Skip to content

Welcome to Wren

New here? Open your project in Claude Code and run:

/setup wren

That starts a guided onboarding — a short tour, a few questions about your research, the right tools set up to match, and your first paper — so you leave with real momentum instead of a blank page. This page is the same tour in written form; read it here or let /setup wren walk you through it.

Wren is an agentic research assistant for Claude Code. You drive everything from the terminal by asking; there is nothing else to open, and nothing external is required to start.

1. A memory that persists

Wren keeps context in a four-layer memory model, each layer with one home so nothing lands in the wrong place:

  1. Behaviour rules (.claude/rules/) — always-on guidance the assistant follows.
  2. Cross-session context (~/.claude/projects/.../memory/) — who you are, durable preferences, and dated learnings, loaded at the start of every session.
  3. Corpus memory (research_memory/, optional add-on) — findings across your literature and projects, queryable with /research-memory.
  4. Per-paper deep work (papers/{id}/) — each paper's status, notes, and methods log.

The rule of thumb: paper and project state lives with the paper, never in general memory. Full detail in Concepts → the four-layer memory model.

2. The 14-phase pipeline and the gold-standard gate

Every paper moves through 14 phases, tracked in its STATUS.md (phase: N), from scoping through data, analysis, drafting, review, and publication.

Phase 7 is the gate: the gold-standard freeze. Before you write a word of prose, the analysis is verified, git-tagged, and archived as canonical. That is the discipline that stops you drafting on numbers that later change — run /gold-standard to do it. See the full pipeline diagram in Concepts → the 14-phase pipeline.

3. The skills

Type these as slash commands once your project is open. All 18 ship in the box — full reference.

  • Orient & manage — /research, /paper, /setup
  • Write & submit — /draft, /cover-letter, /revision-response, /peer-review
  • Discover — /literature-scan
  • Rigor & memory — /gold-standard, /rigor-check, /research-memory, /retro

4. Audit, QC, and peer review

The rigor gates that keep the research honest:

  • /gold-standard — freeze and verify an analysis as canonical (the Phase 7 gate).
  • /peer-review — an agentic reviewer that evaluates a draft against its target journal's guidelines in four stages, and scores it.
  • /rigor-check — a mid-session self-audit that catches drift, shortcuts, and unjustified assumptions during long working sessions.

Two more are planned, not yet shipped: /qc-team (adversarial QC of a draft, phases 6 and 11) and /figure (figure design and critique, phase 9). Until they land, run those phases manually — see the Roadmap.

5. Pluggable tools

Beyond Claude Code itself, nothing in the core needs a database, a reference manager, or any other account. You choose the backends per project in research-config.yml, and /setup wren sets them up to match your data:

  • Database — files (default, zero setup) → sqlite/duckdb (local, structured) → directus-local (Docker, admin UI + API) → managed (any host).
  • References — bibtex (default, universal) or zotero (live, via MCP).

Switching later is a config change plus ./setup.sh. Detail in Databases and References & integrations.

Where to go next

  • Run /setup wren if you haven't — it does everything above and sets up your first paper.
  • Already set up? Run /paper to start or resume a paper, or /research to see the whole workspace.
  • Prefer to read first? Continue to the Getting Started walkthrough.