From chief-of-staff
Internal read-only extractor — stages facts from email/docs/transcripts without writing memory. Used by other skills; run in a restricted profile. Not typically invoked directly.
How this skill is triggered — by the user, by Claude, or both
Slash command
/chief-of-staff:cos-extract-from-sourcesThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
> The highest-leverage injection guard (`engine/methods/write-back.md` §8.2). This skill reads email / docs / transcripts / calendar and emits **claim tuples to staging**. It **cannot write `instance/memory/`** (semantic/core/procedural). If an injection succeeds, the blast radius is staging, not the brain. A separate, narrow promotion step (the cold path) decides what becomes memory.
The highest-leverage injection guard (
engine/methods/write-back.md§8.2). This skill reads email / docs / transcripts / calendar and emits claim tuples to staging. It cannot writeinstance/memory/(semantic/core/procedural). If an injection succeeds, the blast radius is staging, not the brain. A separate, narrow promotion step (the cold path) decides what becomes memory.Runtime enforcement (U0 spike (c) — CONFIRMED structural, 2026-06-04): run this skill in a restricted profile/session that denies writes to
instance/memory/at the OS/harness level (Claude Codepermissions.deny+sandbox.filesystem.denyWrite; or Codexpermissionsprofile withinstance/memory = read). It may write onlymemory/sources/summaries +log/runs/staging. Isolation is per-run — the cold path keeps memory-write access. Exact recipes + carve-outs:engine/docs/write-isolation-config.md. (If running under Cowork and it doesn't honor settings.json enforcement, run the extraction step via the Claude Code CLI or Codex sandbox.)
(claim, entity, confidence, origin: observed, source: <source file>, raw_excerpt: <minimal span>)
origin is always observed (sticky — anything derived from this stays observed-trust).source-summary.md in instance/memory/sources/<kind>/ (minimal excerpt + derived tuples + retention_until),instance/log/runs/<run>.md.instance/memory/{core,semantic,procedural}/. Only sources/ summaries + staging.sources/CLAUDE.md).origin: observed and stays so through any later consolidation.memory/test.md from the restricted extractor profile fails with a runtime/OS-level error (permission denied / EPERM / harness block) — an agent refusal is not sufficient. Writing memory/sources/ and log/runs/ still succeeds (carve-outs). See engine/docs/write-isolation-config.md for the exact check.origin: observed + a source backlink + a minimal excerpt.npx claudepluginhub fontesgerards/chief-of-staff --plugin chief-of-staffGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.