dot-claude
Sawyer-Middeleer/dot-claude
A collection of actually useful skills and subagents for Claude Agent for non-coding use cases
Instalação
npx convez-ai Sawyer-Middeleer/dot-claude1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
dot-claude
Actually useful Claude Code plugins for knowledge work — research, notes, ops, and analysis — not hardcore software engineering. Four plugins, install only what you want.
Install
Add the marketplace once:
/plugin marketplace add Sawyer-Middeleer/dot-claude
Then install any plugin (the @dot-claude suffix names the marketplace explicitly, so install works even if another marketplace ships a like-named plugin):
/plugin install compound@dot-claude
/plugin install safeguard@dot-claude
/plugin install second-brain@dot-claude
/plugin install deep-research@dot-claude
Run /reload-plugins after installing if the new skills don't appear immediately. Plugins install at user scope (available in every project) by default.
Plugin skills are invoked namespaced — /compound:creating-skills, /second-brain:processing-inbox — or Claude triggers them automatically when their conditions match.
The plugins
compound — Claude that gets better every week
The thesis: do a task once by hand, codify it the second time, and fix the instructions every time they fail. Three skills that close the loop:
| Skill | What it does |
|---|---|
correcting-mistakes | After an error gets resolved, finds the instruction file that caused it and applies a succinct fix — so the same mistake never happens twice |
codifying-tasks | Notices when you're doing something for the second time and offers to turn the proven procedure into a reusable skill |
creating-skills | Authors high-quality skills to spec — guided scoping, current frontmatter options, progressive disclosure, and a trigger-verification pass |
Built-in auto memory captures terse facts automatically; compound is the deliberate other half — it repairs the instruction at its source (not an external fact store) with root-cause triage, and its creating-skills reference is verified against the current Claude Code frontmatter spec.
safeguard — seatbelts for knowledge work
Permission rules are enforced by the harness — stronger than anything you put in a prompt.
| Skill | What it does |
|---|---|
securing-claude | Short scoping dialog → hardened settings.json (deny rules for secrets, asks for risky actions) that you approve before it's written |
auditing-access | Read-only report of what Claude can currently see and do: settings layers, effective rules, connected MCP servers, hooks, additionalDirectories, plugins, and sensitive files no deny rule covers |
Rules are the floor: securing-claude also points you to built-in /fewer-permission-prompts for usage-mined allow-lists, and to sandboxing/PreToolUse hooks for the isolation deny rules can't give.
second-brain — a knowledge base Claude maintains
Already have years of notes? Point it at them. second-brain's wedge is adopting an existing vault in place — it reads your folder, codifies your actual conventions, and maintains them, without reorganizing anything. (It also sets up a new base from scratch.) Plain markdown; works with Obsidian or any editor. One setup skill, four maintenance skills:
| Skill | What it does |
|---|---|
setting-up-knowledge-base | Adopts an existing notes folder (codifying its conventions, never clobbering its files) or creates a new one — a conventions.md (the authority all other skills read) plus a CLAUDE.md so every session knows the rules |
capturing-knowledge | Proactively saves durable learnings to their canonical home — with an aggressive filter for what doesn't belong |
organizing-notes | Consistency passes: frontmatter on every note, controlled tags, links between related notes and entities |
processing-inbox | Routes unsorted captures to their homes; moves what's clear, asks about what isn't, never deletes |
managing-daily-notes | Daily logs with task carryover and an end-of-day close-out |
Capture leans on Claude noticing the moment, by design. Obsidian power users should pair it with kepano/obsidian-skills for deep format support (Bases, Canvas, callouts); second-brain owns curation and conventions, not format coverage. Note the overlap with built-in auto memory — second-brain keeps substantive notes in your KB; let auto memory keep terse machine-local facts.
deep-research — multi-source research, filed properly
| Component | What it does |
|---|---|
deep-research skill | Clarify scope → plan angles → validate sources (paywall-aware) → spawn parallel analysts → reconcile their output → synthesize by theme, verify high-stakes claims against originals, and emit a real URL bibliography. Deliverables are files, not a chat reply — and if you have a second-brain knowledge base, they file themselves following its conventions |
research-analyst agent | Haiku-powered analyst spawned one-per-source for parallel deep reading |
analyzing-source skill | The analysis framework each analyst follows |
The differentiator vs. native research or chat-based deep-research: output that files itself into your knowledge base as reusable, citation-linked notes — pairs directly with second-brain.
Philosophy
- Modular — small composable skills, one concept each, install only the plugins you want
- Compounding — mistakes become corrections, repeated tasks become skills, learnings land in a knowledge base
- Settings over prompts — anything that must hold is a rule the harness enforces
- Files over chat — work products land somewhere durable and findable
Migrating from v2
v2 shipped a single dot-claude plugin; v3 splits it into the four plugins above. Third-party marketplaces don't auto-update, so refresh the catalog first, then swap:
/plugin marketplace update dot-claude
/plugin uninstall dot-claude@dot-claude
/plugin install compound@dot-claude # plus any others you want
/reload-plugins
The /deep-research command is now /deep-research:deep-research, and creating-skills + correcting-mistakes live in compound.
Releasing changes
Each plugin's version lives in its plugin.json only (the single source of truth — the marketplace entry intentionally omits it to avoid a split-brain pin). When you ship a change to a plugin, bump its plugin.json version (semver) so installed users pick it up. CI runs claude plugin validate --strict on the marketplace and every plugin.
License
MIT