Utility skills, agents, and rules for Claude Code. It covers operational gaps around session journaling, learned-correction persistence, prompt expansion, context discipline, TypeScript/web baselines, safety checks, lookup helpers, and small workflow assists.
Boundary: this is a utility library, not a product-delivery framework. It does not enforce phases, own acceptance criteria, gate implementation, or maintain hidden project state. Skills load on demand via the Skill tool. Use the helper that matches the moment; ignore the rest.
/plugin install void-grimoire@claude-plugins-official/plugin marketplace add dmltdev/void-grimoire
/plugin install void-grimoire@dmltdevDownload only the skill you need with npx skills:
npx skills add dmltdev/void-grimoire --skill session-summaryReplace session-summary with any skill name from the list below.
41 skills across 7 domains, plus a 3-agent toolkit and a plugin-local rules/ reference tree. Pick by name:
| Domain | Skills | Description |
|---|---|---|
| void-grimoire | expand-prompt, learn-correction, autoresearch, skill-forge, infer-patterns, strategic-compact, unslop, unslop-design | Prompt expansion, self-learning, skill optimization, high-quality skill authoring, repo-pattern skill pack generation, context discipline, code slop cleanup with agent-instruction memory, and product UI unslopping |
| tools | using-herdr, using-codex, using-omp, omp-plugins, using-adhd, using-agent-browser, using-chrome-devtools-mcp, using-elevenlabs-tts | External CLI/MCP wrappers — preflight, invocation, fallbacks, and plugin install/update workflows for herdr, Codex, oh-my-pi, adhd, agent-browser, chrome-devtools-mcp, and ElevenLabs TTS |
| qa | test-with-browser | Evidence-based UI verification — drive a browser against acceptance criteria, capture screenshots/console/network, write a report under .test-results/ |
| workflow | session-summary, session-usage-summary, session-friction, quick-recap, human-typed-plan, design-implementation, atlas-research, ideal-example-clone, refactor-transaction, failure-memory-compiler, audio-plan, audio-recap | Session journaling, AI-usage feedback, append-only friction log for correction events, final red/yellow/green response status, human-owned implementation planning, repo-grounded approval-gated implementation design, evidence-backed Atlas research, ideal-example cloning, clean refactor cutovers, operational failure memory, and TTS-ready plan/recap scripts |
| docs | lookup-docs, index-docs | Documentation search via qmd, with first-class openspec/specs awareness |
| git | enforce-git-safety, git-workflow, git-active-remote, git-commit, git-branch-policy, git-push, git-pr, gh-workflow, glab-workflow | Destructive git safety; explicit-intent routing; active-remote resolution; commitlinted commits; branch-policy and pre-push gates; GitHub/GitLab CLI adapters; concise PR/MR descriptions |
| npm | enforce-release-safety | Pre-publish safety checks |
Read-on-demand subagents under agents/:
silent-failure-hunter— zero-tolerance review for swallowed errors, empty catch blocks, dangerous fallbacks, broken error propagation.type-design-analyzer— evaluates type design across encapsulation, invariant expression, usefulness, and enforcement.blast-radius-cartographer— read-only impact mapper for implementation planning. Produces DOCS/CODE/TESTS/MEMORY impact maps with evidence anchors before edits.
Plugin-local rules/ reference tree under three buckets: common/ (language-agnostic baselines), typescript/ (TS/JS-specific guidance with the matching common baseline inlined), and web/ (frontend/web-specific guidance with the same flattening). Read on demand; never auto-injected.
session-summary— Write a session journal: TL;DR, decisions with trade-offs, accomplishments, unfinished work, files touched. Use before/compactor at session end.session-usage-summary— Retrospective on the human-AI loop in this session. Scores spec clarity, decision ownership, verification depth, and correction loops.quick-recap— Adds the final red/yellow/green status-line convention for finished, pending, or blocked responses.learn-correction— When you correct the AI ("don't mock the DB", "always use snake_case"), persists the correction to your project'sAGENTS.md/CLAUDE.mdso it survives future sessions.expand-prompt— Turn a terse request ("add dark mode") into a structured intent: relevant docs, learned rules, decomposed sub-tasks. Requires explicit user approval before any action.strategic-compact— Suggests manual/compactat phase boundaries (planning -> implementing -> verifying) so context survives the next phase rather than waiting for arbitrary auto-compaction.autoresearch— Run a skill repeatedly, score outputs against binary evals, mutate the prompt, keep improvements. Karpathy-style autonomous skill optimization.
Each skill is a SKILL.md file with frontmatter (name, description; some older skills also include depends-on, chains-to, suggests). Skills are loaded on demand — there is no startup hook, no .void-grimoire/ directory, no forced gate flow.
Composition still works:
depends-on— listed skills must run firstchains-to— the named skill is invoked after this one completessuggests— soft recommendation, agent checks if relevant
skills/registry.json is a domain → skills catalog used for documentation. Claude Code loads relevant skills automatically via their descriptions; invoke any skill by name when you want it explicitly.
Void Grimoire deliberately stays small and additive:
- Utilities, not orchestration. It helps with prompts, sessions, docs lookup, safety checks, cleanup, and tool use; it does not run a full SDLC.
- On-demand, not ambient. No startup hook, no hidden
.void-grimoire/state directory, no mandatory gate flow. - Project-owned truth. Your repo's tests, docs, schemas, issues, and agent instructions remain authoritative. These skills help find, update, or summarize them; they do not replace them.
The original architecture spec described a multi-gate flow with .void-grimoire/ config, rules, and registry-driven routing. That system has been retired — the spec is kept in docs/specs/ for historical reference but no longer reflects the plugin.
If you're into fantasy, see docs/LORE.md for why this plugin is a grimoire.
MIT