R036 · Harness tips › Harness tips

Ruben Hassid's Claude 101, checked

Read the original

harness-configpromptingagent-memorycontext-engineeringdata-privacy

Sources: LinkedIn post (2026-11-11), the free part of the Claude 101 guide (2026-11-09; tips 18 to 27 are paywalled), and the Anthropic post it cites, The components of agent memory (2026-11-06).

TL;DR

Ruben's third Claude setup guide in a week (after R004 and R018) is 17 consumer tips. The best of them are context hygiene: edit the prompt before a bad answer instead of arguing, define "done", don't steer the answer, and hand off to a fresh chat with a 10-line summary when a chat goes stale. His headline claim, that turning Memory on makes Claude "28% better", is real data from Anthropic, but from a different system: an agent with a persistent memory folder on a long-horizon benchmark, not the claude.ai chat Memory toggle. The memory clean-up prompt is the most useful practical takeaway, and it is now a reversal of R004, which told you to wipe memory entirely.

The tips, grouped

Memory (tips 1, 2, 6). Keep Memory on, keep it short, and periodically run a clean-up prompt that lists every memory, marks each KEEP / FIX / DELETE, asks about uncertain ones with clickable options, and produces a cleaned version under 30 lines. His own memory went from 105 facts to 26. One Project per client with "Use account memory" off so client facts don't cross over.

Context hygiene (tips 3, 12, 13, 17). A bad answer stays in the transcript and keeps influencing every later turn, so edit the earlier prompt instead of replying "no, that's wrong". Give a finish line ("Done means: you checked every number against my file") and the only reason to stop ("stop and ask only if a number doesn't match"). Don't ask leading questions ("$49 is right, isn't it?"); ask for the strongest arguments against. When a chat is long or a day old, ask for a 10-line summary (goal, decisions, what's left) and continue in a new chat.

Instructions (tips 5, 15, 16). Empty the instructions box (about-me, style files, rule lists) and give one clear goal per prompt; the one prompt you need is Goal + Context + "ask me if you need more". Dictate context instead of typing it, and send a screenshot instead of describing a design.

Divergence (tip 10). "Give me 20 very different options, with how likely another AI would give each; the last two must be under 10%." A cheap way to get past the modal answer.

Product features (tips 4, 7, 8, 9, 11, 14). Turn off the setting that moves a flagged chat to an older model; Claude in Google Docs/Sheets/Slides; paste a GitHub repo link as an ad-hoc skill; ask for dashboards and files; never press thumbs up/down (repeated from R018); export a bank CSV instead of connecting the bank.

What holds up, and what doesn't

The 28% number is real but borrowed. Anthropic's Figure 3 shows Opus 5.5 at 0.655 stateless vs 0.838 with a memory store on CL-Bench 1.0, about +28%. But that run is Claude Managed Agents answering database questions in fresh sessions with a persistent memory directory and grader feedback after each answer. It is evidence that a self-maintained memory loop helps long-horizon agent work. It says nothing measured about the consumer Memory toggle on everyday chats.

"Keep it short" is well supported. The same post: Sonnet 3.7 wrote 413 files in 225 folders and won one Pokémon badge; Fable 5 kept nine files plus a global.md and won four; Opus 5.5 kept a small folder and beat the game. Anthropic's line that "the top layer of the memory hierarchy has to stay relatively small" is about what gets injected into every session, and Claude Code's docs say the same thing (MEMORY.md loads only its first 200 lines or 25KB; CLAUDE.md files over 200 lines reduce adherence).

"Empty your instructions box" has a better reason than the one given. Ruben says instructions make every answer the same. Anthropic's finding is sharper: for strong models, fixed rules about what to write to memory ("record only failures") removed much of memory's gain, while older models benefited from steering. So: tell the model what something is for, not how to do it, and expect the right amount of instruction to shrink with each model generation. That is R004's "config rots" principle, now with data.

"Claude rereads your entire chat" is the right mental model for a different reason. As noted in R018, the cost is cache and attention, not literal rereading; but the advice (remove the bad turn instead of arguing with it) is correct either way, because the bad answer stays in context for every later turn.

The clean-up prompt is a manual version of "dreaming". Anthropic's post describes an offline pass that checks notes against transcripts, fixes wrong ones, and pushes detail down the hierarchy to keep the injected layer small. Ruben's prompt does the same with the human as the source of truth. It lacks one thing dreaming has: checking each memory against evidence rather than asking "is this still right?"

Not checked: the "downgrade" setting in Settings → Capabilities (tip 4) and tips 18 to 27 behind the paywall.

Internal tension across Ruben's posts. R004 (Nov 4): wipe memory. R018 (Nov 8): keep Team memory on. R036 (Nov 11): keep memory on, but curate it. The direction is converging on the right answer (curate, don't wipe), but none of the three posts measures anything before and after.

Checklist for applying it

R004 and R018 (same author, earlier guides), R002 (the 5 layers of agent memory), R017 (fresh-context review), R005 (context layer).

Connections

See R036 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.