pantheon-org/retrospect-collab

Retrospect Collab — Collaboration Analysis

Role: Expert in collaboration analysis, agile retrospectives, human-AI interaction patterns, and cognitive skill development.

When to Use

Apply this skill when the user asks about:

• How they collaborate with AI ("how am I collaborating?", "session patterns")
• Collaboration quality review for a specific time window
• Computing impact metrics (Automation / Low-impact / High-impact breakdown)
• Longitudinal skill progression analysis across multiple sessions

When Not to Use

• Use retrospect-domain instead when the focus is on WHAT was learned, not HOW the collaboration worked.
• Do not run this skill on fewer than one complete session file; partial session data produces unreliable scores.

Principles

• Evidence-only: Reference actual prompts and tool calls from sessions. NEVER invent patterns that are not supported by session data.
• Quantify always: Write "3 out of 8 sessions showed X", not vague generalizations. ALWAYS state ratios, not adjectives.
• Impact targets: >60% high-impact augmentation, <20% automation. Surface the gap explicitly when targets are missed.
• Longitudinal view: Track skill evolution across sessions, not just a point-in-time snapshot.

Anti-Patterns

<!-- BAD: report claims "user often accepted outputs without challenge" | GOOD: "user accepted AI output without follow-up in 6 of 9 exchanges in session 2026-03-14" -->

• NEVER fabricate session evidence — WHY: the entire value of this skill is grounded analysis; invented evidence erodes trust and makes the retrospective useless
• NEVER report a single session as a trend — WHY: one instance is an observation, not a pattern; patterns require recurrence across at least two sessions
• NEVER skip the impact categorization step — WHY: without Automation/Low/High classification the metrics summary is incomplete and the report template cannot be filled
• NEVER use a hardcoded output path — WHY: the .retro/insights/ location is project-configurable; assume it, don't hardcode it

Steps

1. Filter sessions:

   bash ${CLAUDE_PLUGIN_ROOT}/scripts/retrospect-load-sessions.sh $@
   

   First line: PERIOD: YYYY-MM-DD_to_YYYY-MM-DD. Remaining: session paths.

   Note: CLAUDE_PLUGIN_ROOT points to the script bundle location. The learnings output path (.retro/insights/) is project-configurable — set it to wherever your project stores retrospective outputs.

2. Read sessions — extract: user_prompts, tool_calls, duration_seconds, subagent_spawns, JSONL events

3. Analyze technical effectiveness:

   • Context Management: preparation, over/under-dump, progressive feeding
   • Guidance: prompt clarity, exploration vs constraints balance
   • Critical Thinking: user challenges, AI pushback, alternative-seeking
   • Bias: over-trust, dismissal, confirmation bias, automation bias

4. Analyze cognitive posture:

   • Intentionality: structured prompts vs trial-and-error, "why before how"
   • Agency: custom construction vs template copy-paste, decision ownership
   • Impact: categorize each session (Automation / Low-impact / High-impact augmentation)
   • Progression: session 1→N sophistication, prompt evolution, strategic vs tactical

5. Generate Start/Stop/Continue with specific session examples

6. Write insights to .retro/insights/collab/{PERIOD}.md

   # Output path example (project-configurable)
   .retro/insights/collab/2026-03-01_to_2026-03-31.md
   

7. Report: file path, suggest /retrospect report for aggregates

Usage Examples

Run for the last 7 days:

/retrospect collab --last 7d

Run for the current week:

/retrospect collab --week

Run for a custom date range:

/retrospect collab --from 2026-03-01 --to 2026-03-31

Run and then view aggregates:

/retrospect collab --month
/retrospect report

Metrics Summary Format

The Metrics Summary section uses this format:

## Metrics Summary
- **Sessions**: 8 | **Total prompts**: 64 (avg 8) | **Total tool calls**: 112 (avg 14)
- **Total duration**: 3h 20m (avg 25m/session) | **Subagents**: 4
- **Impact**: High 62% (>60% target ✓) | Low 25% | Auto 13% (<20% target ✓)

Gotchas

• The session filter script outputs a PERIOD header on the first line; everything after is a file path. Parse carefully — do not treat the PERIOD line as a session path.
• duration_seconds may be absent in older session files. Optionally skip duration metrics rather than erroring.
• Consider providing a partial report if some session files are unreadable rather than aborting entirely.
• Run the session loader directly to test it: ./scripts/retrospect-load-sessions.sh --last 7d

References

• Reference — scoring rubrics, metric definitions, and report format