The AI Tool Drop
Take-Home Pack
Two halves. How to pick a model and ask it for work, then how to audit what you have already installed. That second half is where most people's output is quietly going wrong.
Match the model to the job
The number is the generation. Sol, Terra and Luna are permanent tiers that carry forward.
| Model | Give it | Watch for | Price per million |
|---|---|---|---|
| GPT-5.6 Sol | Long, hard, multi-step jobs | Drifts on very long runs | $5 / $30 |
| GPT-5.6 Terra | Everyday work, most of the time | The sensible default | $2.50 / $15 |
| GPT-5.6 Luna | High volume, speed over depth | Shallower reasoning | $1 / $6 |
| Claude Opus 5 | Novel problems, agentic runs | Talks a great deal | $5 / $25 |
| Claude Fable 5 | Large codebases, deep analysis | Twice the price of Opus 5 | $10 / $50 |
| Claude Sonnet 5 | Fast everyday work | Goes up 50% on Aug 31 | $2 / $10 |
Prices are per million tokens, input then output, as published by each vendor on July 27, 2026.
How to actually decide, in four questions
1. One task, or a chain of them? One task, take the cheap fast tier. A chain that has to hold together, take the top tier.
2. Right, or fast? Both cost money. Only one costs you a customer.
3. Will it run while I am not watching? If yes, pick the model that narrates and self-corrects. Unattended is exactly when you want that.
4. Am I thinking, or producing? Thinking is a chat. Producing is Cowork, or ChatGPT Work.
Stop asking questions. Give it a job.
A question gets you an answer you still have to go and do something with. A job has a deliverable, a standard, and a way to check it.
Question-shaped, weak
- "How do I write a follow-up email?"
- "What should I post on LinkedIn?"
- "Can you analyze my sales data?"
- "Help me with my pricing."
- "Write a job description."
Job-shaped, strong
- "Here are the last five follow-ups that got replies. Write the sixth for this client, same voice and length. Then tell me what the five have in common so I can stop asking you."
- "Read these three posts of mine that performed and the two that didn't. Draft one about this week's job, match the winners, then name the pattern."
- "Here is the spreadsheet. Calculate month over month by product line, flag anything moving more than 15 percent, and show me the formula so I can check it."
- "Here are my rates, my last twelve proposals, and which closed. Find what the closed ones share at each price point. Do not recommend anything until you have shown me the pattern."
- "Here are two descriptions that brought good candidates. Write one for this role in the same register, then list what you copied and what you invented so I can check the invented parts."
The tell of a good job prompt
It names the deliverable, gives real examples to work from, and asks the model to show its work so you can judge it. Every one of those five ends with a check.
And the rule underneath: the prompt gets shorter as the context gets better. If you find yourself writing longer and longer instructions, you do not have a prompting problem. You have a context problem, which is Part Two.
Audit what you already installed
For about four months my drafts came back sounding like somebody else. Hashtag blocks, manufactured urgency, a voice that was not mine. I rewrote them every time and never found the cause.
The cause was a file I had never opened, inside a plugin I had forgotten I installed, telling it exactly how to write. It was not broken. It was following instructions perfectly, and it never once mentioned it.
When I finally audited: 21 plugins, 7 duplicated across two marketplaces, and zero carrying a workflow I actually depend on. Three stale copies of my own voice document were living inside them, none aware the real one existed.
Problem one: bloat
- Too much installed, wasting context and slowing things down.
- Annoying, visible, and easy to fix.
- This is what most cleanup advice covers.
Problem two: conflict
- Two instructions that disagree, where the model silently picks one.
- Invisible, and it is the one that changes your output.
- These five prompts go after this one.
The honest inventory
You cannot audit what you cannot see. The instruction that cost me four months never appeared in any menu.
List everything currently loaded into this conversation that could influence your output. For each item give me: its name, where it lives, what it instructs you to do, and whether it loaded automatically or because I invoked it.
Then separate them into two groups. Capability: things that connect you to a tool or a source of data, holding no opinion about how I work. Methodology: things that tell you how to do my job, what to write, what tone to take, or what format to use.
Do not summarize. I want the actual list, including anything that does not appear in a menu.
The conflict hunt
The differentiated one. Most cleanup tooling measures bloat. Almost none looks for semantic conflict, and semantic conflict is what silently changes your work.
Compare every instruction currently loaded and find the places where two of them disagree.
For each conflict give me: both sources by name, the exact instruction from each, what you actually do when they collide, and which one currently wins.
Include soft conflicts, not just direct contradictions. Two sources that set a different tone, a different format, or a different level of formality are in conflict even if neither mentions the other.
If nothing conflicts, say so plainly rather than inventing something.
The drift test
Run this the moment something sounds off. The before and after is what makes the cause undeniable, and it is the fastest diagnosis in the pack.
Here is something you wrote for me that did not sound like me: [paste it]
Here are two or three things I wrote myself: [paste them]
Work backwards. What instruction currently loaded would have produced your version rather than mine? Name the source if you can find it. If more than one is contributing, rank them by how much each one moved the output.
Then show me the same piece rewritten with that instruction ignored, so I can see the difference.
The load-bearing test
Duplicates are the hard part to catch by eye, because a standalone tool and a copy bundled inside a plugin live in different places and never sit next to each other.
Go through everything installed and tell me, honestly, which of these are actually carrying work I depend on.
For each one: when did it last do something for me, what would break tomorrow if I removed it, and is there another copy of the same capability somewhere else in my setup.
Sort them into keep, duplicate, and never used. Be blunt. I would rather delete something useful and reinstall it than keep twenty things I cannot account for.
Write the one page that replaces most of it
This is the fix. It took me an afternoon and it repaired what four months of rewriting could not.
Interview me until you can write one page of context for my business. Ask one question at a time and keep going until you have enough. Cover what I sell, who I sell it to, how I talk, what I never say, and what a good piece of work looks like when I make it.
When you have it, write it as a single page in plain language, addressed to an AI that has never met me.
Then tell me which of my currently installed methodology tools that page makes unnecessary.
Save that page where the model reads it every time: a Claude Project, a Cowork skill, a Gemini Gem, or a ChatGPT project.
The staged answer on what to install
Capability tools: always. They connect you to something otherwise out of reach and hold no opinion about your work. They age well.
Methodology tools: as training wheels, for about a month. A generic framework genuinely beats a blank page when you are starting.
Then write your own context and expect to shed them. That last step is the one nobody teaches, and it is the whole point.
Two things other people built
Neither is ours and both are free.
Stack Cleaner
- Open source, runs locally, sends nothing anywhere.
- One command inventories what you have installed and counts what you actually use by reading your own transcripts.
- Flags the duplicates that plugins quietly bundle, then exports a cleanup plan you review and run yourself.
- stackcleaner.com is the best tool for the duplicate problem specifically.
The Ultimate Claude Code Guide audit prompt
- A serious eight-dimension, hundred-point audit by Florian Bruniaux.
- Covers memory, rules, skills, security posture and connected servers.
- Written for engineers auditing a codebase, so expect shell commands and terminology.
- Start here if you are technical.
Why the prompts above are ours and not theirs
Both of those are excellent at measuring bloat, and both assume a developer at a terminal. Neither looks for the failure that actually cost me four months, which was two instructions disagreeing about voice and the model silently picking the wrong one.
That is a different problem, it is invisible, and it is the one most likely to be quietly costing you.
Come sit with us
The AI Exchange meets every Tuesday at Workuity 24th & Camelback and every Wednesday at Workuity Chandler, noon, lunch on us.
Bring a laptop and a workflow you want to speed up.