Kelvin’s Engineering Wiki

Engineering Wiki

5 min read

A wiki the LLM compiles from the sources in raw/ and the legacy articles, following the conventions in the repository’s CLAUDE.md. Every page lists its sources; pages marked status: draft have not been reviewed yet.

  • Domains - Concept pages grouped by domain.
  • Sources - One summary page per ingested source.
  • Syntheses - Answers filed back from queries.
  • Log - What changed and when, newest first.

Concept

  • AI adoption maturity - A four-stage ladder from individual chat use to end-to-end processes, driven by growing user fluency, system access, and governance.
  • Agent skill - A folder of task-specific instructions and optional resources that Claude Code loads only when a request matches its description.
  • Auto Mode - A Claude Code permission mode where low-risk actions run directly and a separate classifier reviews higher-risk ones against user intent and a configured trust boundary.
  • Context isolation - Keeping an agent’s intermediate work out of the main context window, at the cost of losing whatever the summary leaves out.
  • Claude Code extension mechanisms - How CLAUDE.md, skills, subagents, hooks, and MCP servers differ, and which job each one owns.
  • Prompt injection - Instructions hidden in content an agent reads, such as web pages, files, or issue comments, that try to redirect it away from the user’s request.
  • Subagent - A worker agent that Claude Code hands a bounded task to, which runs in its own context and returns only a focused result.

Pattern

  • AI-native SDLC - A software delivery loop where each stage leaves a committed artifact, agents work between human approval gates, and production evidence returns as new intent.
  • Measuring an AI rollout - Judge an AI tooling rollout by comparing concurrent cohorts against pre-set baselines, and lead with expansion rather than hours saved.
  • Agents propose, people and policy accept - Route every agent change through pull requests so that required checks, code owners, and approval gates, not the agent, decide what is accepted.
  • Risk-based autonomy - Give agents more autonomy on low-risk, reversible work and keep human approval for high-risk and production changes, widening scope gradually.
  • Self-improving skill loop - A scheduled agent reads human feedback on another agent’s output and opens a pull request that edits that agent’s skill file.
  • Delegation contract - What a subagent must be told up front: when it is used, what its output looks like, and which obstacles it must report.
  • Least-privilege tool access - Grant an agent only the tools its job requires, starting from what it must do.
  • Progressive disclosure - Expose only a short summary up front and load detailed material into context only when the task needs it.
  • When to delegate - Delegate when only the result matters to the main thread; keep work in one context when its intermediate steps matter.

Tool

  • Claude Code GitHub Actions - The anthropics/claude-code-action workflow step that runs Claude Code inside a GitHub Actions job, triggered by @claude mentions or a fixed prompt.
  • Claude Projects - In the September 2026 redesign, a Claude project is one long-running conversation whose coordinator splits a goal into parallel threads sharing memory and a library.
  • Pull request auto-fix - A Claude Code cloud feature that watches a pull request and responds to CI failures and review comments, with known blind spots.

Service

  • Claude GitHub App - The GitHub App that gives Claude features repository access, and which features depend on it rather than on other sign-in methods.
  • Claude Managed Agents - An Anthropic-hosted agent harness that runs the agent loop, sandbox, and tools for long-running tasks, driven by events instead of your own runtime.
  • Cloud session - A Claude Code session that runs on an Anthropic-managed VM instead of your machine, cloning your repository from GitHub and running after you disconnect.

Configuration

  • Cloud environment - The saved configuration that sets network access, environment variables, and setup scripts for Claude Code cloud sessions.
  • Skill configuration - The SKILL.md frontmatter fields and directory layout that define a Claude Code agent skill.
  • Subagent configuration file - The Markdown file with YAML frontmatter that defines a custom Claude Code subagent, and how to create it with /agents.

Command

Source Summary

Graph View