context-budget

Overview of context-budget skill

What the context-budget skill does

The context-budget skill audits how much Claude Code session context is being consumed by agents, skills, rules, and MCP servers. Its job is simple: find where tokens are being spent, estimate what is actually expensive, and turn that into prioritized cleanup actions instead of vague “reduce prompt size” advice.

Who should install context-budget

This skill is best for people maintaining a non-trivial Claude Code setup: multiple custom agents, a growing skills/ folder, layered rule files, or several MCP servers. If your sessions feel slower, outputs drift, or you are unsure whether adding another tool will hurt quality, context-budget is a strong fit. It is less useful for minimal setups with only a few files.

Why users choose it over a generic prompt

A generic audit prompt may tell the model to “look for bloat,” but context-budget gives a structured workflow: inventory components, estimate token load, flag common hotspots, avoid double-counting copied skills, and produce ranked savings opportunities. That makes the context-budget skill useful for Skill Testing because it reduces guesswork and creates a repeatable review path.

Key limits to know before install

This is an estimation and triage skill, not a byte-perfect tokenizer or an automatic refactor tool. It focuses on practical overhead signals such as oversized files, verbose frontmatter, overlapping rules, and high-tool-count MCP servers. If you need exact runtime token accounting, treat its numbers as directional and use them to prioritize what to inspect manually.

How to Use context-budget skill

Install context-budget in your skill workflow

The repository does not expose a separate package just for this skill. In practice, users install it from the parent repo and invoke the context-budget skill from there. A common starting point is:

npx skills add affaan-m/everything-claude-code --skill context-budget

Then confirm the skill exists under your installed skills and open skills/context-budget/SKILL.md in the source repo if you want to understand the intended audit flow before running it.

What input the context-budget skill needs

The context-budget skill works best when you point it at a real Claude Code workspace structure. Useful inputs include:

• agents/*.md
• skills/*/SKILL.md
• rules/**/*.md
• .mcp.json or active MCP configuration
• notes on which components are actually loaded in normal sessions

Do not ask only “audit my context.” Better: “Audit token overhead across agents/, skills/, rules/, and .mcp.json. Estimate heavy files, duplicated skills, overlapping rules, and MCP tool schema cost. Return top 5 savings ranked by impact vs effort.”

Turn a rough goal into a strong context-budget prompt

A weak prompt asks for a generic summary. A strong context-budget usage prompt defines scope, output format, and decision criteria. Example:

“Use the context-budget skill on this repo. Inventory all agents, skills, rules, and MCP servers. Estimate token consumption with clear assumptions, skip identical duplicated skills if mirrored under .agents/skills/, flag heavy files and redundant rule content, and recommend immediate, medium, and low-priority cuts. Include likely savings and risk of removing each item.”

That prompt improves quality because it forces the skill to separate measurement, deduplication, and prioritization.

Best workflow and what to read first

Read SKILL.md first; it contains the full method. Focus on:

1. When to Use
2. How It Works
3. Phase 1: Inventory
4. Phase 2 classification and recommendations

In use, the best workflow is:

1. Run one broad audit.
2. Validate the biggest hotspots manually.
3. Remove or consolidate one category at a time.
4. Re-run the context-budget skill after each change.

This prevents large cleanups from breaking useful behavior while still reclaiming context headroom quickly.

context-budget skill FAQ

Is context-budget worth using on a small setup?

Usually not. If you have only a few skills and no MCP sprawl, ordinary inspection is often enough. The context-budget skill becomes more valuable once your environment has enough moving parts that hidden overhead affects quality or makes future additions risky.

How is context-budget different from a normal repo review prompt?

A normal prompt often reviews content quality; context-budget reviews context cost. It is tuned for overhead sources that are easy to miss, like duplicate installed skills, overlong descriptions, overlapping rules, and MCP tool schema bulk. That narrower focus is why it can outperform a broad “optimize my setup” request.

Can beginners use the context-budget skill?

Yes, if they can identify their Claude Code config files and folders. You do not need deep tokenization knowledge. The main beginner risk is over-deleting useful guidance after seeing “large file” warnings. Use the skill to rank candidates, then inspect high-impact items before removing them.

When should I not use context-budget?

Skip it if your problem is model behavior unrelated to context load, such as weak task instructions, bad examples, or missing permissions. Also skip it when you need exact tokenizer math or automated edits; the context-budget guide is for diagnosis and prioritization, not exact accounting or one-click remediation.

How to Improve context-budget skill

Give better repo context and active-load assumptions

The best way to improve context-budget results is to tell it what is actually loaded in real sessions. Distinguish always-on rules from rarely used ones, and active MCP servers from configured but idle ones. Without that, the skill may overstate low-impact components and understate true bottlenecks.

Ask for ranked tradeoffs, not just raw counts

Raw token estimates are not enough. Ask the context-budget skill to score findings by impact, effort, and risk. For example: “Prefer recommendations that save meaningful context without removing unique capabilities.” This helps avoid bad advice like deleting a long but essential skill before trimming duplicated rule text.

Watch for common failure modes

The biggest failure modes are double-counting mirrored files, treating all tools as equally expensive, and assuming large files are automatically wasteful. Prompt the skill to separate:

• duplicated vs unique content
• static overhead vs frequently used context
• bloated frontmatter vs necessary operational instructions

That makes the output more decision-ready.

Re-run context-budget after each cleanup

Do not batch many edits and hope for the best. Use the context-budget skill iteratively: audit, change one cluster, re-audit, compare. This reveals whether the cleanup actually created usable context headroom or simply moved complexity elsewhere. For Skill Testing, iterative runs are the fastest way to confirm the skill is improving a real working setup rather than producing a one-off report.