent premature crystallization: 5+ sessions across 3+ calendar days
4. The Foundry proposes new or refined strategies based on correction fitness
5. Proposals are staged — never auto-applied. Review with /lcars:foundry

Tool Registry

LCARS 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.

Skills