%%{init: {'theme': 'dark', 'themeVariables': {'primaryColor': '#14142a', 'primaryTextColor': '#d0d0e0', 'primaryBorderColor': '#b44dff', 'lineColor': '#b44dff', 'secondaryColor': '#0c0c18', 'tertiaryColor': '#080810', 'edgeLabelBackground': '#080810'}}}%%
flowchart TD
RUN[Session running]:::step -->|context fills up| PRE[PreCompact fires]:::hook
PRE --> SAVE[Saves plan, task, decisions]:::step
SAVE --> COMPACT[Auto-compaction happens]:::event
COMPACT --> POST[SessionStart fires]:::hook
POST --> RESTORE[Restores context summary]:::step
RESTORE --> RESUME[Claude knows where it left off]:::approved
classDef step fill:#14142a,stroke:#b44dff,color:#b44dff,stroke-width:2px
classDef hook fill:#14142a,stroke:#00f0ff,color:#00f0ff,stroke-width:2px
classDef event fill:#14142a,stroke:#ff2d7b,color:#ff2d7b,stroke-width:2px
classDef approved fill:#14142a,stroke:#00cc66,color:#00cc66,stroke-width:2px
Customization Guide
Make the Clo-Author Yours
Overview
The Clo-Author works out of the box, but it’s designed to be customized. Five configuration layers control how Claude behaves in your project — from the high-level project constitution to low-level event hooks.
| Layer | Location | Loaded When |
|---|---|---|
| CLAUDE.md | Project root | Every session |
| Rules | .claude/rules/ |
Always-on or path-scoped |
| Skills | .claude/skills/ |
On demand (slash commands) |
| Agents | .claude/agents/ |
On demand (via skills or orchestrator) |
| Hooks | .claude/settings.json |
On events (automatic) |
Claude reliably follows about 100–150 custom instructions. Your system prompt uses ~50, leaving ~100 for your project. CLAUDE.md and always-on rules share this budget. Path-scoped rules, skills, and agents load on demand — they don’t count against the limit.
CLAUDE.md — The Project Constitution
CLAUDE.md is the single most important file. Claude reads it at the start of every session. It should be a slim constitution — short directives and pointers, not comprehensive documentation. Aim for ~120 lines:
- Core principles — 5–6 bullets (plan-first, verify-after, quality gates, worker-critic pairs, LEARN tags)
- Folder structure — where everything lives
- Commands — compilation, deployment
- Skill quick reference — table of 10 available commands
- Customization tables — Beamer environments, notation
- Current state — what’s done, what’s in progress
If CLAUDE.md exceeds ~150 lines, Claude starts ignoring rules silently. Put detailed standards in path-scoped rules instead.
Rules — Auto-Loading Domain Knowledge
Rules are markdown files in .claude/rules/ that Claude loads automatically. Consolidated into 8 focused files for clarity.
| Rule | Purpose |
|---|---|
workflow.md |
Plan-first protocol, orchestrator loop, dependency graph, standalone access |
quality.md |
Weighted scoring formula, severity gradient by phase, quality gates |
agents.md |
Worker-critic pairs, separation of powers, three-strikes escalation |
logging.md |
Session logging, session reporting, research journal |
content-standards.md |
Table generation, PDF processing, exploration workflow |
revision.md |
R&R cycle comment classification and routing |
meta-governance.md |
Template vs working project distinction |
domain-profile.md |
Field-specific calibration (user-customized) |
Domain Profile
The domain profile (.claude/rules/domain-profile.md) calibrates all agents to your specific research field. It can be populated two ways:
/discover(interactive interview) — generates it automatically- Manual editing — fill in the template directly
All agents read this file for:
- Target journal tiers (for Orchestrator and referees)
- Common data sources (for Explorer)
- Identification strategies (for Strategist and strategist-critic)
- Field conventions (for Coder and Writer)
- Seminal references (for Librarian)
- Referee concerns (for domain-referee and methods-referee)
Journal Profiles
The journal profiles (.claude/rules/journal-profiles.md) calibrate the domain-referee and methods-referee to a specific journal’s review culture when you run /review --peer [journal].
15 journals are pre-populated (AER, Econometrica, JPE, QJE, REStud, AEJ:Applied, AEJ:Policy, JHR, JHE, RAND, JPubE, JLE, JDE, RESTAT, AER:Insights). Each profile describes the journal’s bar, what its referees prioritize, and typical concerns — in plain language that the LLM interprets naturally.
For journals not in the list, referees still adapt using the journal name + your domain profile. You can add your own profiles using the template at the bottom of the file.
Skills — Available Commands
Skills are multi-step workflows invoked with /command. Each skill lives in .claude/skills/[name]/SKILL.md and dispatches the appropriate agents.
There are 10 skills covering the full research pipeline: /new-project, /discover, /strategize, /analyze, /write, /review, /revise, /talk, /submit, and /tools.
Skills can be invoked two ways: explicitly (you type /review Paper/main.tex) or automatically (Claude invokes them when relevant). Most of the time, you just describe what you want.
See the Command Reference for every command, flag, and subcommand. See the User Guide for detailed workflow descriptions.
Hooks — Smart Defaults
Hooks live in .claude/settings.json and fire on events, regardless of context state. Reduced to 4 essential hooks that stay out of the way.
Design Principles
- Use command-based hooks for fast, mechanical checks
- Use rules for nuanced judgment
- Avoid prompt-based hooks that trigger LLM calls on every response
All 4 Hooks
| Hook | Event | What It Does |
|---|---|---|
| File protection | PreToolUse |
Blocks edits to bibliography and settings |
| Context state capture | PreCompact |
Saves plan state + context survival checklist |
| Context restoration | SessionStart[compact\|resume] |
Restores context after compaction |
| Post-merge reminder | Git post-merge | Reminds to run /tools learn after merges |
Memory System
Four persistent layers plus ephemeral conversation:
| Layer | File | Updated When | Purpose |
|---|---|---|---|
| Project context | CLAUDE.md |
Rarely | Project rules, folder structure |
| Corrections | MEMORY.md |
On [LEARN] tag |
Prevent repeating mistakes |
| Task strategy | quality_reports/plans/ |
Once per task | Plan survives handoff |
| Decision reasoning | quality_reports/session_logs/ |
Incrementally | Record why decisions were made |
| Conversation | Context window | Every response | Current working memory (not persistent) |
Context Survival System
When context compaction happens, two hooks preserve and restore state: