ITI Compress

Toggle: compress to activate  ·  verbose to deactivate

Purpose

Claude Code and Cursor sessions in the ITI monorepo are token-expensive by nature — 2,800-line playbook, 284-entry .agents/skills/, n8n workflows, and CLAUDE.md files across 26+ products in scope on every session open.

This skill strips all token waste from conversational output while leaving code, file content, and deliverable artifacts completely unchanged. Goal: same intelligence, 40–60% fewer commentary tokens per session.

Scope — What Gets Compressed

Compression applies only to the words the agent says between tool calls and deliverables. It never touches:

• Code blocks (PHP, TypeScript, Python, Rust, Swift, SQL, shell)
• File content being written or edited
• CLAUDE.md, SKILL.md, requirements docs, or any .md artifact being produced as output
• Commit messages — must follow ITI type(scope): description format regardless
• Error messages and audit output — full sentences, they must be actionable
• Any output for a GHI (HIPAA) or Emily Ward (attorney-client) context

Compression Rules

Do

• Result first. No preamble.
• Drop articles (“the”, “a”, “an”)
• Drop filler (“Sure!”, “Great question!”, “I’d be happy to”, “Let me know if”)
• Drop self-narration (“I found”, “I searched”, “Let me”, “I’ll now”)
• Drop hedging (“I think”, “perhaps”, “it seems”)
• Drop transitions (“Furthermore”, “Additionally”, “Moving on”)
• Drop question restatement and closing summaries
• Fragments valid: “Works. Fast. Done.”
• Symbols over words: “→” not “leads to”, “&” not “and”, “3” not “three”
• Never announce tool use before or after — just do it, show result, stop

Never Compress

• Code blocks (any language)
• File content being written or edited
• CLAUDE.md, SKILL.md, build plans, requirements docs — any .md artifact produced as output
• ITI commit messages — always type(scope): description
• Error messages and AUT-01..11 audit findings — must be full sentences
• [CONTEXT-UPDATE] flags — reproduce the structured template exactly
• Any response involving GHI (HIPAA), Emily Ward (attorney-client), or financial data

File paths are always full. ITI/products/scuba-gpt/includes/class-api.php is data, not prose — never truncate.

ITI-Specific Rules

Rule	Reason
Always show full file paths in commentary	Paths are data — ITI/products/scuba-gpt/includes/class-api.php must be exact
Never compress deliverable artifacts	SKILL.md bodies, CLAUDE.md updates, build plans are outputs, not commentary
Never compress security findings	AUT-06 output, HIPAA scope, attorney-client work must use full sentences
Commit messages: always ITI format	feat(product): description regardless of compression state
File/tool counts: use numerals	“14 files changed” not “fourteen files”
[CONTEXT-UPDATE] flags: always full format	Structured artifact — template must be reproduced exactly

Toggle Behavior

Trigger	Effect
"compress"	Activate for remainder of session
"verbose"	Deactivate — return to normal prose
Detected HIPAA/legal context	Auto-deactivate for that response only; re-activate after
“explain in detail” or “walk me through”	Full sentences for that response only

Compression state persists for the session. No need to re-invoke between messages.

Examples

Recommended Usage Patterns

Long context sessions

"compress" → [load CLAUDE.md, start work]

Heavy refactor sessions

"compress" → "Migrate all AI clients in products/ to extend shared library (B12)"

Back to normal for deliverables

"verbose" → "Now write the updated CLAUDE.md for scuba-gpt"

One-off detail request without toggling

"explain in detail how the ITI_Workflow_Adapter fallback works"
→ [full explanation for this response only, then returns to compressed]

Where This Skill Fits

Attribute	Value
Storage	Personal (~/.cursor/skills/) — not project-scoped
AUT-01 sync target	No — personal utility, not part of .agents/skills/
CLAUDE.md mention	Not required — invoke by name as needed
Related skills	session-context-protocol, scope-control — both token-heavy; pair with compress for long sessions