I ran retention tests on myself to find out how Claude should talk to me

I'll open the source article so the paragraphs stay factual and don't invent numbers or claims.Danilo Velasquez published **Make Claude speak TO you, not LIKE you** after noticing that language models mirror how a user writes, not how that user retains information. He had told Claude that a sentence made him feel stupid so he stopped reading; that moment became the signal for where the writing failed. He then ran retention tests on himself instead of relying on self-report, and turned the results into a personal **communication protocol** file his agent loads at the start of every session. The piece is also listed on Hacker News under a related title about those retention tests, with a small early thread.

The protocol is not a vague “be concise” preference. Each rule had to beat an alternative in a small retention check: short cause-and-effect sentences over bullet lists of loose facts; **bold** on key terms and values so re-reads are cheap; and framing any explanation as a consequence for the reader before the mechanism. He illustrates the difference with the same **debounce** example for a search input: the default version opens with the abstract definition; the tuned version opens with what breaks (typing **“laptop”** fires six stale requests), bolds identifiers, walks a causal chain, and only then defines debounce with a short delay such as **300ms**. Interest itself acted as a gate—content that did not connect to something he cared about did not stick no matter the formatting.

For engineers and builders who live in long agent sessions, the cost is attention, not tokens. Every answer that must be re-decoded or re-scanned spends capacity that should go to the task. A file that encodes how *you* actually retain technical material—especially when learning versus when Claude is reporting what it just did—cuts review fatigue across every later session. The work also surfaces conditional rules people usually flatten: analogies helped Velasquez when learning a new concept, but were dead weight mid-task when he only needed to know what changed.

Market context is the default assistant voice aimed at an average reader. Claude and similar tools already mirror register and style; that is useful for collaboration and weak for retention when the user’s speaking style and learning style diverge. Velasquez notes the same protocol helps other assistants and agents, with stronger results in **Claude Code**, where fewer competing system prompts dilute the instructions, and still useful (if noisier) in the web UI. The competitive angle is not a new model feature but a user-side calibration layer that sits above whichever tool you use.

The practical path is the four-phase prompt he published: interview one question at a time about how you take information, then run at least four retention checks that explain the same content two ways and see what you can repeat back, then derive one rule per constraint with a one-line reason, then write a finished **CLAUDE.md** (or custom-instructions block) with a profile of cheap vs expensive mental operations, numbered rules, a session “reading signal,” and a low-capacity fallback format. Budget twenty to thirty minutes; be blunt when something bores you or you abandon a sentence. What to watch next is whether other people report a large interest-versus-neutral retention gap and whether domain material from their own stack changes results the same way—if those patterns hold, personal communication protocols become a standard agent setup step rather than a one-off blog experiment.