workflow-improver

name	workflow-improver
description	Evaluate session using OpenAI eval-skills framework (Outcome/Process/Style/Efficiency). Analyzes session transcript vs Claude Code config to score performance and generate improvement recommendations. Creates GitHub issue with rubric scores and actionable plan.
context	fork
agent	general-purpose
allowed-tools	["Read","Glob","Grep","Bash(gh:*)"]

Workflow Improver

Analyze the current session transcript alongside Claude Code configuration to compare what was intended versus what actually happened. Generate improvement recommendations and create a GitHub issue.

Session Discovery

Claude Code stores session transcripts at:

~/.claude/projects/{project-path-encoded}/{session-id}.jsonl

Find the current session:

# Project path encoded (replace / with -)
!`pwd | tr '/' '-' | sed 's/^-//'`

# Find session directory
!`ls -t ~/.claude/projects/ | head -20`

# Current session ID
${CLAUDE_SESSION_ID}

Read the session transcript: Use the Read tool to read: ~/.claude/projects/!pwd | tr '/' '-' | sed 's/^-//'/${CLAUDE_SESSION_ID}.jsonl

Analysis Framework (OpenAI Eval-Skills)

1. Outcome Goals (Task Completion)

Analyze the transcript for:

• User requests: What did the user ask for?
• Final artifacts: What was produced?
• Completion status: Was the task fully completed?

Checklist:

• User's primary request fulfilled
• All acceptance criteria met
• No partial implementations left

2. Process Goals (Correct Steps)

Analyze the transcript for:

• Skill loading: Were skills loaded? At the right time?
• Gate compliance: Were all gates passed? Any skipped?
• Tool selection: Were the right tools used?

Checklist:

• Required skills loaded (TDD, verification, etc.)
• All gates output with proofs
• Tools used appropriately (no grep when Grep tool available, etc.)
• Task management used for multi-step work

3. Style Goals (Convention Conformance)

Analyze the transcript for:

• Iteration quality: Were iterations meaningful or boilerplate?
• Output format: Did responses follow required formats?
• Documentation: Were changes properly documented?

Checklist:

• Gate outputs follow required format
• Iteration counts tracked
• Evidence provided for each gate checkbox

4. Efficiency Goals (Resource Optimization)

Analyze the transcript for:

• Context management: Was context used efficiently?
• Tool calls: Were there unnecessary/redundant calls?
• Parallelization: Were parallel opportunities exploited?

Checklist:

• No redundant file reads
• Parallel tasks dispatched in parallel
• Explore agent used for broad searches

Configuration Files to Review

Read these files to understand the intended workflow:

1. CLAUDE.md: ./CLAUDE.md
2. Crew skill: .agents/skills-local/crew-claude/SKILL.md (contains workflows, hard requirements, anti-patterns, and skill routing)

Scoring Rubric

For each category, assign a score (0-100) and pass/fail:

Category	Score Range	Pass Threshold
Outcome	0-100	≥70
Process	0-100	≥60
Style	0-100	≥50
Efficiency	0-100	≥50

Scoring Guidelines:

Outcome:

• 90-100: All requests fulfilled, high quality output
• 70-89: Primary request fulfilled, minor gaps
• 50-69: Partial completion, significant gaps
• 0-49: Failed to complete primary request

Process:

• 90-100: All skills loaded, all gates passed with proofs
• 70-89: Most skills loaded, gates mostly complete
• 50-69: Some skills missing, gates incomplete
• 0-49: Major process violations

Style:

• 90-100: Perfect format compliance, meaningful iterations
• 70-89: Good format, some boilerplate iterations
• 50-69: Format issues, iterations lack depth
• 0-49: Format violations, no meaningful iteration

Efficiency:

• 90-100: Optimal tool usage, perfect parallelization
• 70-89: Good efficiency, minor redundancy
• 50-69: Some wasted operations
• 0-49: Significant inefficiency

Output Format

Generate a structured analysis:

{
  "session_id": "${CLAUDE_SESSION_ID}",
  "outcome": { "score": <N>, "pass": <bool>, "notes": "<summary>" },
  "process": { "score": <N>, "pass": <bool>, "notes": "<summary>" },
  "style": { "score": <N>, "pass": <bool>, "notes": "<summary>" },
  "efficiency": { "score": <N>, "pass": <bool>, "notes": "<summary>" },
  "overall": <average>,
  "improvements": [
    { "priority": 1, "category": "<O/P/S/E>", "file": "<path>", "suggestion": "<change>" }
  ]
}

GitHub Issue Creation

After analysis, create a GitHub issue:

gh issue create --repo settlemint/agent-marketplace \
  --title "Workflow Improvement: Session ${CLAUDE_SESSION_ID}" \
  --label "workflow-improvement,session-memory" \
  --body "$(cat <<'EOF'
## Workflow Improvement Recommendations

**Session analyzed:** ${CLAUDE_SESSION_ID}
**Project:** !`basename $(pwd)`
**Date:** !`date +%Y-%m-%d`

### Session Evaluation Scores

| Category | Score | Pass | Notes |
|----------|-------|------|-------|
| Outcome | <score>/100 | <emoji> | <notes> |
| Process | <score>/100 | <emoji> | <notes> |
| Style | <score>/100 | <emoji> | <notes> |
| Efficiency | <score>/100 | <emoji> | <notes> |
| **Overall** | **<avg>** | | |

### Session Summary
<Brief summary of what the user wanted to accomplish>

### Intent vs Execution Analysis
| User Intent | What Happened | Category | Gap |
|-------------|---------------|----------|-----|
| <request 1> | <execution>   | <O/P/S/E> | <gap> |

### Improvements by Priority

#### P1 (Critical)
- [ ] [File: path] [Change: description]

#### P2 (Important)
- [ ] [File: path] [Change: description]

#### P3 (Nice-to-have)
- [ ] [File: path] [Change: description]

### Suggested New Skills/Routing
- [ ] Add skill: [name] - addresses [category] gap
- [ ] Add routing: [trigger] -> [skill]

### Workflow Adjustments
- [ ] [Phase/gate modification]

---
**Files analyzed:**
- Session: ~/.claude/projects/[path]/[session].jsonl
- Config: CLAUDE.md, .claude/skills/*, .agents/skills-local/crew-claude/*
**Framework:** OpenAI eval-skills (Outcome/Process/Style/Efficiency)
EOF
)"

Execution Steps

1. Discover session: Find and read the current session transcript
2. Analyze transcript: Parse user messages, tool calls, and outputs
3. Score each category: Apply rubric to generate scores
4. Read config files: Understand intended workflow
5. Identify gaps: Compare intent vs execution
6. Generate improvements: Prioritized, actionable suggestions
7. Create issue: GitHub issue with full analysis

Begin by discovering the session transcript.