TL;DR
Hassid's claim: accumulated harness config (instructions, skills, memory, folder scaffolding) was written for older models and now mostly costs you, so periodically delete it, default to Auto mode, and prompt with a goal instead of steps. The core idea (prune config written for a weaker model, and re-baseline against the current one) is sound and matches advice he attributes to Boris Cherny. The execution is a blanket wipe with no measurement, and the post contradicts its own infographic twice. Treat it as a prompt to run an ablation, not as a procedure.
Source: LinkedIn post by Ruben Hassid, published 2026-11-04 (about 106 reactions and 92 comments when read, 34 minutes after posting). Clean-up prompt: ruben.substack.com/p/one (not opened). Infographic credited to how-to-ai.guide. Read signed out, so the author headline wasn't visible; the text below is paraphrased.
The seven steps
Prerequisites he names: a personal account, the desktop app, and trying Claude Code. He says to redo the whole thing in 5 months.
One Claude. Check Settings, Organization Settings, Capabilities so everything is enabled in one place.
Auto, not Manual. Stop approving every step. His default is Opus 5.5 at High effort with Auto (the image says Medium).
Everything is code. Slides, images and video are code, so ask Claude to build the thing. Design, then Slides, with an uploaded reference.
Delete the old setup. Empty account instructions, remove skills, turn off "writing" skills, turn off Memory, flatten over-organized folders, delete global instructions (he claims Claude ignores them).
One clear goal. Start each chat with Goal (what and why) and Context (audience, what exists, what good looks like), then ask Claude to restate the goal and ask questions.
HTML, not .docx. Ask for a page (an Artifact) and share the link, instead of Word, PowerPoint or PDF.
Clean house. Point the desktop app at the old AI folder once and run his clean-up prompt: ask what still matters, keep one line, delete the rest.
The funnel: subscribe to how-to-ai.guide and repost. There is no keyword or deadline.
Critique
The principle holds; the procedure throws away the evidence you'd need to know whether it worked.
What holds
Config rots against model upgrades. Instructions that patched a weaker model's failure ("always run tests", "don't invent APIs") become noise, and sometimes over-constrain a stronger one. One commenter's line is the best summary: every instruction you keep was written for a model that no longer exists.
Goal plus context beats step lists. Stating the outcome and what good looks like, then asking for a restatement, is cheap and catches misreads before work starts.
Fewer approval prompts. For low-risk, reversible work, Auto mode removes the human as a per-step bottleneck.
HTML over .docx. A linked page is diffable, versioned and editable by the agent; a binary attachment is not.
What doesn't
No baseline, no eval. "Delete everything and see" with no before/after task set means you can't tell regression from improvement. The Cherny framing (delete and observe what the model does) is an ablation; Hassid's version skips the observing.
Conflates three kinds of config. Style preferences, capability patches and project facts decay at different rates. Project facts (repo layout, conventions, where things live, decisions and why) don't get obsolete when the model improves, because the model can't infer them.
"Claude ignores global instructions" is overstated. Ignored instructions are usually a symptom of too many, conflicting or stale ones, which is an argument for pruning, not deletion.
Auto is a permission decision, not a UX one. It's fine for reversible work; it's not a default for anything touching production, money or outbound messages.
Self-contradiction. Text says High effort and Memory off; the image says Medium and Memory on with per-chat opt-out. The image's Memory advice is the more defensible one.
Comment section is an echo chamber. All visible comments agree; none report a measured result. Shivani Virdi (author of R002) says to keep the context that matters outside the session, which is the opposite of "delete memory".
Doing the CTA safely
Run the clean-up as a reversible ablation: snapshot, strip, measure, then restore only what earns its place.
Snapshot everything first: account instructions, project instructions, skills list, CLAUDE.md files (dotfiles repo), memory exports. Commit them to a branch so deletion is a revert away.
Pick 3 to 5 representative tasks you ran this week (a research summary, a writing task, a coding task) as the before/after set.
Sort each piece of config into: capability patch (candidate to delete), style preference (keep one line), project fact or rationale (keep; the model can't infer it).
Read the clean-up prompt at ruben.substack.com/p/one and adapt it to output that three-way sort, not "keep one line".
Strip the capability patches and duplicate instructions; leave project facts and memory banks alone.
Rerun the task set and compare against the snapshot. Restore anything whose absence caused a regression.
Decide Auto vs Manual per surface: Auto where work is reversible, Manual (or allowlists) where it isn't.
Put a recurring reminder in about 5 months (early April 2027) to repeat the ablation after the next model upgrade.
Related entries
R002 The 5 layers of agent memory: argues memory fails when the maintenance step (dedupe, reflection, garbage collection) is missing. Hassid's wipe is garbage collection done by hand, all at once, with no eval gate.
Sources
Ruben Hassid on LinkedIn, read 2026-11-04 signed out.
Clean-up prompt: ruben.substack.com/p/one (linked, not opened). The Boris Cherny quote is as attributed by Hassid and not verified.