Token Optimization Check

Token Optimization Check

Audits the current repo against the GitHub Copilot token-optimization ruleset and reports concrete violations with file paths, line counts, estimated token cost, and fix suggestions.

When to use

• User asks: "check token optimization", "audit copilot costs", "is this repo token efficient", "review copilot-instructions", "AGENTS.md too big", "MCP overhead check".
• After adding/editing .github/copilot-instructions.md, AGENTS.md, CLAUDE.md, .vscode/mcp.json, or .copilot/ config.
• Before publishing a template repo or onboarding new contributors.

Rules checked

Rule IDs map to the token-optimization guide. Each finding = rule ID + severity + path + evidence + fix.

R1 — Always-on context size (HIGH)

• Files: .github/copilot-instructions.md, AGENTS.md, CLAUDE.md, .cursorrules, .github/instructions/**/*.md
• Limits:
  • copilot-instructions.md: ≤ 80 lines AND ≤ 600 tokens (2400 chars)
  • AGENTS.md / CLAUDE.md: ≤ 120 lines
• Token estimate: chars / 4
• Flag if any file exceeds limits.

R2 — Duplicate always-on files (HIGH)

• If both AGENTS.md AND .github/copilot-instructions.md exist, diff them. Flag overlapping content (>30% line similarity) as "paid twice".

R3 — Verbose / non-caveman instructions (MEDIUM)

• Scan always-on files for filler patterns:
  • Articles overload: lines starting with "The ", "A ", "An " > 20% of lines
  • Pleasantries: "please", "kindly", "feel free", "I'd like you to", "could you"
  • Hedging: "maybe", "perhaps", "might want to", "you should probably", "it's important to"
  • Filler: "just", "really", "basically", "actually", "simply"
• Flag if ≥ 3 hits across these categories.

R4 — Discoverable / redundant facts (MEDIUM)

• Flag lines that restate what code reveals:
  • "This project uses TypeScript" when tsconfig.json exists
  • "We use PostgreSQL/Node/Python/etc" when manifest declares it
  • "Tests are in tests/" when tests/ exists
  • "Main branch is protected" / "Repo uses Git"
• Pattern: \b(this (project|repo) (uses|is)|we use|tests? (are|live) in)\b

R5 — Missing output-control directive (HIGH, single-line fix)

• Always-on files SHOULD contain at least one of:
  • "code only" / "no explanation" / "be concise" / "terse" / "bullets over paragraphs"
• If none found → flag. Suggested fix: append Be concise. Code only for generation. No explanations unless asked.

R6 — MCP server bloat (HIGH)

• Files: .vscode/mcp.json, .copilot/mcp-config.json, .mcp.json
• Count servers. Estimate servers × 5 tools × 200 tokens per-step cost.
• Flag if > 5 servers configured at workspace scope (global is user's concern).
• For each server, note name + flag if obviously off-topic for the repo's language/stack.

R7 — Open-tab / large-file risks (LOW)

• List source files > 1000 lines under src/, lib/, app/. These bloat auto-context when open.

R8 — Non-English always-on content (LOW)

• Detect CJK / Cyrillic / Hebrew / Arabic characters in always-on files. CJK costs ~2x English tokens.

R9 — Missing Conventional Commits / terse commit hint (LOW)

• If .github/copilot-instructions.md covers commit guidance, check for "conventional commits" or "≤50 chars" / "subject line ≤ 72".

R10 — Wenyan / classical-Chinese prompts (LOW)

• Flag files containing dense classical Chinese markers (e.g., 之, 也, 矣, 乎) used as prompt style. Recommend terse English.

How to use

1. Read the repo root listing. Identify which always-on files exist.
2. For each existing file, read fully and run rules R1–R5, R8, R9, R10.
3. Read MCP config files if present → run R6.
4. Scan workspace for files > 1000 lines (R7). Use file_search/grep, not full reads.
5. Compute an overall score:
   • 100 baseline
   • HIGH violation: −20
   • MEDIUM: −10
   • LOW: −3
   • Floor at 0.
6. Output a single Markdown report (see Output).

Output

Emit ONE markdown report. No preamble, no postamble. Structure:

# Token Optimization Audit — <repo name>

**Score:** <0–100> / 100  |  **Findings:** <H> high · <M> medium · <L> low

## Always-on context budget
| File | Lines | ~Tokens | Limit | Status |
|---|---:|---:|---|---|
| .github/copilot-instructions.md | 142 | ~980 | 80 / ~600 | ❌ Over |
| AGENTS.md | — | — | — | n/a |

## Findings

### [R1·HIGH] copilot-instructions.md exceeds size budget
- File: `.github/copilot-instructions.md` (142 lines, ~980 tokens)
- Cost: ~980 tokens × every interaction
- Fix: Compress to ≤80 lines. Drop discoverable facts. See R4 candidates below.

### [R5·HIGH] No output-control directive
- Fix: Append one line:

Be concise. Code only for generation. No explanations unless asked.

(... one block per finding ...)

## Quick wins (apply in order)
1. <highest-ROI fix first>
2. ...

## Estimated savings if all fixes applied
- Per-interaction input: −<N> tokens
- Per agent step (tool defs): −<N> tokens

Notes

• Token counts are estimates (chars / 4). Do not claim exact billing numbers.
• Do NOT modify any files. Report only. The user applies fixes.
• If a file is absent, omit its row rather than flagging absence (except R5).
• Keep the report itself terse. Caveman style allowed in finding text.

Source: sujithq/skills — distributed by TomeVault.