Session Insights

End-to-end pipeline: search sessions by natural language prompt, analyze all matches, synthesize findings.

Usage

/phx:session-insights "all my Elixir Phoenix sessions"
/phx:session-insights "sessions where I debugged LiveView"
/phx:session-insights --project my_app
/phx:session-insights "Oban worker problems" --after 2026-01-15
/phx:session-insights "sessions with test failures" --limit 10

Requirements

Requires ccrider MCP. If not available, tell the user:

ccrider MCP is required. See: https://github.com/neilberkman/ccrider

Pipeline

Step 1: Parse intent from `$ARGUMENTS`

Extract from $ARGUMENTS:

Prompt (the non-flag text): Natural language description of what to find
--project NAME: Filter by project (match substring in path)
--after DATE: Only sessions after this date (ISO format)
--limit N: Max sessions to analyze (default: 15)

If no prompt given, default to: "all Elixir and Phoenix development sessions"

Step 2: Translate prompt to search queries

Based on the user's prompt, generate 2-5 ccrider search queries that cover the intent. Think about what terms would appear in sessions matching the description.

Examples:

User Prompt	Queries to Run
"all my Elixir sessions"	`mix compile`, `\.ex`, `Phoenix`, `Ecto`
"LiveView debugging"	`LiveView`, `_live.ex`, `assign_async`, `handle_event`
"Oban worker problems"	`Oban`, `worker`, `perform`, `queue`
"sessions where I used the plugin"	`/phx:`, `elixir-phoenix:`, `Iron Law`
"database work"	`Ecto`, `migration`, `Repo.`, `schema`
"project my_app"	(use --project filter, no query needed)
"test failures"	`mix test`, `test failed`, `assertion`, `ExUnit`

Run each query with mcp__ccrider__search_sessions, passing --project and --after filters if provided.

Step 3: Merge and deduplicate

Combine results from all queries. Deduplicate by session ID. Sort by date (most recent first).

Report to user:

Found {N} unique sessions matching "{prompt}" (from {Q} searches). Projects: {list of unique project names} Date range: {earliest} to {latest}

If more than --limit sessions found, take the most recent N.

If 0 sessions found, suggest broader search terms and stop.

Step 4: Confirm with user

Show the session list as a compact table:

| #  | ID       | Project | Date       | Msgs | Summary                    |
|----|----------|---------|------------|------|----------------------------|
| 1  | 90a74843 | my_app   | 2026-02-09 | 30   | Fix gettext translations   |

Ask:

"Analyze all {N} sessions? Or enter specific numbers (e.g., 1,3,5) to select a subset."

Step 5: Fetch transcripts

For each confirmed session:

Call mcp__ccrider__get_session_messages(session_id: id)
If >200 messages, use last_n: 150 and note the truncation
Write transcript to .claude/session-analysis/{short-id}-transcript.md

Format each transcript file:

# Session: {short-id}

Project: {project}
Date: {date}
Messages: {count}

## Messages

### User (seq 1)

{content}

### Assistant (seq 2)

{content}

Step 6: Analyze sessions

Locate the analysis template by searching for it:

Glob: **/analyze-session/references/analysis-template.md

Read the template contents. If not found, use the analysis questions inline from this skill's instructions.

Strategy by session count:

1-2 sessions: Analyze directly in context
3-6 sessions: Spawn sonnet subagents (one per session). Each reads transcript + template, writes report.
7+ sessions: Spawn haiku subagents for speed

Subagent prompt:

Read the analysis template at {template_path}. Read the session transcript at {transcript_path}. Apply the template to analyze this session. Write your report (under 200 lines) to {report_path}.

Write reports to .claude/session-analysis/{short-id}-report.md.

Step 7: Synthesize

After all reports are written, read them all and produce a synthesis:

.claude/session-analysis/insights-{date}.md

Structure:

# Session Insights: {prompt}

Date: {date}
Sessions analyzed: {N}
Projects: {list}

## Key Findings

### Most Common Friction Patterns

1. {pattern} — seen in {N} sessions
2. {pattern} — seen in {N} sessions

### Plugin Skills You Should Try

| Skill              | Why      | Sessions Where It Would Help |
| ------------------ | -------- | ---------------------------- |
| `/phx:investigate` | {reason} | {short-ids}                  |
| `/phx:plan`        | {reason} | {short-ids}                  |

### Plugin Improvement Opportunities

Aggregate findings from individual reports section 5. Group by type:

**Missing automation:**

- {concrete automation opportunity} — seen in {N} sessions

**Missing skills or agents:**

- {what should exist} — {evidence from sessions}

**Iron Law candidates:**

- {pattern that caused bugs} — should be caught automatically

**Auto-loading gaps:**

- {file pattern where skill should load but doesn't}

### Workflow Patterns

- {observation about how developers work}
- {observation about tool usage}
- {sessions with plugin vs without — what difference did it make?}

## Per-Session Summaries

### {short-id} — {project} ({date})

**Goal**: {1-line}
**Friction**: {key pain point}
**Plugin opportunity**: {top improvement suggestion}

(repeat for each session)

Step 8: Present results

Show the user the Key Findings section directly in the conversation. Tell them:

Full report: .claude/session-analysis/insights-{date}.md Per-session reports: .claude/session-analysis/{id}-report.md

Iron Laws

ALWAYS confirm before analyzing — show the session list, get user approval
NEVER skip the synthesis — individual reports are useful but the cross-session patterns are the real value
Respect limits — default 15 sessions max to keep analysis manageable

session-insights