Load Cutter
Skill Installation Guide — Step-by-step for first-time users
Load Cutter
Skill Installation Guide — Step-by-step for first-time users
Load Cutter is a skill for Claude's Cowork mode that prevents context overload on complex, long-running tasks. Without it, Claude accumulates conversation history and file content in its active memory until it hits a limit — slowing down, degrading quality, or stopping entirely.
Once installed, it automatically breaks large tasks into focused phases, saves reusable context to knowledge files instead of keeping it in the chat window, and routes each part of the work to the right model tier. The result: the same output quality, at a fraction of the token cost.
Download load-cutter.skill
The installer package is a single file. Download it, save it somewhere easy to find, then follow the five steps below to add Load Cutter to your Claude Cowork setup.
| You need | Details |
|---|---|
| Claude desktop app | Mac or Windows. Download at claude.ai/download if you don't have it yet. |
| Cowork mode | Open the app and confirm you see a "Cowork" option in the left sidebar. It is currently a research preview. |
| The .skill file | You should have received load-cutter.skill — keep it somewhere easy to find (Desktop or Downloads). |
Find the file you received: load-cutter.skill
If it came via a shared folder or download link, save it to your Desktop or Downloads — somewhere you can reach it quickly from a file picker in the next steps.
- Launch the Claude desktop app.
- In the left sidebar, click Cowork. If you don't see it, make sure you have the latest app version installed.
- Start or open a Cowork session. You should see the main chat area.
Skills are managed through the plugin panel inside Cowork mode:
- Look for a Plugins or Skills button — typically in the top-right corner of the Cowork interface, or accessible via the ⋯ (more options) menu.
- Click it to open the plugin manager panel.
- You'll see a list of any currently installed skills and an option to add a new one.
- In the plugin manager, click "Install from file" or "Add skill" (the exact label may vary by app version).
- A file picker will open. Navigate to where you saved
load-cutter.skill. - Select the file and click Open or Confirm.
- Wait a few seconds. You should see load-cutter appear in your list of installed skills.
Start a fresh Cowork session and try one of the phrases below. Claude should respond by analysing the task and producing a structured phase plan — not just diving straight into work.
Claude should break the task into phases and identify what context to extract.
Claude should produce a phase plan and suggest knowledge files to create.
Claude should offer to compact and restructure the remaining work.
How it works
Phase breakdown
Splits complex tasks into clear stages. Each phase loads only what it actually needs — not the entire conversation history.
Knowledge files
Saves reusable context (schemas, decisions, project rules) to .md files and loads them on demand — not permanently in the chat window.
Model routing
Uses the strongest model for planning and a leaner model for execution. You stop paying for heavy reasoning on routine tasks.
Quick trigger reference
These phrases reliably activate Load Cutter in any Cowork session. It also triggers automatically on complex multi-step or context-heavy tasks.
Troubleshooting
| Problem | Fix |
|---|---|
| Skill doesn't appear after install | Restart the Claude desktop app completely, then check the plugin manager again. |
| File picker won't accept the file | Confirm the filename ends in .skill and hasn't been renamed to .zip or .txt by your browser or OS. |
| Claude doesn't respond to trigger phrases | Open a brand new Cowork session (not a continuation of an old one) and try again. |
| Plugin manager isn't visible | Update the Claude desktop app to the latest version at claude.ai/download. |
| Error during install | Re-download the .skill file and retry. If the error persists, contact whoever shared the file with you. |
Token savings analysis
Modelled scenario: Developing a 15-page AI governance framework for a regulated financial services client — 30 exchanges on Claude Sonnet 4.5 ($3/M input · $15/M output). Initial context load: 6,000 tokens (client brief, AI systems inventory, regulatory scope). Average tokens added per exchange: 2,800.
Every message in an unmanaged Claude session re-sends the full conversation history. Context at exchange n = initial brief + all prior exchanges. This makes token consumption grow quadratically — not linearly — with session length. Load Cutter breaks that compounding by resetting context at phase boundaries.
| Metric | Without Load Cutter | With Load Cutter | Delta |
|---|---|---|---|
| Session structure | 30 linear exchanges | 6 phases × 5 exchanges | — |
| Context carried per exchange | Full accumulation | Compact + knowledge file | ~4,400 → ~2,600 tok |
| Input tokens | 1,413,000 | 278,000 | −80% |
| Output tokens | 69,000 | 78,300 * | +13% |
| Total tokens | 1,482,000 | 356,300 | −76% |
| Estimated cost | $5.28 | $2.00 | −62% |
* Output tokens are slightly higher with Load Cutter due to knowledge file writes and compact summaries between phases — a small, necessary overhead. The cost reduction (62%) is lower than the token reduction (76%) because output tokens cost 5× more than input tokens on Sonnet, and the output volume stays roughly constant regardless of approach.
Savings scale with task complexity. The break-even point is roughly 8–10 exchanges — below that, the structure isn't worth the overhead. At 50 exchanges, the token reduction climbs to approximately 82%, because the quadratic penalty of unmanaged context growth compounds harder at scale.