Cleanup jobs worth spending Fable 5.2 on
Private deliverable. Enter the hub password to continue.
That's not right. Try again.
Cleanup jobs worth spending Fable 5.2 on
Fable 5.2 is the first Claude model that will tidy a codebase instead of politely working around the mess, and one sentence is what flips it.
Fable 5.2 landed on November 9. The interesting thing for you isn't the benchmark scores, it's a finding buried in Anthropic's own prompting guide: by default, Fable treats your repo like a house it's visiting. It bolts new functions onto what's there, walks past duplicated code, and mentions nearby bugs instead of fixing them.
Tell it the code belongs to it, and that changes. In Anthropic's testing, one sentence took duplicate-code consolidation from under 10 percent of runs to about 75 percent. An independent tester on YouTube ran the same experiment 16 times on his own repo and got zero merges out of eight runs without the sentence and eight out of eight with it.
That makes Fable 5.2 genuinely good at a category of work you've been carrying for months across 57 project folders: duplicated scaffolding, dead code nobody dares delete, stale rules in your memory files, leftover branches, and keys sitting in old folders.
Here's what to paste, what it costs, and where to keep your hands on the wheel.
Two blocks to set up first
1. The ownership sentence
You are the owner and long-term maintainer of this codebase, and its design is up to you.
Put it in the project's CLAUDE.md, or in the prompt itself for a one-off job.
What comes with it, straight from the guide: output tokens go up 70 to 80 percent, and the model makes changes nobody asked for, mostly new tests and docs. In some runs it changed how an existing feature behaves. The closing summary often leaves small changes out, so read the diff, not the summary.
Leave the sentence out when you want the smallest safe diff, like a hotfix on a live site. And don't try to have it both ways: a "no drive-by refactors" rule doesn't cancel it. Anthropic saw Fable consolidate anyway in more than half of runs, and the YouTube tester saw the opposite, with the rule beating the sentence six times out of six. Either way, stacking the two gives you an unpredictable result rather than a careful one.
2. The finishing block
Fable 5.2 has a habit of stopping partway through a long job and handing you a to-do list. Anthropic's fix, abridged from their guide:
# Finishing work
Please avoid stopping while you still owe work the user asked for. Be persistent and do not stop unless needed.
Status notes are welcome, and so is clarifying open decisions, but carry on with work that does not depend on my answer.
If you notice yourself inviting me to redirect you or offering to wait, proceed to the next part of the task instead. Ordinary obstacles like errors, timeouts, locked files, empty results and failing tools are not blockers: diagnose them, then work through them with the access you have.
If something is a real hard blocker, say plainly what you found and look for another safe way to finish.
In their replay tests this halved the number of sessions that ended with work left on the table, at about 30 percent more output tokens. It also makes the model ask permission less often, which is exactly why every prompt below keeps its own confirmation step on anything that deletes.
The eight cleanup jobs
1. The cross-repo duplication sweep
You already know this problem exists. open_md.py had six copies in three
versions, and gq.py exists because 57 repos share the same Astro and Worker
layout. That's the single best-fit job for this model.
Effort: high. Ownership sentence: yes.
You are the owner and long-term maintainer of these projects, and their design is up to you.
Read the merged knowledge graph at ~/.graphify/globalws/graphify-out/graph.json, using py -3.13 scripts/gq.py from MyExecAssistant so every hit is tagged with its repo.
Find the code that genuinely repeats across projects: same function, same helper, same Worker handler, same build step, copied rather than imported. For each cluster, tell me:
- which repos hold it, and which copy is the newest and most correct
- whether the copies have drifted, and what the differences actually do
- whether it belongs in one shared place, and where that place should be
Then write me one consolidation plan, ordered by how much breakage each move risks. Do not change any file yet. Flag any cluster where the copies have drifted so far that merging them would change behavior on a live site.
Hand the approved plan to Opus or DeepSeek to execute, one repo at a time, then have Fable review the result. That split is what the people getting the most out of this model are all doing: Fable decides, cheaper models type.
2. The dead code question nobody asks
One team paid for Fable 5, Opus 4.8 and GPT 5.5 to audit the same backend. Two models wrote careful refactoring plans for a 900 line subsystem. Fable asked who calls this, found the answer was nobody, and told them to delete it. About 1,800 lines of provably dead code, plus the dependencies that only existed to serve it.
Effort: high. Ownership sentence: no, this one stays read-only.
Audit this repo for code that is provably dead. For every candidate, answer the question "who calls this?" with file paths and line numbers, and distinguish:
- defined and never called anywhere, including tests, scripts and CI
- called only by other dead code
- reachable only behind a flag or path that was abandoned
- looks important but was replaced by a newer execution path
Separate what you verified from what you suspect. Do not delete anything and do not refactor anything. Give me a deletion list ordered from safest to riskiest, with the test or command that would prove each one is safe to remove.
3. The memory and rules audit
This is the one I'd run first. Anthropic's guide has a section called "Keep caution from piling up," and it describes your setup exactly: cautious instructions accumulate across sessions in memory, notes and instruction files, and Fable 5.2 follows every one of them literally, so it starts asking permission for routine work. Their advice is to delete stale notes, name the steps that need no approval, narrow your stopping conditions, and rewrite absolutes. "Ask for approval before deleting files" works. "Never delete files" gets treated as absolute.
Your trio is full of rules written for older models, plus rules you've explicitly killed that keep resurfacing from stale copies.
Effort: high. Ownership sentence: no.
Read ~/memory/claude.md, ~/memory/annette.md, ~/memory/soul.md, every file in ~/memory/topics/, and every CLAUDE.md and AGENTS.md under ~/ClaudeProjects.
Audit them as instruction sets for a current model. Find and list:
- rules that contradict another rule in the same set, quoting both
- rules a newer decision already overturned, including any file still repeating a rule that another file says was deleted
- absolute "never" rules that should be "ask me first" instead
- stopping conditions and validation gates that would block routine autonomous work
- workarounds written for older models that are no longer needed
- rules that point at a file, script, hook or flag that no longer exists on this machine, verified by checking
For each finding, quote the line, name the file, and say what you'd replace it with. Change nothing yet. Rank by how often the rule would fire in a normal week.
4. The session and history audit
Your own typing is a dataset. Claude Code keeps one folder per project under
~/.claude/projects and every prompt you've typed in ~/.claude/history.jsonl.
Sessions older than 30 days get deleted by default, so raising cleanupPeriodDays
in settings now gives the next pass more to work with.
Effort: high. Ownership sentence: no.
Read my Claude Code sessions in ~/.claude/projects and my prompt history in ~/.claude/history.jsonl. Leave out any secret, key or token you encounter: never quote one, and refer to it by file and line instead.
Map how I actually work:
- prompts I type again and again, grouped by what they're really asking for
- corrections I keep giving you, and which ones I've given more than three times
- tasks I'm still doing by hand that a script or hook could do
- sessions that stalled, looped, or ended without finishing
- skills, agents and rules I have installed and never use
Then give me one plan that turns that map into new skills for the prompts I repeat, hooks for the manual tasks, memory rules for the corrections, and an archive list for what I never touch. Wait for my review before changing any file.
5. The stranger-with-access sweep
Someone ran a version of this on his laptop last week and found 14 API keys sitting
in plain text across old project folders, one of them pushed to a public repo. You
have a known live example: YT_API_KEY is still in YouTubeIntelligenceFeed git
history, and that rotation has been sitting open since July.
Effort: high. Ownership sentence: no. Read-only, every time.
Review this machine the way a stranger who just got access to it would. Look at ~/ClaudeProjects, ~/.claude, ~/memory and the git history of every repo in them.
Find and list, with file path and line:
- credentials, API keys, tokens and passwords in tracked files
- credentials in git history even where the current file is clean, naming the commit
- .env files that are tracked, or that a .gitignore stopped covering at some point
- secrets inside session logs, notes, handoffs and markdown
- anything published to a public repo or a live site that should not be there
For each one, say which service it belongs to, whether the repo is public or private, and what rotating it would break. Quote no secret value, not even partially. Change nothing. Order the list by blast radius.
Then rotate by hand at each provider. That part stays yours.
6. The automation drift check
You register every recurring job in ROUTINES.md, which only helps while it matches
what's really scheduled. Hooks are the same story: one set lived in the wrong file
for three months and never fired once.
Effort: medium. Ownership sentence: no.
Compare what ~/memory/ROUTINES.md claims is running against what is actually scheduled on this machine and in my Cloudflare account.
Check Windows Task Scheduler with schtasks /Query, the hooks keys in ~/.claude/settings.json and every project settings.json, and the cron triggers on my Cloudflare Workers.
Report four lists: running and registered, running but unregistered, registered but not running, and registered and running but last succeeded more than a week ago. For each hook, say how you verified it fires, by side effect rather than by reading its config. Change nothing.
7. The branch and unpushed-work sweep
MyExecAssistant alone has two branches idle for 104 days, each a few hundred commits behind main. Across 57 repos there's more.
Effort: medium. Ownership sentence: no.
Go through every repo under ~/ClaudeProjects. For each one report:
- commits on local branches that were never pushed anywhere
- branches with unmerged commits, how far behind the default branch they are, and how long they've been idle
- whether each branch's work already landed on main by another route, checked with git cherry rather than assumed
- uncommitted changes in the working tree, and whether they look like real work or scratch
Classify each branch as already-landed, worth-merging, or abandoned, and say what you based it on. Watch for reused migration numbers or duplicate file names, which git will merge without reporting a conflict. Push nothing, merge nothing, delete nothing.
8. The projects graveyard
PROJECT_REGISTRY.md is supposed to hold active projects only, and there are 37
Cloudflare Pages projects behind 25 domains. Some of both are finished, superseded
or dead.
Effort: high. Ownership sentence: no.
Reconcile three things: the projects listed in ~/ClaudeProjects/PROJECT_REGISTRY.md, the folders that actually exist under ~/ClaudeProjects, and the Cloudflare Pages projects and custom domains on my account.
For each project, report last meaningful commit, whether it has a live URL that resolves, whether anything else depends on it, and which of the three lists it appears in. Treat the 11 domains in my dead-domains note as intentionally dead and exclude them.
Then classify each one as active, dormant, superseded, or dead, with the evidence. Propose what to archive and what to delete, as a list for me to approve. Delete nothing and deploy nothing.
What this costs, and where to spend it
Fable 5.2 runs at $8 per million input tokens and $40 per million output, down from $10 and $50. Addy Osmani's caveat matters more than the price cut: it uses so many more tokens that the per-task cost comes out about the same as 5.1.
Effort level is where the real money is. Medium is the default in Claude Code,
/effort changes it mid-session without a restart, and on Pro and Team seats Fable
is included up to half your weekly limit.
| Job | Effort | Why |
|---|---|---|
| Sorting, grouping, renaming, cleaning an export | low | 5.2 at low beats older models at high on coding evals |
| Routine audits, reading a few files, drift checks | medium | Enough for jobs 6 and 7 |
| The cross-repo and dead-code work, long unattended runs | high | The default on the API, and where the judgment shows up |
| One job you'd be embarrassed to get wrong | xhigh | Slightly better scores for many more tokens |
Two things not to do with it. Don't use Fable for the typing: let it find the work and write the plan, then hand execution to Opus 5.5 or DeepSeek and bring Fable back to review the result. And don't ask it to paste its internal reasoning, which it refuses. "Explain your reasoning" is fine.
Garry Tan's measurement is worth keeping in view: across 52 coding tasks, Fable 5.2 solved about as many as Opus 5.5 (82.7 percent against 84.6) at 2.8 times the cost per solved task. For ordinary building, Opus stays your daily driver. These eight jobs are the other thing, the ones where nobody has been willing to decide what gets deleted.
One honest caveat
Every number Anthropic published came from their own testing on small coding tasks, and nobody has reproduced it independently. The one person who tried in public got a stronger result than Anthropic on the ownership sentence and the opposite result on the restraining rule. So run job 3 or job 6 first, somewhere low-stakes, and see what your own repos do before you paste the ownership line into all 57 projects.
Sources
- Prompting Claude Fable 5.2, Anthropic's official guide
- Introducing Claude Fable 5.2 and Claude Mythos 5.2
- Anthropic Says This One Sentence Fixes Fable 5.2. I Tested It.
- The 10 New Anthropic Rules for Prompting Fable 5.2
- bebu on the ownership finding
- Moritz Kremb, the 10 rules in plain language
- Shann Holmberg, auditing your Claude sessions
- Defileo, the laptop and security audits
- Shimecki, what $139 of audits found, including the 1,800 dead lines
- Addy Osmani on per-task cost
- Garry Tan's 52-task comparison, quoted in the X trending roundup
Published to Annette's hub. Rebuilt from the source markdown, so edit the source and rerun rather than editing this page.