The First Prompt To Run On Every New AI Model (Claude Fable 5.2)

This tip comes from Alex Albert, who works in research at Anthropic. The day Fable 5.2 launched, he posted that every time they have a new model, he runs the same simple prompt on all his projects. Here it is, word for word.

Alex’s Original Prompt

Read this entire project and find everything that's broken, inconsistent, inefficient, or no longer needed. Fix what you're sure about and list the rest for me to review.

I run my own version so I can see what the new model catches that the last one missed. Anthropic’s Fable 5.2 prompting guide agrees with the timing. It says rules written for older models can raise your cost or stop doing anything, so reread them when you switch.

Only want to check your instruction files? Use my AI instructions checkup. This page audits the whole project.

Switch to the new model

Open Claude Code in your project folder and type /model fable. No plan uses Fable by default, so you have to pick it. (Model docs)

On Max and Ultra, Fable can use up to half your weekly limit at no extra cost. On Pro it runs on usage credits, and the picker says "Requires usage credits" when that applies. (Fable on your plan)

Fable 5.2 starts at medium effort in Claude Code. For an audit I type /effort high, since Anthropic says that’s where most of the gain shows up.

Run the audit

Commit your work first, then paste this. The branch means you can throw the whole audit away if you don’t like it.

Whole Project Audit

Read this entire project and find everything that's broken, inconsistent, inefficient, or no longer needed. Before you change anything, make a new git branch called model-audit so I can undo the whole thing in one step. Fix only what you're sure about, keep each fix small, and check that it works. List everything else for me to review, riskiest first, with the file and what you'd do about it. Last, tell me which of these the model I used before you probably missed, and why.

I left out Anthropic’s "you are the owner of this codebase" line on purpose. Their guide says it makes Fable clean up more, and it also made changes nobody asked for and used 70 to 80 percent more tokens.

Approve by number

Anthropic also says Fable 5.2’s summaries can leave out small changes, so ask Claude to show you the full diff before you merge the branch. Then reply with the review items you want done, like "do 2, 4 and 7."

No Claude Code? A Project works too, since it’s just instructions plus files. The AI can’t edit your files there, so this version hands you fixed copies to paste back in.

Open the Project. Click Projects in the sidebar of Claude or ChatGPT.
Start a new chat on the newest model. Pick it from the model menu before you send anything.
Paste this and send.

Project Audit

Read every file in this project and the project instructions. Find everything that's broken, out of date, contradicting something else, or no longer needed. For anything you're sure about, give me the corrected version ready to paste back in. List the rest for me to review, most important first, with the file name and why. Last, tell me what you caught that an older model probably missed.

To swap in fixed instructions, click the instructions box in a Claude Project. In ChatGPT, open the three-dot menu at the top of the project, then project settings.

↓ Download the skill (.zip)

This is the long version. It checks for unsaved work, makes the branch, reads everything, and keeps a model-audit-log.md file. That log lets the next model compare its findings with this one’s, so "what did it catch that the last one missed" gets a real answer.

Claude Code. Unzip it into .claude/skills/ in your project, or ~/.claude/skills/ for every project. Then type /new-model-audit. (Skills docs)
Claude app. Go to Customize (or Plugins), then Skills, the + button, Create skill, Upload a skill, and pick the ZIP without unzipping it.

new-model-audit / SKILL.md

--- name: new-model-audit description: Run when a new AI model drops. Reads the whole project, fixes only what it's sure about, lists the rest for review, and logs what this model caught that the last one missed. --- # New Model Audit Use this the first time I work in a project after switching to a new model, or when I say "run the new model audit." ## Before you start 1. Tell me in one line which model you are. 2. If this is a git project, check for unsaved or uncommitted work. If there is any, stop and ask me to commit it first. Then make a new branch called model-audit-[today's date], so I can undo every change in one step. 3. If you can't edit files here (for example, a Claude or ChatGPT Project), don't try. Hand me corrected versions to paste back in instead. 4. Look for a file called model-audit-log.md in the top folder of the project. If it's there, read it so you know what earlier models found, fixed and left for me. ## Read everything Read the whole project. That means code, docs, instruction files (CLAUDE.md, AGENTS.md, skills, subagents), settings, notes and memory files. Don't skim or sample. If the project is too big for one pass, split it into areas, go area by area (or give each area to a subagent if you can), and put the findings together at the end. ## What to look for - **Broken.** Anything that errors or fails a test, plus any file, link, path or command it points to that no longer exists. Check before you call something missing. - **Inconsistent.** Two files that disagree, or the same rule written several different ways. Names that don't match count too. - **Inefficient.** Duplicated code and steps that run twice. Anything slow or wasteful. - **No longer needed.** Dead code and old files. Also workarounds or cautions written for an older model that you don't need. ## Fix or list Fix something yourself only when you're sure it's wrong, the fix is small, it doesn't change what my users or customers see, and you can check that it worked by running it, testing it or rereading it. Everything else goes on the review list. Always ask before you delete a file or touch payments, pricing, logins, or anything that sends messages to people. ## Report back Keep it short and in plain words. 1. **Fixed.** List each fix with its file and how you checked it. 2. **For my review.** Number them, riskiest first. Give the file, what's wrong and what you'd do. 3. **What this model caught.** Name the findings you think the previous model would have missed, and why. Use model-audit-log.md to compare if it exists. If it doesn't, give your best guess and say it's a guess. Then add an entry to model-audit-log.md with today's date, your model name, what you fixed, and what's waiting on me. Create the file if it isn't there. Wait for me to approve review items by number before you do any of them.

Fable 5.2 came out November 9, and in Anthropic’s tests it beat Opus 5.5 on almost everything, like 69.9% vs 62.3% on the FrontierSWE coding test. It also costs exactly twice as much on the API: $8 per million tokens in and $40 out, against $4 and $20 for Opus 5.5.

It’s the third Fable release this year, after Fable 5 in June and 5.1 in September. Anthropic’s own models page still says to start with Opus 5.5 for most work and save Fable for the hardest, longest jobs.

Their prompting guide says early testers did well using Fable 5.2 as the planner for other models, and warns that helpers can be a big part of your bill. So I let Fable plan and check, and hand the heavy reading to Opus 5.5.

Make an Opus helper

Create a file at .claude/agents/opus-worker.md with this in it, or ask Claude to make it for you. model: opus runs it on the latest Opus, which is Opus 5.5 today. Write claude-opus-5-5 instead to lock in that exact model. (Subagent docs)

opus-worker.md

--- name: opus-worker description: Reads one area of the project and reports back. Use when the main session splits a big audit, review or research job into areas. tools: Read, Grep, Glob model: opus --- You handle one area of a bigger job. Read every file in your area and do exactly what your brief asks. Don't edit anything. Report back what you found, with the file and line for each item, in plain words, sorted with the most serious first.

Start on Fable and hand off

Type /model fable, then send this.

Fable Plans, Opus Reads

Run a full audit of this project for anything broken, inconsistent, inefficient, or no longer needed, but split up the work. You plan it, then send each area of the project to its own opus-worker subagent with a clear brief on which files to read and what to look for. Check what they send back yourself before you fix anything. Fix only what you're sure about on a new git branch called model-audit, and list the rest for me to review.

Check who’s doing the work

Type /tasks while it runs to see which model each helper is on. Want every helper on Opus without making files? Set CLAUDE_CODE_SUBAGENT_MODEL to claude-opus-5-5 in the env section of your settings. It covers any helper that doesn’t name its own model.

The First Prompt ForEvery New Model

The First Prompt For
Every New Model