Compound Learning Capture
Capture solved problems as durable repo knowledge right after the work, so every subsequent unit of engineering gets easier.
The skill is a documentation generation tool, not involving high-privilege operations, but it executes external commands (Node script) and constructs environment variables from path arguments, potentially vulnerable to path injection. In interactive mode, it asks for user consent before editing files; non-interactive mode does not edit instruction files. No data exfiltration or sensitive information handling. Publisher identity unverified, but license is clear (MIT). Deductions: no explicit principle of least privilege, external command execution lacks sandboxing; no rollback mechanism; insufficient data flow disclosure.
The skill description is detailed, with multiple modes (interactive/non-interactive/lightweight/full), and provides fallback for subagent coordination failures (e.g., #956). However, no test suite is provided, and it relies on external script (context.mjs) whose behavior is unverified. Error handling for abnormal input is only partial (e.g., unknown depth token generates failure report), but overall failure feedback is not nuanced. Static review cannot run tests; deductions: no test evidence, robustness of parallel subagent coordination unverified.
Target scenario is clear: documenting solved problems. The skill defines applicability boundaries (one learning per run, CONCEPTS.md bootstrap redirection), and supports Chinese users (skill itself is in English, but not optimized for China networks). Trigger conditions are clear (e.g., '/ce-compound' command). Deductions: documentation is English only, no Chinese version; reliance on external services (e.g., GitHub) may be unreachable from mainland China, affecting usability.
Information architecture is good, using layered documentation (SKILL.md, references, templates), with usage examples. Provides configuration docs (docs_root) and version number (package.json), MIT license. But lacks changelog, maintenance responsibility unclear (publisher unverified). Deductions: no CHANGELOG, no version history file, governance incomplete.
The skill has clear goals (generate structured docs) and a complete workflow (multi-stage subagents). But the value proposition (knowledge compounding) is unverified, static review cannot confirm output quality. Deductions: no real output examples, no evidence of being more effective than manual recording.
Repository has CI (GitHub Actions PR title validation), but does not cover key paths of this skill (no unit tests). Internal scripts (validate-frontmatter.py etc.) exist but without tests. Static review cannot independently reproduce. Deductions: no third-party execution evidence, coverage insufficient.
- The skill runs external Node scripts during execution; ensure the path arguments are trusted to prevent path injection.
- Publisher is not verified by FollowSkills' curated enterprise registry; assess before use.
- Skill documentation is English-only, which may be less accessible to Chinese mainland users, and relies on online services that may be unreachable.
- Static review did not execute any tests; actual runtime behavior should be verified by the user.
What does this skill do, and when should you use it?
This skill automatically captures problem solutions while context is fresh, using parallel subagents to research and structure the information into a Markdown document with YAML frontmatter, stored under `<root>/solutions/`. It supports interactive and non-interactive modes, as well as full and lightweight depths, and can probe session history to incorporate relevant prior attempts. Part of a 38-skill collection, this skill focuses on knowledge compounding to accelerate future work.
Parses conversation history to extract problem, symptoms, solution, and prevention; generates YAML frontmatter with category, problem type, and component; writes documents to <root>/solutions/<category>/; runs multiple subagents in parallel (Context Analyzer, Solution Extractor, Related Docs Finder) to gather context; runs a session-history probe to discover relevant prior sessions; updates existing docs or creates new ones based on overlap; optionally updates CONCEPTS.md for vocabulary capture.
- A developer who just fixed a tricky bug wants to document the root cause and fix to avoid future rework.
- A team wants to capture lessons learned from a complex troubleshooting session for future reference.
- An automation pipeline runs in CI and needs to capture learnings without human intervention, using non-interactive mode.
- A developer with limited context budget chooses lightweight mode to quickly log key takeaways without subagents.
- When session history contains relevant failed attempts, the skill integrates them to prevent repeating mistakes.
What are this skill's strengths and limitations?
- Automated parallel research saves time
- Structured output ensures long-term reusability
- Supports non-interactive mode for automation
- Integrates prior session history to enrich docs
- Configurable `docs_root` for repo flexibility
- Built-in YAML frontmatter validation prevents corruption
- Requires Node.js and Python 3 runtimes
- Needs shell access to run scripts and subagents
- Session-history probe may add complexity in large repos
- Full mode may consume significant tokens, though usually worth it
- Related docs finder depends on GitHub CLI (`gh`); if missing, issue search is skipped
How do you install this skill?
Install the Compound Engineering plugin (EveryInc/compound-engineering-plugin) via your client's marketplace (Claude Code, Cursor, Codex, etc.). The skill is located at skills/ce-compound/SKILL.md within the plugin.
How do you use this skill?
In Claude Code, simply type /ce-compound optionally followed by context, e.g., /ce-compound record the recent database migration issue. For non-interactive runs, use mode:non-interactive with optional depth:lightweight or depth:full. The skill automatically selects full or lightweight mode and orchestrates subagents to produce the documentation.
How does this skill compare with similar options?
Compared to other documentation skills, this focuses on learning capture with parallel subagents and session-history integration. Within the same plugin, ce-compound-refresh handles refreshing stale docs, while ce-compound creates new ones. External tools like Notion AI lack this deep integration.