gregoryho/arc-reflecting

Diary Reflection & Pattern Extraction

Overview

Analyze multiple diary entries to identify recurring patterns. Save insights to ~/.arcforge/diaryed/ for user review.

Quick Reference

| Task | Command | |------|---------| | Determine strategy | node "${SKILL_ROOT}/scripts/reflect.js" strategy --project {project} | | Scan diaries | node "${SKILL_ROOT}/scripts/reflect.js" scan --project {p} --strategy {s} | | Update log | node "${SKILL_ROOT}/scripts/reflect.js" update-log --project {p} --diaries "{f}" --reflection "{id}" | | Pattern threshold | 3+ occurrences = Pattern, 1-2 = Observation | | Rule violations | Check CLAUDE.md first, report violations with evidence | | Save instinct | node "${SKILL_ROOT}/scripts/reflect.js" save-instinct --project {p} --id {id} --trigger "..." --action "..." [--domain D] [--evidence "..."] [--evidence-count N] | | Strategy modes | unprocessed (5+ new) | project_focused (5+ total) | recent_window (fallback) |

Infrastructure Commands

Node.js utilities handle diary scanning and processed.log management.

Set SKILL_ROOT from skill loader header (# SKILL_ROOT: ...):

: "${SKILL_ROOT:=${ARCFORGE_ROOT:-}/skills/arc-reflecting}"
if [ ! -d "$SKILL_ROOT" ]; then
  echo "ERROR: SKILL_ROOT=$SKILL_ROOT does not exist. Set ARCFORGE_ROOT or SKILL_ROOT manually." >&2
  exit 1
fi

Determine strategy (auto-detect):

node "${SKILL_ROOT}/scripts/reflect.js" strategy --project {project}
# Returns: unprocessed | project_focused | recent_window

Scan for diaries:

node "${SKILL_ROOT}/scripts/reflect.js" scan \
  --project {project} \
  --strategy unprocessed
# Returns: List of unprocessed diary file paths

Update processed.log after reflection:

node "${SKILL_ROOT}/scripts/reflect.js" update-log \
  --project {project} \
  --diaries "diary-1.md,diary-2.md,diary-3.md" \
  --reflection "2026-01-reflection-2.md"

Subagent for Diary Analysis

For large diary sets, use the diary-analyzer subagent (see diary-analyzer.md) to read diaries in an isolated context without polluting the main conversation.

Integration with instincts:

• /reflect → ~/.arcforge/diaryed/ (reflections) + instincts saved via save-instinct
• /recall → retrieves instincts and learned patterns

Core principle: Patterns must appear 3+ times across diary entries to be considered "Pattern". 1-2 occurrences are labeled "Observation".

When to Use

• User runs /reflect
• 5+ diary entries accumulated
• User asks "what have I learned?" or "show me patterns"
• User wants to review preferences across sessions

When NOT to Use

• Fewer than 3 diary entries exist
• User wants patterns auto-loaded (use /recall instead)
• Single-session insights (use /journal instead)
• No meaningful patterns found

Storage

~/.arcforge/diaryed/
├── global/                              # Cross-project patterns
│   ├── processed.log                    # Tracks which diaries were processed
│   └── prefers-explicit-errors.md
└── {project}/                           # Project-specific patterns
    ├── processed.log                    # Per-project tracking (no cross-pollution)
    └── architecture-decisions.md

processed.log format:

# processed.log for {project}
# Format: diary_filename | processed_date | reflection_filename

diary-abc123.md | 2025-01-24 | 2025-01-reflection-1.md
diary-def456.md | 2025-01-24 | 2025-01-reflection-1.md

NOT auto-loaded by Claude. User must manually review.

Process

1. Smart Filter Selection (Auto)

Before reading diaries, determine the optimal filtering strategy:

Strategy Selection Algorithm:
┌─────────────────────────────────────────────────────────────┐
│ 1. Check processed.log for project                         │
│ 2. Count unprocessed diaries                                │
│                                                             │
│ IF unprocessed >= 5:                                        │
│   → Mode: "unprocessed" - analyze only new diaries          │
│ ELIF current_project has 5+ total diaries:                  │
│   → Mode: "project_focused" - analyze project diaries       │
│ ELSE:                                                       │
│   → Mode: "recent_window" - analyze recent 10 diaries       │
└─────────────────────────────────────────────────────────────┘

Output the strategy header at start of reflection:

## Reflect Strategy
**Mode:** {unprocessed|project_focused|recent_window}
**Diaries analyzed:** {count}
**Reason:** {why this mode was selected}
**Projects covered:** {project} (count), ...

2. Locate Diary Entries

Search for diary files:

~/.arcforge/diaries/{project}/*/diary-*.md

Count entries. If fewer than 3:

"Found only X diary entries. Run more sessions with /journal before reflecting."

3. Read CLAUDE.md Rules (if exists)

Before analyzing diaries, read the project's CLAUDE.md to:

• Extract existing rules and conventions
• Enable detection of rule violations in diary entries

This enables the skill to detect when diaries show user corrections for breaking existing rules.

4. Read and Analyze Diaries

Read each diary entry. Look for:

• Repeated decisions
• Consistent preferences
• Recurring challenges (and solutions if marked Generalizable)
• Common techniques
• Rule violations: Cases where user corrected Claude for breaking a CLAUDE.md rule

Observation Cross-Reference: When observations are available (~/.arcforge/observations/{project}/observations.jsonl), cross-reference diary patterns with tool call data for stronger evidence. Tool usage sequences that match diary-reported techniques provide quantitative backing for patterns. For example, if a diary mentions "always grep before editing", check observations for Grep→Read→Edit sequences to confirm frequency.

5. Identify Patterns and Violations

Pattern threshold (3+ occurrences): A pattern MUST appear in 3+ diary entries to be labeled "Pattern".

Observation (1-2 occurrences): Noted but not promoted to pattern status.

For each potential pattern, track:

• Which diary entries contain it
• How it manifested each time
• Whether it's a preference, technique, or decision

For rule violations, track:

• Which CLAUDE.md rule was violated
• Which diary entries show user corrections
• Specific correction quotes as evidence

6. Draft Reflection Output

Use this structure (with strategy header from step 1):

## Reflect Strategy
**Mode:** {mode}
**Diaries analyzed:** {count}
**Reason:** {reason}
**Projects covered:** {list}

---

## Rule Violations Detected (PRIORITY)

### Violation: {rule-name}
**Existing Rule:** "{quoted from CLAUDE.md}"
**Violation Pattern:** User corrected Claude in N sessions
**Evidence:**
- [YYYY-MM-DD] diary-{id}: "{correction quote}"
- [YYYY-MM-DD] diary-{id}: "{correction quote}"
**Suggested Action:** Strengthen rule in CLAUDE.md (user decides)

---

## Patterns Identified (3+ occurrences)

### Pattern: {pattern-name}
**Occurrences:** N sessions
**Evidence:**
- [YYYY-MM-DD] Session {id}: {how it appeared}
- [YYYY-MM-DD] Session {id}: {how it appeared}
**Implication:** {what this suggests}
**Confidence:** High/Medium

---

## Observations (1-2 occurrences)

- {observation}: seen in {N} session(s)

Note: Rule violations appear FIRST (priority) before patterns.

Instincts to Create

<!-- For each pattern found, create an instinct -->

• ID: [pattern-name]
• Trigger: [when does this apply]
• Action: [what to do]
• Domain: [category]
• Evidence count: [how many times seen]

7. Auto-Save and Inform

Reflections and instincts are auto-saved. Inform user of what was saved.

"I found these patterns/violations across X diary entries. Saving reflection and instincts."

For rule violations, additionally inform:

"These rule violations suggest strengthening CLAUDE.md. Would you like to update those rules?"

8. Save Reflections and Instincts

1. Ensure ~/.arcforge/diaryed/{project}/ or ~/.arcforge/diaryed/global/ exists
2. Write the reflection markdown file (e.g., YYYY-MM-reflection-N.md)
3. Update processed.log with each diary that was analyzed:

   diary-abc123.md | 2025-01-24 | 2025-01-reflection-1.md
   

4. For each Pattern, save an instinct:

   node "${SKILL_ROOT}/scripts/reflect.js" save-instinct \
     --project {project} \
     --id {pattern-name} \
     --trigger "{when this applies}" \
     --action "{what to do}" \
     --domain {category} \
     --evidence "{source diary references}" \
     --evidence-count {N}
   

5. Confirm save location, processed.log update, and instincts saved

Key Principles

Evidence-Based Only

Every pattern MUST cite specific diary entries as evidence. No patterns based on general assumptions.

Auto-Save with Notification

Reflections and instincts are auto-saved. Inform user of what was saved.

Non-Prescriptive

Insights are observations, not rules. User decides how to act on them.

Separate from Recall

Diaryed patterns are for reflection. If user wants auto-loading, redirect to /recall.

Common Mistakes

Extracting Patterns from Few Entries

Wrong: "You mentioned preferring TypeScript twice" → calling it a Pattern Right: 1-2 occurrences = Observation, 3+ occurrences = Pattern

Not Citing Evidence

Wrong: "You seem to prefer X" Right: "Based on sessions from Jan 15 and Jan 20, you consistently chose X"

Confusing with /recall

Wrong: Putting reusable techniques here Right: Observations go to diaryed; techniques go to learned

Making Prescriptive Rules

Wrong: "You should always use TypeScript" Right: "Observed: Chose TypeScript in 4/5 new projects"

Skipping processed.log Update

Wrong: Saving reflection but not updating processed.log Right: Always append processed diary filenames to log after save

Auto-Updating CLAUDE.md

Wrong: Automatically updating CLAUDE.md when violations found Right: Report violations and ask user if they want to update rules

Ignoring Strategy Header

Wrong: Starting reflection without showing the filtering strategy Right: Always output strategy header (Mode, Diaries, Reason, Projects)

Example Output

Input: 5 Diary Entries

Diary 1 (2025-01-15)

## Decisions Made
- Chose PostgreSQL for JSON support
- Used connection pooling

## Challenges & Solutions
- **Challenge**: User corrected "Added with AI assistance" in commit
- **Solution**: Removed AI attribution
- **Generalizable?**: Yes

Diary 2 (2025-01-18)

## User Preferences Observed
- Prefers explicit error handling
- Likes PostgreSQL for complex queries

Diary 3 (2025-01-20)

## Decisions Made
- Selected PostgreSQL again for new service

Diary 4 (2025-01-22)

## Challenges & Solutions
- **Challenge**: User said "Don't mention Claude in PR description"
- **Solution**: Removed AI mention
- **Generalizable?**: Yes

Extracted Reflection

## Reflect Strategy
**Mode:** unprocessed
**Diaries analyzed:** 5
**Reason:** 5 new diaries since last reflection (2025-01-10)
**Projects covered:** my-api-project (5)

---

## Rule Violations Detected (PRIORITY)

### Violation: AI Attribution
**Existing Rule:** "Never add AI attribution to commits" (from CLAUDE.md)
**Violation Pattern:** User corrected Claude in 2 sessions
**Evidence:**
- [2025-01-15] diary-abc123: "User corrected 'Added with AI assistance' in commit"
- [2025-01-22] diary-ghi789: "User said 'Don't mention Claude in PR description'"
**Suggested Action:** Strengthen rule in CLAUDE.md to include PR descriptions

---

## Patterns Identified (3+ occurrences)

### Pattern: prefers-postgresql
**Occurrences:** 3 sessions
**Evidence:**
- [2025-01-15] Session abc123: Chose PostgreSQL for JSON support
- [2025-01-18] Session def456: Expressed preference for PostgreSQL complex queries
- [2025-01-20] Session ghi789: Selected PostgreSQL for new service
**Implication:** PostgreSQL is the preferred default database
**Confidence:** High

---

## Observations (1-2 occurrences)

- Prefers connection pooling: seen in 1 session
- Prefers explicit error handling: seen in 1 session

Red Flags - DO NOT Extract as Pattern

• Pattern based on fewer than 3 diary entries (use "Observation" label instead)
• No specific diary citations
• Prescriptive language ("always do X")
• Already extracted in previous reflect session (check processed.log)
• Already captured as instinct (check for duplicates)
• Auto-updating CLAUDE.md without user approval