progressive-disclosure

Progressive Disclosure

Overview

A general methodology for organizing any document an LLM consumes — CLAUDE.md, AGENTS.md, system prompts, runbooks, or any instruction file loaded into context. The core principle: keep high-signal content always loaded, defer low-frequency detail to on-demand files.

Also known as: lazy prompting (Teresa Torres), progressive discovery.

Complementary skill: After optimizing, use /revise-claude-md (from claude-md-management plugin) to capture future learnings into the right tier — it prevents re-bloat by routing new information appropriately.

---

Interactive Workflow

Step 0: Detect Target

Determine which document(s) to optimize:

1. Ask the user which document to optimize, or auto-detect common targets:
   • Project CLAUDE.md: ./CLAUDE.md or ./.claude/CLAUDE.md
   • User CLAUDE.md: ~/.claude/CLAUDE.md
   • AGENTS.md: ./AGENTS.md
   • System prompts or other instruction files the user specifies
2. Show file path and line count

ASK USER: "Found {path} ({N} lines). Optimize this file? Or specify a different document."

Step 1: Back Up

Create a timestamped backup before making any changes:

cp {target} {target}.bak.$(date +%Y%m%d_%H%M%S)

Confirm backup was created.

Step 2: Analyze & Classify

Read the entire file. For each section, apply the classification decision tree:

Question	Yes	No
Used frequently (most sessions)?	Level 1	next question
Severe consequences if violated?	Level 1	next question
Contains code patterns for direct copying?	Level 1 (keep the pattern)	next question
Scoped to specific file paths or conditions?	Level 1.5	next question
Has a clear trigger condition?	Level 2 + trigger in Level 1	next question
Historical or reference material?	Level 2	Consider removing

Level 1.5 depends on document type:

• CLAUDE.md → .claude/rules/ files with paths: frontmatter (Anthropic's native lazy loading)
• AGENTS.md → conditional sections or linked sub-agent docs
• System prompts → conditional includes or modular prompt segments
• Other docs → sub-documents, appendices, or linked references

Present a classification table:

Section	Lines	Tier	Reason
(section name)	(count)	L1 / L1.5 / L2	(brief justification)

Show summary: "X lines staying in L1, Y lines moving to L1.5, Z lines moving to L2"

ASK USER: "Does this classification look right? Want to adjust any rows before I proceed?"

Step 3: Propose Level 1.5 Moves

For each section classified as Level 1.5, propose the destination based on document type:

For CLAUDE.md — .claude/rules/ files:

File: .claude/rules/{domain}.md
---
paths:
  - "{matching-glob-pattern}"
---
{content to move}

For AGENTS.md — conditional sub-documents:

File: docs/agents/{topic}.md
Activation: When agent handles {specific task type}

For other documents — linked sub-documents:

File: {doc-root}/{topic}.md
Include condition: {when this section is relevant}

ASK USER: "Create these files? Approve, adjust, or skip any."

Step 4: Propose Level 2 Moves

For each section classified as Level 2, show the proposed reference file:

File: {references-dir}/{topic}-reference.md
Trigger condition: {when this file should be loaded}
Content summary: {keywords describing what's inside}

References directory depends on document type:

• CLAUDE.md → docs/references/
• User-level CLAUDE.md → ~/.claude/references/
• Other documents → {doc-root}/references/ or co-located

ASK USER: "Create these reference files? Approve, adjust, or skip any."

Step 5: Generate Structural Additions

Preview each structural element to add to the optimized document:

1. Information Recording Principles (near top) — self-governing rules that prevent re-bloat by routing new information to the correct tier
2. Reference Index (top section) — problem-triggered lookup: "I hit an issue — where do I look?"
3. Pre-Modification Checklist (middle section) — proactive lookup: "I'm about to change X — what should I watch for?"
4. Reference Trigger Index (bottom section) — reminder during long conversations (addresses the "lost in the middle" attention curve)

Show a preview of each element populated with the document's actual references.

ASK USER: "Add all structural elements, or skip any?"

Step 6: Apply Changes

Execute the approved changes:

1. Create Level 1.5 files (.claude/rules/, sub-documents, etc.)
2. Create Level 2 reference files
3. Rewrite the target document: remove moved sections, add approved structural elements
4. Show a diff summary of what changed

Step 7: Verify

Run verification checks:

• Confirm all referenced files exist on disk
• Report: "Before: {N} lines. After: {M} lines. {K} files at Level 1.5, {J} files at Level 2."

Step 8: Next Steps

• For CLAUDE.md: use /revise-claude-md (from claude-md-management plugin) to capture future learnings — it routes new information to the correct tier automatically
• The Information Recording Principles section makes the LLM self-governing: when asked to "remember this," it routes to L1, L1.5, or L2 based on frequency and scope
• Re-run /progressive-disclosure periodically if the file grows past the target line count again

---

Progressive Disclosure Principles (Reference)

The content below serves as background reference. It activates when trigger conditions match outside of explicit /progressive-disclosure invocation.

Core Philosophy

"Find the smallest set of high-signal tokens that maximizes the probability of the desired outcome." — Anthropic

The goal is to maximize LLM working efficiency, not to minimize line count.

Documents accumulate knowledge — SOPs, diagnostics, code patterns, edge cases — that often pushes past manageable sizes. Progressive disclosure solves this tension: keep high-signal content in Level 1 (always loaded) and move detailed reference material to Level 2 (loaded on demand when triggered).

Three-Tier Architecture

Level 1 (Primary Document) — Loaded every session
+-- Information Recording Principles   <- Self-governing rules to prevent future bloat
+-- Reference Index (top)              <- Entry 1: "I hit an error -- where do I look?"
+-- Core content (commands, rules, patterns)
+-- Pre-modification checklist          <- Entry 2: "I'm about to change X -- watch for what?"
+-- Reference Trigger Index (bottom)    <- Entry 3: reminder after long conversations

Level 1.5 (Conditional Loading) — Loaded when conditions match
  For CLAUDE.md:  .claude/rules/ with paths: frontmatter (Anthropic's native lazy loading)
  For AGENTS.md:  conditional sub-agent docs
  For other docs: modular sections, conditional includes

Level 2 (On-Demand) — Loaded via explicit trigger
+-- Detailed SOP workflows
+-- Edge case handling
+-- Full configuration examples
+-- Historical decision records

Level 1.5 is the most underused mechanism. For CLAUDE.md, .claude/rules/ files with paths: frontmatter load automatically when the LLM touches matching files — zero token cost otherwise. For other document types, equivalent conditional-loading mechanisms exist.

The Multi-Entry Principle

A single Level 2 resource can have multiple entry points that serve different lookup paths:

Entry Point	Position	Trigger Scenario	User Mindset
Reference Index	Top	Hit an error or problem	"Something broke — which doc should I read?"
Pre-modification checklist	Middle	About to change code	"I'm changing X — what pitfalls exist?"
Reference Trigger Index	Bottom	Orientation during long conversation	"What was that reference doc again?"

This is not duplication — it's multiple entry points. Like a book having a table of contents (by chapter), an index (by keyword), and a quick-reference card (by task).

The reason this matters: LLM attention follows a U-shaped curve — the beginning and end of context receive stronger attention than the middle (the "lost in the middle" phenomenon). Placing trigger indexes at both top and bottom ensures they're noticed regardless of conversation length.

Classification Decision Tree

For each section, ask these questions in order:

Question	Yes	No
Used frequently?	Level 1	next question
Severe consequences if violated?	Level 1	next question
Contains code patterns that need direct copying?	Level 1 (keep the pattern)	next question
Scoped to specific file paths or conditions?	Level 1.5	next question
Has a clear trigger condition?	Level 2 + trigger in Level 1	next question
Historical or reference material?	Level 2	Consider removing

Content Retention Rules (what stays in L1)

Content Type	Reason
Core commands	High-frequency use
Hard rules / prohibitions	Severe consequences if violated — must always be visible
Code patterns	LLM needs to copy directly; avoids re-derivation
Error diagnostics	Complete symptom -> cause -> fix flow
Directory/structure map	Helps LLM locate files quickly
Trigger index tables	Helps LLM find Level 2 during long conversations

What Moves to Level 2

Content Type	Level 1 keeps	Level 2 gets
SOP workflows	Trigger condition + key pitfalls	Full step-by-step
Config examples	The 1-2 most common	Complete configuration
API documentation	Common method signatures	Full parameter reference
Historical decisions	Nothing (or a one-liner)	Full rationale
Performance data	Nothing	Full benchmarks
Edge cases	Nothing	Detailed handling

Reference Formats (4 varieties)

Use a mix of these formats — variety helps the LLM distinguish different types of references.

1. Detailed Format (important in-body references):

**When to read `{references-dir}/xxx-sop.md`**:
- [specific error message]
- [specific scenario]

> Contains: [keyword 1], [keyword 2], [code template].

2. Problem Trigger Table (top/bottom index):

## Reference Index
| Trigger | Document | Key Content |
|---|---|---|
| `ERR_DLOPEN_FAILED` | `native-modules-sop.md` | ABI mechanism, lazy loading |

3. Task Trigger Table (pre-modification checklist):

## Pre-Modification Checklist
| What you're changing | Read this first | Key pitfalls |
|---|---|---|
| Native module code | `native-modules-sop.md` | Must lazy-load |

4. Inline Format (brief references):

Full workflow in `database-sop.md` (FTS5 escaping, health checks).

Anti-Patterns

Anti-Pattern	Example	Fix
Over-Compression	2,937 lines compressed to 165	Keep all high-frequency content; 482 organized lines beats 165 stripped lines
References Without Triggers	See xxx.md	Always pair with trigger condition + content summary
Code Patterns in Level 2	Moving frequently-used code examples to reference files	High-frequency code patterns stay in Level 1
Deleting Instead of Moving	Removing "unimportant" sections entirely	Move to Level 2 with a trigger; never delete knowledge

Information Density Checks

Check	Passing Criteria
Daily commands	No need to read Level 2
Common errors	Has complete diagnostic flow
Code writing	Has copy-paste patterns
Specific problems	Knows which Level 2 to read
Trigger indexes	Table format at top and bottom

Too little: LLM repeatedly asks same questions, re-derives patterns, user corrects same rules. Too much: Low-frequency workflows in Level 1, identical content duplicated (not multi-entry), edge cases mixed with common cases.

Document-Specific Guidelines

Dimension	User-Level CLAUDE.md	Project-Level CLAUDE.md	AGENTS.md	Other Docs
Target L1 lines	100-200	200-400	150-300	Varies
L1.5 mechanism	~/.claude/rules/	.claude/rules/	Sub-agent docs	Conditional includes
L2 location	~/.claude/references/	docs/references/	docs/agents/	Co-located references/
Shares with	Just you	Team via source control	Team	Depends on doc

Quick Checklist

After optimizing, verify:

• "Information Recording Principles" near the top (prevents future bloat)
• Reference Index near the top (Entry 1: problem-triggered lookup)
• Core content is complete (commands, rules, patterns)
• Hard rules / prohibitions include WRONG/RIGHT examples (not just prose)
• Common errors have complete diagnostic flow (symptom -> cause -> fix)
• Code patterns are copy-paste ready
• Level 1.5 files created for path-scoped or conditional content
• "Pre-Modification Checklist" table (Entry 2: indexed by "what you're changing")
• Reference Trigger Index near the bottom (Entry 3: reminder in long conversations)
• Every Level 2 reference has a trigger condition
• All referenced files exist on disk

For detailed case studies and lessons learned, see references/progressive-disclosure-principles.md.