Context Engineering Prover MCP Connector for Claude

Context engineering is not "include everything" — it is the disciplined selection, ordering, and budgeting of information to maximize model performance. You must: (1) audit RELEVANCE — for each context block, state its specific purpose and what BREAKS if the block is removed. "All files are relevant" is the opposite of engineering. Every included block must pass the REMOVAL TEST: remove it, run the task, observe degradation. If performance does not degrade, the block is waste that dilutes attention on critical context, (2) map STRUCTURE — priority-order blocks from highest to lowest importance, use semantic delimiters (<SYSTEM_CONTEXT>, <SCHEMA>, <EXAMPLES>), and label each block's role (required/reference/fallback). First-position tokens receive 3x attention weight vs middle-position — put critical context first. "Paste it in" is not structure, (3) specify BOUNDS — total token budget, per-block allocation (tokens and percentage), response headroom (minimum 10% for model output), and waste ratio (included-but-unreferenced tokens). "It fits in the window" ignores that attention quality degrades in long contexts — a 4K context with high relevance outperforms a 32K context with 60% waste, (4) ground INSTRUCTIONS — cite evidence for each major decision: test results (A/B, eval suite), documented patterns (research papers, framework docs), or measured improvements (accuracy delta). "Best practice" and "usually works" are not evidence — show the measurement, (5) define MEASUREMENT — metric name, measurement method, baseline (before context engineering), target (after), and measurement cadence (when do you re-evaluate?). "The output looks better" is not measurement. "Task accuracy: 62% → 85% on 50 SQL eval cases" is measurement. If rejected, your context has a structural flaw that degrades model performance. Structured reflection tool for rigorous context engineering before constructing any prompt. Forces the agent to audit every context block for relevance, structure context with priority ordering and semantic delimiters, specify token budgets with per-block allocation and waste analysis, ground instruction decisions in measurable evidence, and define quantifiable quality metrics. Catches Context Dumping (including everything without justification — "all files are relevant" when 40% of tokens go to unreferenced context), Unstructured Context (no priority ordering, no delimiters, no role labels — "just paste it in" causes attention decay on critical information), Unbounded Context (no token budget — "it fits in the window" ignores that attention degrades 15-20% for middle-position tokens in long contexts), Ungrounded Instructions (vibes-based choices — "best practice" without citing test results, documented patterns, or measured improvements), and Unmeasured Quality (no metric — "looks good" instead of task accuracy on eval cases with baseline and target). Call once per prompt/context construction or evaluation