/project-state

Domain expert for the CLAUDE.md orientation layer. Reads frontmatter (routing contract: Linear IDs, knowledge intake, description, status) and the Re-entry Cue (the one-sentence heads-up for session resumption). Writes the Re-entry Cue and the updated timestamp — nothing else.

Session-level memory (current state, waiting-for, decisions needed) lives in Linear, not in CLAUDE.md. This skill does not read or write those fields.

Identity

This skill owns the structured orientation that lives in a project's CLAUDE.md — the thinnest layer of the three-layer memory model (CLAUDE.md = orientation, Linear issues = item-level, Linear Project Updates = session-level).

Discipline rules:

• Re-entry Cue is ONE sentence, only when work is mid-flight. If the last session closed cleanly, the cue is "No work in progress" or absent. Not a recap; a heads-up.
• Three-layer separation: item-level decisions go on the issue; session-level narrative goes in Project Updates; orientation + re-entry goes here. Nothing else.
• Routing fields live in frontmatter. linear_project_id, linear_url, knowledge_intake, description, status, updated — all parsed from frontmatter, never from body text.

Intent

Objective. Parse and write the CLAUDE.md orientation layer so consumers (session-start, session-closeout, new-project) get structured fields without re-implementing parsing. Enforce Re-entry Cue format.

Health Metrics:

• Re-entry Cue one-sentence invariant (validated pre-write; fail on violation)
• Frontmatter fields parsed cleanly — missing required fields surfaced as data errors, never silently null

Decision Authority:

• Autonomous: all read parsing; Re-entry Cue validation; frontmatter updated writes
• Escalate: CLAUDE.md missing → return null with error; hub type → return minimal structure; missing linear_project_id → data error; missing description → warning

Stop Rules:

• CLAUDE.md not found → fail with clear error
• Re-entry Cue multi-sentence on write → fail; caller collapses before re-invoking
• Hub when caller expected project → return type:hub; caller decides

Navigation

Operation	Input	Output	Playbook
read	project_root (abs path) OR "cwd" OR project name	Structured: type, description, status, last_updated, re_entry_cue, linear_project_id, linear_project_url, knowledge_layer_declared, knowledge_index_path	playbooks/read.md
write	project_root + last_updated + re_entry_cue	Mutation: frontmatter updated + Re-entry Cue body; fields_updated list returned	playbooks/write.md

Cross-cutting

Project resolution. When input is "cwd", walk up from cwd until finding a CLAUDE.md whose frontmatter tags contain type/claude-project or type/claude-hub. When input is a project name, append to workspace_root (global CLAUDE.md > Configuration). When input is an absolute path, use directly. Fail if no CLAUDE.md exists.

Hub vs. project. Frontmatter tag type/claude-hub → return type: hub with minimal structure; caller decides whether to recurse.

Frontmatter is the routing contract. linear_project_id missing in a non-hub is a data error. description missing is a warning (project is invisible to routing). knowledge_intake absent defaults to false.

What this skill does NOT do

• Does NOT write to Linear (that's /linear)
• Does NOT scan Knowledge layer (that's /knowledge-layer)
• Does NOT compose the Re-entry Cue's content — the caller writes it; this skill enforces format
• Does NOT read or write Current State, Waiting For, or Decisions Needed — those live in Linear
• Does NOT manage Intake sections — routing fields are in frontmatter; /new-project creates the initial frontmatter

References

• Project template: path configured in global CLAUDE.md > Configuration > templates.project
• [[linear-discipline]] — three-layer memory model