usage-limit-reducer

A Claude Code skill that stops you from burning through your usage limit.

Claude re-reads the entire conversation on every turn. One developer measured it: 98.5% of his tokens went to re-reading history, only 1.5% to generating responses. This skill diagnoses where your tokens are actually going and applies Dubi’s 11 rules for cutting usage — using real data from the JSONL logs Claude Code already writes to your disk.

What it does

When you trigger the skill, it:

1. Measures — runs scripts/usage-report.py against ~/.claude/projects/*/*.jsonl and shows token/cost breakdowns by model, project, and day (with cache-hit % and heuristic callouts).
2. Diagnoses the current session — conversation length, CLAUDE.md presence, model in use.
3. Applies the 2–4 rules most relevant to what it finds. No dumping all 11 on you.
4. Suggests concrete next actions (/compact, /clear, /model claude-haiku-4-5, create CLAUDE.md, etc.) and acts only on confirmation.

Install

git clone  ~/.claude/skills/usage-limit-reducer

Or, if you already cloned it elsewhere:

cp -r . ~/.claude/skills/usage-limit-reducer/

Restart Claude Code. It will show up in the skills list and auto-trigger on phrases like:

• “hit my limit” / “usage limit” / “running out of tokens”
• “save tokens” / “reduce usage” / “am I wasting tokens”
• “this chat is getting long”
• “which model should I use”

You can also invoke it directly: /usage-limit-reducer.

Standalone: the token report

The script is useful on its own, outside the skill:

python3 ~/.claude/skills/usage-limit-reducer/scripts/usage-report.py --days 7
python3 ~/.claude/skills/usage-limit-reducer/scripts/usage-report.py --days 30 --project myrepo
python3 ~/.claude/skills/usage-limit-reducer/scripts/usage-report.py --json   # machine-readable

Example output:

Claude Code usage — last 7 day(s)
------------------------------------------------------------------------
Turns:        110
Input:        274 fresh  +  6.16M cache-read  +  934.3K cache-write
Output:       152.9K
Cache hit:    86.8% of input tokens came from cache
Estimated $:  ~$38.22 (API list prices; subscription is flat)

By model (top by cost):
  claude-opus-4-7     turns= 109  in=   274  out= 152.9K  cache_r= 6.16M (86.8%)  ~$38.22

Heuristics:
  - Cache hit is 86.8% — healthy.
  - Opus is 100% of spend. Consider Haiku (/model claude-haiku-4-5) for simple edits.

What’s read: every JSONL file under ~/.claude/projects/. What leaves your machine: nothing. No network calls.

Pricing note: the $ figure uses API list prices as a relative-comparison tool. If you’re on a Claude Code subscription plan, billing is flat — treat the number as “what this would cost on the API.”

The 11 rules, and what each maps to

Not implementable in Claude Code (flagged honestly by the skill):

• Rule #1’s “edit the original prompt” is a claude.ai web UI feature.
• Rule #5’s “upload to Projects” is also web-only.
• Rule #11’s Overage toggle is plan-level.

Layout

usage-limit-reducer/
├── SKILL.md                  # what Claude reads when the skill triggers
├── README.md                 # this file (for humans)
└── scripts/
    ├── usage-report.py       # the local token dashboard
    └── test_usage_report.py  # 16 unit + CLI tests

Tests

python3 scripts/test_usage_report.py

Covers price lookups, aggregation math, --days cutoff, --project filter, malformed-JSONL handling, CLI text/JSON output, and a regression test for the defaultdict(lambda: dict(totals)) shared-state bug.

Credit

The 11 rules are from Dubi’s article. This skill just wires them into Claude Code with real measurement.