R018 · Harness tips › Harness tips

Ruben Hassid's six levels of Claude, checked

Read the original

harness-configpromptingprompt-cachingagent-memoryagent-governancedata-privacy

Source: LinkedIn post by Ruben Hassid ("Nobody has ever explained all of Claude to you", 2026-11-08, about 760 reactions and 125 comments 8 hours in). Text only, no infographic. Its one link is the author's skill library, skillsclau.de.

TL;DR

A consumer-level setup guide in six "levels", from the same author as R004, four days later. Four of the six are reasonable defaults; two are worth knowing even at staff level, and both check out against Anthropic's own pages: thumbs up/down sends the whole conversation to Anthropic for up to 5 years and it may be used for training, and on Team and Enterprise, an owner turning memory off for the organization permanently deletes every member's memory. The model-switching advice is right for the wrong reason (it is a cache miss, not a "re-read"). The post contradicts itself on skills: delete yours, then install his, from a 33,971-skill email-gated library. It also quietly reverses R004's "default to Auto" in favor of a fixed Opus 5.5 on high.

What the post says

  1. Use the desktop app. It works with local folders without uploading files one by one, and it pulls you toward Claude Code.

  2. Don't choose a model each time. Default to Opus 5.5 on high, and don't switch mid-chat, because "Claude re-reads the entire chat" and burns tokens.

  3. Effort is Medium or High, nothing else. Medium for light work, High for most; Extra High mostly burns your usage limit.

  4. Delete your old setup. Clear the custom-instructions box and old skills, since they were written for older models. Give one goal, and end prompts with "Before you start, restate my goal in your own words."

  5. Ask for HTML, not a document. Add "Make it an HTML page" and "for [who will read it]". Claims there is then "no scope of making mistakes".

  6. Protect what you plug in. Never click thumbs up or down, since feedback may go to training. On a Team plan, leave memory on once it's on: if an owner turns it off, every member's memories are deleted permanently.

Then a 20-minute plan: install the desktop app, no settings, connect one tool, use Claude Code more, delete installed skills and add new ones, and visit his library.

Claim check

#ClaimVerdictWhy
1Desktop app beats the browserMostly rightLocal folder access and Claude Code remove the upload step. The trade-off it skips: the agent can now read and write your disk, so permission prompts and folder scoping matter (R006's permissions by blast radius).
2Don't switch models mid-chat; it "re-reads the chat"Right advice, wrong mechanismEvery turn already sends the whole conversation; prompt caching is what makes that cheap. Caches are per model, so a switch is a cold cache: the full history is billed at full input price once, and again on any switch back. The cost scales with conversation length. A second, quieter cost: the new model inherits a transcript shaped by the old one's style.
2bDefault to Opus 5.5 on highReasonable, unmeasuredSimple and stable, which also helps caching. It contradicts R004 (same author, 2026-11-04), which said to default to Auto. Neither post shows evidence.
3Medium or High effort onlyPlausible heuristicHigher effort spends more thinking tokens; whether it pays depends on the task. Long agentic or hard-to-verify work can repay Extra High by avoiding retries; no data is offered either way.
4Delete old instructions and skillsSame as R004, same gapPruning stale config after a model upgrade is sound, but wiping without a before/after eval also deletes project facts and rationale. Skills cost little until they trigger (only their descriptions sit in context), so the risk is a stale skill that triggers and overrides newer model behavior, not idle ones.
4bEnd with "restate my goal in your own words"GoodA cheap alignment gate: misreads surface before work starts. It is the one-line version of EGM's comprehension test (R017).
5Ask for HTML, "for [audience]"Half rightNaming the reader is the real win. HTML gives rich layout but is harder to diff, edit and review than Markdown or a doc; it does nothing about factual errors. "No scope of making mistakes" is false.
6aThumbs up/down can go to trainingConfirmedAnthropic's Privacy Center (consumer plans): feedback stores the entire related conversation, including content and style preferences, for up to 5 years, de-linked from your user ID, and it may be used to train models. It excludes raw connector content unless pasted into the chat.
6bTeam owner turning memory off deletes everyone's memoryConfirmedClaude Help Center: when an owner turns memory off for the organization, all existing memory entries for all users are deleted immediately and permanently. Owners can't view or edit members' memories.

Takeaways for a staff engineer

Review checklist

Source notes

The post is paraphrased. Comments seen: a handful of short ones (praise, and one asking how a skill differs from an agent); none add technical substance. Privacy claims checked against Anthropic's Privacy Center article "Is my data used for model training?" (consumer plans; commercial plans have a separate article) and the Help Center article on chat search and memory. skillsclau.de was opened: a library of 33,971 skills, sorted by job, download requires an email sign-in; individual skills were not reviewed. Effort-level and caching statements are from general knowledge and are not re-measured.

Connections

See R018 on the knowledge map

Part of the Research notes on nowucca.com. Figures and claims are the source's own unless the summary says otherwise.