Agent Context Impact Analyzer — AGENTS.md / CLAUDE.md Efficiency Benchmark

📝 Repository Prompt Context (AGENTS.md / CLAUDE.md) 342 Tokens

🎯 Injection Strategy Strategy

Full AGENTS.md File Inject complete repo instructions verbatim into context.

Modular On-Demand Include architecture & test rules only for active file scope.

System Prompt Direct Core constraints injected directly as system message.

Minimal Baseline Strip all formatting/style guidelines; retain core logic.

🔍 Rule Category Toggles & Noise Diagnostics

Toggle categories off to simulate context optimization. High formatting overhead produces zero accuracy lift.

Total Prompt Tokens 342 Context overhead

Signal / Noise Ratio 38.5% 210 noise tokens

Primary Bottleneck Formatting & Style 61.4% total noise

📊 Token Category Distribution Breakdown

Context Token Allocation by Rule Type

🧪 Simulated Task Correctness vs Context Overhead

Strategy	Tokens	Claude Code	Codex	Acc Delta

* Based on 288 gold-test evaluation runs. Correctness does not move statistically between full context and minimal baseline (+0.3-0.5%).

✅ Active Optimization Proof

Signal Ratio 38.5%

Redundant Tokens 210

Claude Delta +0.5%

🚀 Optimized AGENTS.md Output

Context analyzer: token heuristics, strategy assumptions and fixed scores

Your browser does not support video.

Read the explanation

This saved context analyzer estimates token count by splitting trimmed text on whitespace, multiplying the word count by one point three, and rounding. Ten words therefore estimate thirteen tokens, while twenty estimate twenty six. Bars show one hundred thirty and two hundred sixty pixels at ten pixels per estimated token. Empty or whitespace only text estimates zero. This is a heuristic, not an actual tokenizer, and it computes each section separately before summing. Sections begin only at lines starting with two hash marks followed by a space. Architecture and testing are classified by their heading words, while task or execution marks the task category. Every other heading, plus introductory text, defaults to formatting and is labeled noise. That automatic label does not prove an instruction is unnecessary. A security requirement under an unrecognized heading could be categorized as formatting by this simple parser. Enabled categories contribute either signal or noise tokens. The modular strategy multiplies the combined total by zero point four five and rounds; system direct multiplies by zero point three five. For an illustrative total of one hundred, the display becomes forty five or thirty five. Bars compare two hundred seventy and two hundred ten pixels at six pixels per estimated token. These multipliers are assumed savings, not measured retrieval or a real model context. The signal ratio is still signal divided by the original enabled total, without using the adjusted effective token count. The exported Markdown is assembled from enabled original sections. Selecting modular or system direct does not shorten those sections by its numerical multiplier. Minimal strategy instead disables formatting; returning to full enables formatting again. Other category toggles are maintained, and reset restores all categories. Preserve necessary project instructions rather than trusting this heuristic to decide their value. The comparison table stores fixed correctness values. Its Claude baseline is sixty four point two percent and full file is sixty four point seven, a difference of zero point five percentage points. Bars use four pixels per stored percentage point, giving two hundred fifty six point eight and two hundred fifty eight point eight pixels. These are fixture values, not evaluations performed on edited instructions. The Codex values are likewise fixed at sixty one point five and sixty one point eight. Only the currently selected row substitutes the calculated token total; other rows retain hard coded token counts. The bottleneck prose and proof delta also remain fixed as the text changes. This offline copy lacks the D3 chart dependency, so chart rendering raises an error after earlier text updates. Event listeners are attached before that first computation, but the graph and later matrix rendering require separate evidence. The Markdown download uses the filtered original sections, with no model generation or backend request in this local run.