kit/plugins/*/README.md

kit/plugins/*/README.md · 19 files docs

@@ ## Installation — all 18 plugins @@
	## Installation
-	```bash
-	# Or install via marketplace
-	/plugin marketplace add /path/to/agentics
-	/plugin install git-agent@agentics-kit
-	```
+	### Via Marketplace (recommended)
+	```bash
+	/plugin install git-agent@agentics-kit
+	```
+	### Local Development
+	```bash
+	claude --plugin-dir ~/devbox/agentics/kit/plugins/git-agent
+	```
@@ ## Usage — skill activation modes @@
-	All skills activate automatically when intent matches.
+	\| Skill \| Activation \| Trigger \|
+	\| branch-agent \| Auto-activated \| "create a branch..." \|
+	\| commit-agent \| Manual invoke only \| /git-agent:commit-agent \|
+	\| pr-agent \| Manual invoke only \| /git-agent:pr-agent \|
+	\| ship \| Manual invoke only \| /git-agent:ship \|
+	\| ship-autonomous \| Auto-activated \| "ship and watch CI..." \|

+1028 -558

37 agents · 18 plugins · pipeline(), not parallel barriers

Social media post

Just ran a multi-agent workflow across all 18 plugins in my Claude Code marketplace — and watched 37 subagents rewrite every README in parallel. This one got me genuinely excited.

Here's what happened: plugin documentation had quietly drifted. Install sections used placeholder paths. Usage tables were missing whole skills. Manual-invoke-only features were unlabeled as such. Not broken — just stale, in 18 places at once.

The workflow ran a 3-phase pipeline:

Scout — read the marketplace manifest, pull versions + descriptions for all 18 plugins

Audit — per plugin: read the README, plugin.json, and every command/skill/agent file; extract what's accurate vs. stale

Update — rewrite install and usage sections with correct invocation syntax, standardized two-subsection install headers, and explicit "Manual invoke only" labels for skills with disable-model-invocation

Final result: 19 files, 1,028 insertions, 558 deletions. The root README even self-corrected from 17 → 18 plugins and v3.8.0 → v3.9.1 — because the agents actually read the source and cross-referenced the manifest.

The detail that stuck with me: I used pipeline() instead of parallel barriers. Each plugin flows through Audit → Update independently — so fast plugins don't sit blocked waiting for the slowest one. Wall-clock time = slowest single plugin, not the sum of 18. That one structural choice saves real minutes on any fan-out job this size.

Multi-agent orchestration isn't just for code generation. Documentation drift is a real problem at scale, and agentic pipelines are a deeply underrated fix.

Follow for more Claude Code plugin patterns and agentic workflow deep-dives.

#ClaudeCode #AIAgents #PluginDev #DeveloperTools