Meta Ways¶
The meta domain holds guidance about the system itself and about how Claude works with the human: how ways and skills are written, how work is delegated, how state survives compaction, and how trust is handled. This page lists the top-level meta ways. The way files under hooks/ways/meta/ are the source; ways session ways shows which fired in a session, and ways author graph exports the whole tree.
| Way | Covers | Fires on |
|---|---|---|
attend/* |
Responding to attend signals: a finished build, context pressure, an overdue reflection | attend signals |
choices |
Putting real decisions to the human as explicit choices | prompt pattern ("which option", "let me decide") |
compaction-checkpoint |
Summarizing and checking in with the human before compaction | context 85% full |
deployment |
How agent-ways installs, updates and reconciles into ~/.claude |
prompt pattern (agent-ways, ways update, ~/.claude) |
develop |
Routing a piece of work through the development loop and the /develop skill |
prompt pattern ("how should we approach") |
goals |
When to set a /goal and what goal mode changes |
semantic |
governance |
Citing the real controls behind a practice | semantic |
introspection |
Looking back over the session at pull-request time for guidance worth keeping | gh pr create, prompt pattern |
knowledge |
The ways system: ways, skills and hooks, domains, matching. Children cover authoring, frontmatter, refire, trees, the keyword lane, locale stubs and vocabulary tuning. | prompt pattern ("ways"), editing way files |
memory |
The persistent memory files: what to record and when | context 80% full, "save to memory", editing memory files |
skills |
House conventions for writing Claude Code skills | prompt pattern (SKILL.md, "author a skill") |
start |
Recognizing session start and routing to /start |
prompt pattern ("pick up where we left off") |
subagents |
When to delegate to a sub-agent, writing the brief, what comes back | prompt pattern ("subagent", "delegate") |
think |
Structured reasoning for hard decisions and the /think skill |
prompt pattern ("trade-off", "I'm stuck") |
todos |
Capturing unfinished work as tasks before compaction | context 75% full with no task list |
trust/* |
Trust between Claude and the human: earned autonomy, acting through the human's accounts, whose voice to write in, long-form prose | semantic, prompt pattern |
workflows |
When to reach for the Workflow tool for multi-agent orchestration | prompt pattern ("fan out", "pipeline") |
wrap |
Recognizing session end and routing to /wrap |
prompt pattern ("wrap up") |
Ways and skills¶
The knowledge way draws the line between the two:
| Skills | Ways | |
|---|---|---|
| Discovery | Semantic (Claude decides) | Triggered (patterns, tools, state) |
| Activation | Claude matches user intent to the description | A hook event fires and a trigger matches |
| Use case | Specialized procedures | Workflow guardrails, conventions |
| Can detect | User intent | Tool execution, file edits, session state |
They complement each other. Skills handle "the user wants to do X". Ways handle "Claude is about to do Y". A skill cannot see that git commit is about to run; a way cannot tell that a vague request is really about API design.
Sub-agents¶
The subagents way covers delegating to the agents shipped in agents/:
| Agent | Purpose |
|---|---|
| requirements-analyst | Capture requirements as GitHub issues |
| system-architect | Draft ADRs and evaluate design trade-offs |
| task-planner | Plan multi-branch implementations |
| code-reviewer | Review PRs for quality and SOLID compliance |
| workflow-orchestrator | Coordinate the ADR-driven workflow |
| workspace-curator | Organize docs/ and .claude/ |
| skeptic | Try to refute a finished deliverable's claims from primary sources, read-only |
Sub-agents are for delegating token-heavy work, such as reviewing a long diff with fresh context, not for every action.
Team coordination for teammates is the collaboration/teams way; see teams.md.