# Limit Saver: make your Claude and ChatGPT plan last longer

You pay for a plan, and it runs out faster than it should. Most of that is not you. It is
default behavior: long answers, whole documents rewritten to change one line, giant chats
that resend everything on every message, and tools firing when you did not need them.

This kit turns those defaults off. Install once, about 2 minutes.

Honest note: how much this saves depends on how you use your AI. It points usage in the
right direction. It is not a guaranteed number.

## What actually eats your limit

Anthropic lists these as what counts against your Claude usage: message length, file
attachment size, current conversation length, tool usage (research, web search), model
choice, effort level, artifacts, and multi-step tasks like running code or creating files.
ChatGPT works the same way in practice.

The big one most people miss: **every message resends the whole chat.** Message 40 in a long
thread costs far more than message 1 in a fresh one, even if you type the same thing.

## What is in the kit

| File | What it does | Where it goes |
|---|---|---|
| `always-on-instructions.txt` | 7 rules that make every answer cheaper | Claude + ChatGPT settings, pasted once |
| `limit-saver-skill.zip` | Claude skill: handoff, limit check, lean mode | Uploaded to Claude once |
| `chatgpt-handoff-prompt.txt` | The handoff for ChatGPT | Saved prompt, paste when a chat gets long |
| `claude-code/CLAUDE.md` | Lean rules file for Claude Code | Your project root or `~/.claude/` |
| `claude-code/handoff/SKILL.md` | `/handoff` skill for Claude Code | `~/.claude/skills/handoff/` |

## Step 1: paste the always-on block (Claude and ChatGPT)

This is the part that works on every single answer.

**Claude:** click your initials (bottom left), open **Settings**, find **Instructions for
Claude**, paste the block, save.

**ChatGPT:** open **Settings**, then **Personalization**, then **Custom instructions**, and
paste it into the box about how you want ChatGPT to respond. (Menu names move around. If you
cannot find it, search settings for "custom instructions".)

What the 7 rules do:

1. Answer first, no intro or recap. Shorter answers cost less and read faster.
2. Match length to the question.
3. Show only what changed instead of rewriting the whole thing. This is the biggest single
   saver if you edit documents or code.
4. Ask up to 2 questions before starting when the request is unclear. One good answer beats
   three do-overs.
5. No web search, code running, or file creation unless you ask or it is needed.
6. One recommendation, not five options.
7. Warn you when the chat is getting long so you can hand off.

## Step 2: install the skill (Claude)

1. In Claude, open **Settings**, then **Capabilities**, and turn on code execution. Skills
   need it.
2. Go to **Customize**, then **Skills**, click **+**, choose **Create skill**, then
   **Upload a skill**.
3. Upload `limit-saver-skill.zip`. Make sure it is toggled on.

Skill uploads work on Free, Pro, Max, Team and Enterprise plans. On Team and Enterprise an
owner has to enable skills for the organization first.

Now you have three commands, in plain words:

- **"handoff"** writes a short brief of this chat (goal, decisions, key details, next step).
  Copy it, open a new chat, paste it. You keep the thread and drop the weight.
- **"limit check"** tells you what is eating your limit in this chat, biggest drain first,
  with a one-line fix for each.
- **"lean mode"** makes every answer in this chat short and cheap until you say
  "lean mode off".

**ChatGPT** does not take skill uploads the same way, so use `chatgpt-handoff-prompt.txt`.
Save it somewhere you can reach fast (a note, a text shortcut), and paste it when a chat
gets long.

## Step 3: match the model to the job

Using the biggest, deepest-thinking model for everything is the fastest way to drain a plan.
Thinking is billed like output, and it can be a lot.

- Quick questions, rewrites, summaries, formatting: use the smaller, faster model and keep
  thinking off or low.
- Hard problems, planning, tricky code, anything where a wrong answer costs you an hour: use
  the big model with thinking on. That is what you are paying for.

In Claude you change the model, effort and thinking from the model picker in the chat box.

## Step 4: three habits that cost nothing

- **Put reused files in a Project.** Anthropic says content in projects is cached and does
  not count against your limits when reused. Upload the brand guide, the codebase notes, the
  long PDF once. Stop re-attaching it to every chat.
- **Send one full message, not five small ones.** Group your questions. Read your prompt
  once before sending and add what is missing. Every follow-up resends the whole chat.
- **Watch the meter.** In Claude, **Settings**, then **Usage** shows your session and weekly
  limits live, so you can see what a heavy chat actually cost.

## Step 5: Claude Code (for builders)

Claude Code burns the most, because every turn carries files, tool calls and reasoning.

1. Drop `claude-code/CLAUDE.md` into your project root (or merge the rules into yours).
   Keep your CLAUDE.md under about 200 lines. It loads into every session.
2. Copy `claude-code/handoff/` into `~/.claude/skills/` so you have
   `~/.claude/skills/handoff/SKILL.md`.
3. Use these:

| Command | When |
|---|---|
| `/usage` | See what is counting against your plan, including which skills, plugins and MCP servers use the most |
| `/clear` | Every time you switch to an unrelated task. Stale context costs you on every message |
| `/handoff` then `/clear` | When you want a clean session but need to keep the thread |
| `/compact` | At a natural checkpoint in a long task. Note that compacting a huge session is itself a big request |
| `/model` | Sonnet handles most coding. Save Opus for architecture and hard multi-step problems |
| `/effort` | Lower it for simple tasks |
| `/mcp` | Turn off servers you are not using |
| `/context` | See what is filling the window |
| `/rewind` | Claude went the wrong way? Go back instead of spending ten messages correcting it |

Also: for anything big, use plan mode (Shift+Tab) first. Approving a plan is much cheaper
than redoing the work.

## Does this make the answers worse?

Shorter is not dumber. The rules cut padding, repeats and do-overs, not thinking. For hard
problems, say "go deep" or "full version" and you get everything. Rule 4 (ask before
guessing) usually makes answers better, because the model stops building on a wrong guess.

## Sources

- Anthropic, Usage limit best practices: https://support.claude.com/en/articles/9797557-usage-limit-best-practices
- Anthropic, Using skills in Claude: https://support.claude.com/en/articles/12512180-using-skills-in-claude
- Anthropic, Understanding Claude's personalization features: https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-features
- Claude Code docs, Manage costs effectively: https://code.claude.com/docs/en/costs

Made by Artem Novitckii. Free to use and tweak. More at https://novitckii.com/resources/