5 Prompt Tips to Be a Better Meat Proxy

5 prompt tips to be a better meat proxy

A meat proxy copies and forwards AI-generated text, code, or other output without reading, understanding, or validating it. A human-shaped clipboard. The model does the work and the meat proxy (you) forwards it unread. We all know the type. Some mornings, that's us.

If you want to be the best meat proxy in the building, here is how. Please don't. If you want to get more done for less, do the opposite.

• Use the biggest model for everything. Ask the frontier model to rename a variable. Your invoice is the only part of the pipeline anyone will read carefully.
• Run the same session forever. Context is a lifestyle. By hour nine the model is 400,000 tokens deep in a conversation about a bug you fixed on Tuesday, and it is still very sure about Tuesday.
• Review and implement in the same breath. "Write it, then check it" is a great way to have the model grade its own homework. It always gets an A.
• Skip the handoff note. The next agent will figure it out. So will the next human. Nobody will, but the vibes are immaculate.
• Never read the output. Copy, paste, send. If it is wrong, it was the model's fault, and the model has no manager to complain to. The proxy's only job is to be a very expensive cable.

A meat proxy is a pipe with a face. Captain Cache, our hero, reads the diff, runs the tests, and sometimes does the small thing by hand. Grimly competent behavior.

Most of our teams use AI coding tools, but every engineer works with them a little differently. Captain Cache ships more product work and merges more PRs than anyone else on the team. Their AI tools cost about $500 a month, below the team average. Some engineers spend more than $5,000 a month and finish fewer tasks of similar size.

Captain Cache follows two rules. They pick the right model for the job and give each session a clear focus. Hard problems get the most capable models. Routine work gets something faster and cheaper. Nobody summons the ancient oracle to rename a variable.

The invoice leaderboard

Now Captain Cache becomes financially annoying. The table shows the ten highest spenders from June through September, with our hero added for comparison. We removed everyone's names. Captain Cache is our nickname for an engineer on the team.

The highest spender spent 8.7 times as much as Captain Cache. Sadly, invoices do not award frequent-flyer miles. We included engineers who used the tools regularly and left out service accounts.

The biggest spenders did not ship the most

We measure completed product work through Jira delivery, not lines of code. To compare engineers fairly, we turned that data into an internal productivity index from 0 to 100. A higher score means more completed work.

Captain Cache scored 96, the highest output on the team, while ranking 27th in spend. Several engineers spent far more and shipped less. The leaderboard had betrayed them.

The $500 to $2,500 brackets had the highest average output. Engineers below $500 and above $5,000 both averaged about 42 points. More tokens did not mean more completed work. Some of the largest bills came from agents retrying tasks after long sessions had filled up with old context. The meter kept running; the work did not.

Engineers who already shipped steadily tended to complete more work with AI. Engineers who had trouble finishing tasks often kept struggling. Looking at Jira delivery next to agent activity helps us separate useful work from expensive motion.

A few high spenders appeared to spend more time running agents than shipping. We never rewarded spend, but online "fleet of agents" demos can make constant activity look productive. The agents were busy. Jira was less impressed.

Do not send a tank to rename a variable

We call our most capable models frontier models. Opus, Fable, and Astra can handle our hardest problems, including difficult math. A CRUD endpoint is rarely difficult math. For routine work, we usually start with a faster, less expensive model:

• GLM-5.3
• DeepSeek-v4.1-flash
• GPT-6-Luna
• Sonnet-5.5

Give these models a clear task or bug reproduction and they can handle most CRUD endpoints, refactors, tests, and bug fixes. Starting with the most expensive model often raises the bill without improving the result.

When the cheaper model gets stuck, Captain Cache switches to a frontier model, keeps the session short, and switches back. No loyalty oath required.

Captain Cache's operating manual

• Give each session a single job. Captain Cache may work across several repositories, but closes a session as soon as its task is done.
• Separate review from implementation. Combining them makes the goal less clear and asks the model to hold the red pen over its own homework.
• Start fresh after the prompt cache expires. The cache lasts roughly 30 minutes with Codex and 60 minutes with Claude. After that, the next turn has to process the full conversation again. Old context is now full-price context.
• Leave a short handoff before a break. Record the current state and the next step, then continue in a new session from that note. Future you has suffered enough.
• Save a checkpoint before the session fills up. Captain Cache leaves automatic compaction on, then writes a handoff or makes a checkpoint commit as the session nears its limit. The next agent can continue without doing conversational archaeology.
• Work without AI when it is faster. Captain Cache reviews PRs by hand, checks metrics and backlogs, and handles small changes alone. That keeps their coding and review skills sharp.

Long sessions send more history with every turn. Short sessions cost less, use the prompt cache better, and give the model less old context to sort through.

The boring system that works

For routine work, start with a capable, inexpensive model. Give the session a single job and close it when the job is done. Leave a short handoff if someone needs to continue. Bring in a frontier model when the cheaper model gets stuck or the problem demands deeper reasoning. Move up when you need to; you rarely need to start there.