A compound learning system for Claude Code. Not a prompt collection — a system that gets better every time you use it.
18 skills, 7 agents, and a compound learning loop, packaged as a Claude Code plugin. It turns Claude Code into a structured knowledge work environment for research, strategy, writing, and analysis.
The core idea: every session should make the next session better. Skills capture lessons from real use. Insight files accumulate institutional knowledge. Review agents catch what you miss. The system compounds over time.
This grew out of months of daily knowledge work — marketing strategy, competitive analysis, executive communication, research synthesis. The skills encode real methodologies from real practitioners, not generic prompts.
Compound over time. The system has a dedicated "Compound Phase" after significant work. It extracts patterns, templates, preferences, and failure modes into atomic insight files. These insights are searchable and feed back into future work.
Skills self-improve. When a skill produces unexpected results or you correct its output, the correction gets folded back into the skill file itself. Skills with "Observed Patterns" sections are living documents.
Every session makes the next better. Session closing automatically extracts decision logs, insights, and pattern seeds. Weekly reflection synthesizes these into cross-session patterns. The knowledge graph grows without extra effort.
Structure when it helps, not when it doesn't. Quick tasks skip ceremony. High-stakes work gets full review batteries. The system classifies stakes automatically and adjusts rigor accordingly.
Paste both lines into Claude Code:
/plugin marketplace add elrolio/skills
/plugin install claude-knowledge-work@elrolio
That's it. The plugin adds all 18 skills, 7 agents, and the compound-phase hook.
/plugin marketplace add elrolio/skills
/plugin install claude-knowledge-work@elrolio
Installs everything. Skills are available immediately by name.
git clone https://github.com/elrolio/claude-knowledge-work.git
cp -R claude-knowledge-work/skills/* ~/.claude/skills/For people who want to read the source, modify skills, or contribute.
Browse skills/. Copy any individual skill folder into your own .claude/skills/ directory — each one is self-contained. The pack groupings (editing, research, strategy) are documented in docs/packs/.
Every significant task follows: Research → Work → Review → Compound.
- Research: Gather context, check precedents, map constraints. Don't produce anything yet.
- Work: Execute against the plan. Use domain-specific skills.
- Review: Validate before shipping. Rigor scales with stakes.
- Compound: Extract what you learned. Update skills. Capture insights.
Trivial tasks skip to Work. Standard tasks do Research → Work → Review. High-stakes tasks do all four.
The system classifies work into three levels and adjusts review rigor:
| Level | Examples | Review Rigor | Compound |
|---|---|---|---|
| Low | Meeting notes, Slack updates, working drafts | Quick sweep | Skip |
| Medium | Briefs, analyses, internal one-pagers | Business sweeps (clarity 1-5) | Optional |
| High | Exec deliverables, customer-facing, finance docs | Full battery (clarity + interview + accuracy + sensitivity) | Mandatory |
Stakes are auto-detected from context (audience, visibility, reversibility) but you can override.
After significant work, the system extracts four types of learnings:
- Patterns — What approach worked? What's reusable?
- Templates — Did the work produce a reusable structure?
- Preferences — What style/format choices were reinforced or corrected?
- Failures — What went wrong? (Most valuable — anti-patterns prevent repeat mistakes.)
Each learning becomes an atomic insight file with structured YAML frontmatter, stored in insights/[skill-name]/. Searchable, composable, accumulating.
---
type: pattern
skill: clarity-editing
category: structural
date: 2026-02-15
project: quarterly-review
takeaway: "Leading with the recommendation before the evidence doubles exec engagement"
tags: [executive-communication, structure]
---
## Evidence
[What happened, what data supports this.]
## Application
[When and how to apply this learning.]The writing quality layer. Includes a 7-sweep clarity editor, a de-LLMification pass that strips AI-generated patterns, accuracy validation, and platform-specific formatters for blog, LinkedIn, and X.
| Skill | What It Does |
|---|---|
clarity-editing |
7-sweep structural and rhetorical editing pipeline |
de-llmification |
Strips AI writing patterns — hedging, filler, generic transitions |
accuracy-validation |
Evidence and claim verification against sources |
edit-for-blog |
Blog post structure and optimization |
edit-for-linkedin |
LinkedIn post formatting and hooks |
edit-for-x |
X/Twitter thread structure and formatting |
Expert methodologies encoded as skills. Each one implements a specific framework from a specific practitioner — not generic advice.
| Skill | What It Does |
|---|---|
market-health-check |
Market sizing with TL;DR templates and validation |
six-growth-engines |
Growth engine classification and strategy |
pricing-strategy |
Pricing model analysis and recommendations |
launch-strategy |
Launch planning with phased execution |
content-strategy |
Content planning and editorial strategy |
market-anthropology |
Qualitative market research methodology |
barrier-inversion |
Turning objections into selling points |
gaccs-framework |
Campaign planning: Goals, Audience, Channels, Creative, Stakeholders |
Federated search across documents, memory, and connected tools.
| Skill | What It Does |
|---|---|
recall |
Federated search across vault, memory, and connected tools |
claude-knowledge-work/
├── CLAUDE.md # Plugin instructions (loaded by Claude Code)
├── README.md # This file
├── LICENSE # MIT
├── .claude-plugin/
│ └── plugin.json # Plugin manifest
├── skills/ # All 18 skills, one folder each
│ ├── diary/ # Decision-log extraction (10 memory types)
│ ├── reflect/ # Cross-session pattern analysis
│ ├── strategic-interview/ # Socratic problem modeling
│ └── ... # + 15 editing, strategy, and research skills
├── docs/packs/ # Pack groupings: editing, strategy, research
├── agents/
│ ├── compound/ # 4 agents for automated learning extraction
│ └── review/ # 3 agents for validation
├── hooks/ # Event-driven automation
├── templates/
│ └── insight-file.md # Insight file template
└── examples/
├── insights/ # Example insight files by skill
└── decision-logs/ # Example decision-log entries
The examples/ directory contains sample insight files and decision logs showing the compound system in action. These demonstrate the format and level of detail that makes insights searchable and reusable.
The skills are domain-agnostic but were built for B2B marketing and strategy work. To adapt:
- Editing pack: Works as-is for any writing domain.
- Strategy pack: The frameworks are universal (pricing, growth, launch). Swap the example companies in skill files to match your industry.
- Research pack: Configure search tools for your stack (the recall skill routes across whatever tools you have connected).
- Compound system: Works for any knowledge work. The insight file format and diary extraction are domain-independent.
The compound learning architecture was inspired by ideas from Nabeel Hyatt and the coworkpowers community around AI-augmented knowledge work, as well as Every.to's writing on AI workflows.
Strategy skills encode methodologies from their original authors:
- Elena Verna — Growth anti-patterns, growth engines, PLG frameworks
- April Dunford — Positioning and competitive context
- Gibson Biddle — Pricing strategy (DHM model)
- Framework authors are credited in individual skill files
MIT — see LICENSE.