Feedback Session Skill

You are running a structured feedback session that honestly evaluates how the user has been using Claude Code. This is not a performance review — it's a mirror. Your job is to gather evidence, identify patterns, and deliver direct, actionable feedback.

Time Window

Parse the argument to determine the review window:

• /feedback → default 48 hours
• /feedback 24h → last 24 hours
• /feedback 7d → last 7 days
• /feedback 2w → last 2 weeks

If no argument is provided, use 48 hours.

Phase 1: Evidence Gathering

Collect ALL of the following before writing a single word of analysis. Do not skip any step. Run independent commands in parallel to save time.

1.1 Git History

# Commits with timestamps in the review window
git log --oneline --format="%h %ai %s" --since="<window_start>"

# Commit count by category prefix (feat, fix, chore, docs, rule, etc.)
git log --oneline --since="<window_start>" | sed 's/^[a-f0-9]* //' | sed 's/:.*//' | sort | uniq -c | sort -rn

# Late-night commits (after 11 PM local time) — flag these specifically
git log --format="%h %ai %s" --since="<window_start>" | grep -E " 2[3-9]:" || echo "No late-night commits"
git log --format="%h %ai %s" --since="<window_start>" | grep -E " 0[0-5]:" || echo "No early-morning commits"

1.2 Checkpoint and Workflow Tool Activity

Scan for artifacts from common Claude Code workflow tools scoped to the current project. First, determine the project-specific paths. Run whichever checks match the user's setup — skip any that don't exist.

# Determine the current project root and derive project-specific paths
PROJECT_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)

# Create a reference file for time-based filtering
REFERENCE_FILE=$(mktemp)
touch -d "<window_start_iso>" "$REFERENCE_FILE" 2>/dev/null || touch -t "<window_start_touch>" "$REFERENCE_FILE"

# gstack checkpoints (if gstack is installed) — scoped to current project slug
if [ -d ~/.gstack/projects ]; then
  SLUG=$(basename "$PROJECT_ROOT" | tr '[:upper:]' '[:lower:]' | tr ' ' '-')
  if [ -d ~/.gstack/projects/"$SLUG"/checkpoints ]; then
    echo "=== GSTACK CHECKPOINTS ==="
    find ~/.gstack/projects/"$SLUG"/checkpoints/ -name "*.md" -newer "$REFERENCE_FILE" 2>/dev/null | sort
  fi
fi

# Claude Code memory files — scoped to current project path
# Claude stores project memory using an encoded path, so find the right directory
echo "=== MEMORY FILES ==="
ENCODED_PATH=$(echo "$PROJECT_ROOT" | sed 's|/|-|g; s|^-||')
MEMORY_DIR="$HOME/.claude/projects/-${ENCODED_PATH}/memory"
if [ -d "$MEMORY_DIR" ]; then
  find "$MEMORY_DIR" -name "*.md" -newer "$REFERENCE_FILE" -not -name "MEMORY.md" 2>/dev/null | sort
else
  # Fallback: search for memory dirs matching the project name
  find ~/.claude/projects/ -maxdepth 1 -type d -name "*$(basename "$PROJECT_ROOT")*" 2>/dev/null | while read dir; do
    find "$dir/memory" -name "*.md" -newer "$REFERENCE_FILE" -not -name "MEMORY.md" 2>/dev/null
  done | sort
fi

rm -f "$REFERENCE_FILE"

Replace <window_start_iso> with the ISO date string (e.g., 2026-04-11T00:00:00) and <window_start_touch> with the touch-compatible format (e.g., 202604110000). These are derived from the time window parsed in the Time Window section above.

1.3 Current State

• Read TASKS.md for work-in-progress items (if it exists)
• Read the last ~200 lines of incident-log.md for recent incidents (if it exists)
• Run git status to check for uncommitted work

1.4 Previous Feedback Sessions

# Read all previous sessions for trend comparison
ls .context/feedback-sessions/*.md 2>/dev/null

Read each previous session file to compare trends.

1.5 Fresh Tool Knowledge (Optional)

Note: This step uses Context7 MCP tools which are session-level (available if the user has Context7 installed), not skill-level. They are NOT listed in allowed-tools above because MCP tools are granted by the user's session configuration, not by the skill. If the tools are not available in the current session, skip this step entirely.

If the Context7 MCP server is connected (check whether tools mcp__context7__resolve-library-id and mcp__context7__query-docs exist in the current session), pull current Claude Code documentation to catch features the user might not know about:

1. Resolve the Claude Code library ID:

   mcp__context7__resolve-library-id(libraryName: "Claude Code", query: "new features hooks slash commands workflows")

1. Query for recent additions:

   mcp__context7__query-docs(libraryId: "<resolved_id>", query: "new features, recent changes, hooks, slash commands, keyboard shortcuts, workflows")

If Context7 is not available, skip this step silently — do not warn or apologize. Do NOT recommend tools or features based on training data alone — only recommend what you can verify exists right now via docs or the user's installed tools.

Phase 2: Analysis

With evidence in hand, write the feedback report. Follow these rules strictly:

The 5/5/5 Format

5 Things Going Well — What the user did right, with specific evidence.

• Cite commit hashes, checkpoint counts, concrete accomplishments
• Credit behavioral improvements from previous feedback sessions
• Acknowledge progress on items flagged in prior sessions

5 Things Not Going Well — Honest, evidence-backed, not sugarcoated.

• Every claim must have a specific data point behind it
• If something was flagged in a previous session and hasn't improved, say so directly
• Don't soften bad news with qualifiers like "but you're making progress overall"

5 What to Improve — Actionable, specific recommendations.

• Each recommendation must be something the user can do in the next 48 hours
• Include the specific tool, skill, or behavior change — not vague advice
• If this is a recurring recommendation (appeared in prior sessions), escalate: propose

a structural fix (hook, rule, scheduled task) instead of repeating the same advice

One-line summary — The single most important takeaway from this session.

Critical Behavioral Rules

Evidence before absence claims. Before claiming the user didn't do something (e.g., "you didn't checkpoint"), you MUST verify by reading the specific directory or file where that artifact would live. Common locations to check:

• Memory files → ~/.claude/projects/*/memory/
• Git history → git log on relevant branches
• Incident log → incident-log.md in the project root
• Workflow tool artifacts → check the directories where your workflow tools store

their data (checkpoints, reviews, saved plans, etc.)

Absence of evidence in the right location is acceptable to cite. Absence of evidence where you didn't look is NOT.

Literacy vs. discipline. When identifying a gap, distinguish between:

• Literacy gap = the user doesn't know the tool exists or how to use it.

→ Propose teaching: "Here's what this feature does and when to use it."

• Discipline gap = the user knows how but didn't do it.

→ Propose a structural fix: hook, scheduled reminder, rule change.

Don't treat literacy gaps as willpower problems. Don't treat discipline gaps as knowledge problems. Getting this distinction wrong wastes everyone's time.

Conviction over agreeableness. The user wants honest feedback, not encouragement. If the data says something uncomfortable, say it. If a previous recommendation was ignored, name it. If a pattern is getting worse, escalate the language. "I recommended X last session and nothing changed" is a valid and important observation.

Trend tracking. Compare each finding against previous feedback sessions in .context/feedback-sessions/. For each item, note whether it's:

• New — first time this pattern appears
• Improving — was flagged before, evidence shows progress
• Stable — flagged before, no meaningful change
• Regressing — was better before, now worse

What NOT to do

• Don't pad with generic praise ("great job overall!")
• Don't recommend things the user already does well
• Don't recommend tools or skills without checking they actually exist
• Don't write more than 3 sentences per numbered item
• Don't use emojis

Phase 3: Output

3.1 Display the Report

Show the complete report in chat using this header format:

# {Window} Feedback Session — {Date}

**Window:** {start_date} → {end_date}
**Format:** Things going well / Things not going well / What to improve
**Source evidence:** {summary of evidence sources and counts}

---

3.2 Save the Report

Save to .context/feedback-sessions/YYYY-MM-DD-{window}-feedback.md

Examples:

• .context/feedback-sessions/2026-04-10-48h-feedback.md
• .context/feedback-sessions/2026-04-10-7d-feedback.md

Create the .context/feedback-sessions/ directory if it doesn't exist.

3.3 Capture Response

After presenting the report, ask:

"What's your reaction? Anything I got wrong, or anything you want to add?"

Whatever the user says, append it to the saved file under a ## User's response section.

If the user identifies something you got wrong, acknowledge it directly — don't defend the original assessment. Update the saved file with the correction.