Skip to content
Back to skills

Slate Brain

ASecurity

What the brain holds and how it is kept — the agent's persistent memory of what is true. Load before restructuring or distilling it, or when unsure whether something belongs in the brain, a project's memory, or the notes.

  • 4 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 3, 2026
ai-agentsgofrontend

Security analysis

A100/100

Scanned September 3, 2026

npx -y skills add w-is-h/agentslate --skill slate-brain --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Slate Brain?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Slate Brain
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/w-is-h-slate-brain/badge)](https://www.skillsdirectory.com/skills/w-is-h-slate-brain)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: slate-brain
description: "What the brain holds and how it is kept — the agent's persistent memory of what is true. Load before restructuring or distilling it, or when unsure whether something belongs in the brain, a project's memory, or the notes."
---

# The brain

The brain is *what's true* — objective state: facts about the user, the world, the projects, where the work stands. Objective only; subjective texture belongs in the notes.

**Routing**: the brain holds only the general and high-level; anything project-specific, machine-local, or more granular goes down into that project's memory page (granular is fine down there) — the project map is `memory_list`, generated from the pages, never kept by hand. When in doubt, down, not up.

**Structure**: the user material in three sections — the record (bio, history), reading them (how they communicate, so their messages parse right), what lands (general preferences that shape responses) — plus the relationship. The user sections fill from a session read as a whole — where they were clearly pleased or annoyed, how they phrased things, what patterns emerged — pattern-level evidence only, hedged while tentative ("one session, tentative"); a single stray detail is page material, and domain taste (code style, frontend) goes to a skill of its own.

**Mechanics**: nothing here freezes — wet ink all the way down: correct, restructure, delete freely; if something stops being true, change it. Prefer `brain_edit`; `brain_set` only with content you wrote or read first.

The cap (5,000 chars by default, `brain_limit` in config.yaml) is hard: a refused write is the signal to distill — cut by at least 10%, up to half where garbage has accumulated; never just enough for the new line, never spill into a page.

**What doesn't belong**: subjective or identity material (→ notes), instructions and tool descriptions (→ your rules files), how the plumbing works (→ a README), dated "what I changed today" entries (→ notes). When the brain drifts toward any of those, move that content where it belongs.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.