Interscribe

Documentation quality toolkit with three modes: audit, refactor, consolidate.

The Boundary Rule

This is the single most important convention interscribe enforces:

CLAUDE.md    → Claude Code ONLY: plugins, hooks, tool preferences, permissions, Claude-specific settings
AGENTS.md    → Everything else: architecture, workflows, conventions, troubleshooting, standards
  → docs/    → Deep reference files linked from AGENTS.md sections

Any project knowledge in CLAUDE.md is a boundary violation. Project knowledge includes: git workflow, architecture, build instructions, coding conventions, server setup, debugging guides, workflow patterns.

What belongs in CLAUDE.md: Plugin versions and settings, hook configuration, tool preferences (Read vs cat), permission notes, Claude-specific behavioral overrides, links to AGENTS.md.

Input

<interscribe_input> #$ARGUMENTS </interscribe_input>

Parse the input to determine mode and target:

• audit / audit [path] → Audit mode
• refactor / refactor [path] → Refactor mode
• consolidate / consolidate [path] → Consolidate mode
• No argument → Audit mode (default)
• Just a path → Audit that path

Mode 1: Audit

Analyze documentation health and report findings. No changes made.

Step 1: Discover Docs

Find all CLAUDE.md and AGENTS.md files in the target path (default: project root). Also find docs/canon/doc-structure.md if it exists.

Step 2: Boundary Check

For each CLAUDE.md, classify every section:

• Claude Code config — belongs here (plugins, hooks, tool prefs, permissions)
• Project knowledge — boundary violation (architecture, workflows, conventions, git workflow, build, troubleshooting)
• Pointer — links to AGENTS.md sections (good)

Report violations with the section heading and line count.

Step 3: Token Budget

Count lines in each doc. Report:

• CLAUDE.md line count (target: 30-60 lines)
• AGENTS.md line count (no hard cap, but flag sections >100 lines as candidates for extraction to docs/)
• Total context load at session start (sum of auto-loaded docs in the hierarchy)

Step 4: Duplication Check

Look for semantically identical instructions across:

• Global CLAUDE.md ↔ project CLAUDE.md
• CLAUDE.md ↔ AGENTS.md in the same project
• Parent AGENTS.md ↔ subproject AGENTS.md

Report duplicates with locations.

Step 5: Staleness Check

Scan for:

• References to files that don't exist
• References to directories that don't exist
• Version numbers that don't match actual versions
• Instructions about tools/features that are no longer present

Step 6: Report

Output a summary:

## Interscribe Audit Report

**Target:** [path]
**CLAUDE.md:** [N] lines ([M] boundary violations)
**AGENTS.md:** [N] lines ([M] sections > 100 lines)
**Duplicates:** [N] found
**Stale references:** [N] found

### Boundary Violations
[list with section names and line counts]

### Oversized Sections
[AGENTS.md sections that should be extracted to docs/]

### Duplicates
[instruction, locations]

### Stale References
[reference, file, line]

**Health Score:** [A-F] based on violation count and severity

Mode 2: Refactor

Restructure docs for progressive disclosure. Fully automatic — applies all changes. Review via git diff.

Step 1: Audit First

Run the full audit (Mode 1) to understand current state.

Step 2: Boundary Enforcement

For each boundary violation in CLAUDE.md:

1. Find or create the appropriate section in AGENTS.md
2. Move the content from CLAUDE.md → AGENTS.md
3. If the content was a standalone topic, add a pointer in CLAUDE.md: See [Topic](AGENTS.md#section)
4. If the content was inline in a larger section, just remove it from CLAUDE.md

Step 3: AGENTS.md Extraction

For AGENTS.md sections exceeding 100 lines:

1. Create a reference file at docs/[topic].md with the full content
2. Replace the AGENTS.md section with a summary (3-5 lines) + pointer link
3. Preserve the section heading in AGENTS.md for discoverability

Step 4: Deduplication

For each duplicate found:

1. Determine the canonical location (highest common ancestor in the hierarchy)
2. Keep the canonical copy
3. Replace other copies with a pointer or remove entirely
4. If there's a contradiction, keep both and add a <!-- CONTRADICTION: also stated differently in [file] --> comment

Step 5: Stale Cleanup

Remove or fix stale references:

• Dead file links → remove the link (keep the instruction if still valid)
• Dead directory references → update or remove
• Outdated version numbers → update if determinable, flag if not

Step 6: Summary

Report what was changed:

## Interscribe Refactor Summary

**CLAUDE.md:** [before] → [after] lines
**AGENTS.md:** [before] → [after] lines
**Files created:** [list of new docs/ reference files]
**Duplicates removed:** [count]
**Stale references fixed:** [count]

Review changes: `git diff`

Mode 3: Consolidate

Deduplicate across a multi-level doc hierarchy. Use when you have global + project + subproject docs.

Step 1: Map Hierarchy

Build the doc loading tree:

~/.claude/CLAUDE.md (global)
  → project/CLAUDE.md
    → project/subdir/CLAUDE.md
~/.codex/AGENTS.md (global, if exists)
  → project/AGENTS.md
    → project/subdir/AGENTS.md

Step 2: Cross-Level Duplication

For each instruction that appears at multiple levels:

• Identical → keep at highest level, remove from lower levels
• Specialized (lower level adds project-specific detail) → keep both, verify no contradiction
• Contradictory → flag for manual resolution, do not auto-fix contradictions across hierarchy levels

Step 3: Apply and Report

Apply deduplication changes. Report what was consolidated and any contradictions flagged.

Health Score Rubric

Grade	Criteria
A	0 boundary violations, CLAUDE.md <60 lines, 0 duplicates, 0 stale refs
B	1-2 boundary violations OR CLAUDE.md 60-80 lines OR 1-2 duplicates
C	3-5 boundary violations OR CLAUDE.md 80-120 lines OR 3-5 duplicates
D	6+ boundary violations OR CLAUDE.md >120 lines OR 6+ duplicates
F	AGENTS.md missing entirely OR CLAUDE.md contains >50% project knowledge

Conventions

docs/canon/

Foundational project docs that define identity and standards:

docs/canon/
├── PHILOSOPHY.md      # Design bets and tradeoffs
├── doc-structure.md   # This standard (pointer-doc convention, budgets, boundaries)
└── naming.md          # Naming conventions

Root keeps: CLAUDE.md, AGENTS.md (operational, auto-loaded by tools).

Line Budgets

File	Target	Hard Cap	Action
CLAUDE.md	30-60	80	Move project knowledge → AGENTS.md
AGENTS.md section	50-80	100	Extract → docs/ reference file
MEMORY.md	100-150	200	Promote stable facts → AGENTS.md via intermem:synthesize