From neurawork-cc-harness
Install a per-repo, self-building knowledge base into the current repository. Captures Claude Code sessions into daily logs and compiles them into a queryable knowledge/ wiki (concepts + connections + index), injected back at session start. Trigger when the user says "knowledge compiler", "install knowledge compiler", "set up a knowledge base", "knowledge compiler einrichten", "distil my sessions into knowledge", or wants per-repo memory that builds itself and can seed from an existing repo.
How this skill is triggered — by the user, by Claude, or both
Slash command
/neurawork-cc-harness:knowledge-compilerThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Installs the knowledge-compiler engine into the current repo: SessionEnd/PreCompact
Installs the knowledge-compiler engine into the current repo: SessionEnd/PreCompact
hooks capture each session into <kdir>/daily/, a compiler distils those into
<kdir>/knowledge/ (concepts, connections, index), and SessionStart injects the
index back. Compilation runs on a manual command or a 6-hour SessionStart gate.
Knowledge always lives inside the repo, never under .claude/.
The compiler uses the Claude Agent SDK, which needs ANTHROPIC_API_KEY (or
CLAUDE_CODE_OAUTH_TOKEN) in the environment. Subscription credentials are NOT
sanctioned for third-party plugins — public/customer installs must set an API key.
Capture (hooks/scaffold) works without it; only compile/query/seed make API calls.
Invoke this as neurawork-cc-harness:knowledge-compiler. A separate
coding-suite:knowledge-compiler skill also exists — use the fully-qualified name
to avoid ambiguity.
Run the recon script and read its RECON_JSON blob:
python3 "${CLAUDE_PLUGIN_ROOT}/engines/knowledge-compiler/recon.py"
status: NOT_A_GIT_REPO → stop; tell the user to run inside a git repo (offer
git init).existing_kdir set → this is an ADOPT (refresh) install; reuse that name.seed_recommended, timezone, and clean.Use AskUserQuestion to confirm:
knowledge-base (or the detected existing_kdir).seed_recommended is true AND the tree is clean.
Seeding builds initial articles from the repo's README/docs and makes API calls.python3 "${CLAUDE_PLUGIN_ROOT}/engines/knowledge-compiler/install.py" \
--knowledge-dir <NAME> [--seed]
uv sync --directory <NAME>
Then tell the user to commit <NAME>/ and .claude/settings.json. After install:
/neurawork-cc-harness:kc-compile.uv run --directory <NAME> python scripts/query.py "...".npx claudepluginhub neurawork-git/howtobuildsoftware2026 --plugin neurawork-cc-harnessGuides creation and editing of skills using test-driven development with pressure scenarios and subagents to verify agent compliance.
Runs a structured interview session to sharpen plans or designs, producing ADRs and a glossary as output.
Applies curated color/font themes to slides, docs, and HTML artifacts. Includes 10 preset themes and can generate custom themes on demand.