Repeatable software development workflows for Claude Code that keep design decisions traceable through implementation, tests, and review. Each phase runs in a fresh agent context and hands off through explicit artifacts.
@shinprFAQ
rashomon is a Claude Code plugin with 6 hand-picked skills for testing work, indexed on Flowy. Install it with the command on its page. It includes prompt-optimization, knowledge-base, prompt-optimization. Its skills do not fire on their own yet. Request auto-invocation to have Flowy route them as you prompt. Free and open source.
Repo: shinpr/rashomon
Know whether your skills actually improve agent behavior — not just look different.
Inspired by the Rashomon effect — the idea that the same event can produce different outcomes depending on perspective. rashomon makes those differences explicit and comparable.
rashomon evaluates skills and prompts through blind comparison — running tasks with and without your changes in isolated environments, then comparing real outputs without knowing which version produced which.
rashomon is designed for:
Not ideal if:
/recipe-eval-skill create
Creates a skill through interactive dialog, then evaluates effectiveness:
What the evaluation report looks like:
Skill Quality: Grade A
- Project-specific rules clearly encoded, no critical issues
Trigger Check: pass (loaded through the Skill tool or its SKILL.md file)
Execution Effectiveness:
- Winner: with-skill
- Assessment: repeatable structural improvement across valid pairs
- Key difference: Retry constraints and three-stage catch ordering were applied
consistently (linked to skill Rules 3 and 6)
Recommendation: ship
/recipe-eval-skill api-error-handling skill's scope needs adjustment
Updates an existing skill, then evaluates old vs new version side by side.
See a real-world example: I Built a Skill Reviewer. Then I Ran It on Itself.
/recipe-eval-prompt Write a function to sort an array
Analyzes prompt issues, generates an improved version, and runs the original and optimized prompts in up to three blind, paired trials. The report highlights differences that recur across trials.
1. Analysis
- BP-002 (already satisfied): No consumer requirement selects a language,
ordering, or error policy, so those choices remain flexible.
2. Final Prompt
Write a function to sort an array
Result: Original sufficient - Rashomon stops before paired execution because no outcome-relevant ambiguity was found.
Requires Claude Code (this is a Claude Code plugin)
# 1. Start Claude Code
claude
# 2. Install the marketplace
/plugin marketplace add shinpr/rashomon
# 3. Install plugin
/plugin install rashomon@rashomon
# 4. Restart session (required)
# Exit and restart Claude Code
/recipe-eval-skill create
Create a new skill and evaluate its effectiveness.
/recipe-eval-skill my-skill-name what to change
Update an existing skill and compare old vs new.
/recipe-eval-prompt Your prompt here
From a file:
/recipe-eval-prompt Generate code following this skill: ./prompts/my-skill.md
For complex tasks that need more time, just mention it in natural language:
/recipe-eval-prompt Refactor the entire authentication module. This might take a while.
Skill Evaluation (/recipe-eval-skill)
├── skill-creator (generates/modifies skills)
├── skill-reviewer (grades quality A/B/C)
├── eval-executor (up to 3 valid pairs, sequential within each pair)
└── skill-eval-reporter (blind A/B comparison)
Prompt Evaluation (/recipe-eval-prompt)
├── prompt-analyzer (analyzes and optimizes)
├── prompt-executor (up to 3 valid pairs, parallel within each pair)
└── report-generator (compares results)
rashomon uses git worktrees to run both versions in completely separate environments. A worktree is a Git feature that creates independent working directories from the same repository—this ensures the two executions don't interfere with each other.
Not all differences are improvements. rashomon classifies results into four categories:
| Classification | Meaning | Recommendation |
|---|---|---|
| Structural | Real improvement in accuracy, completeness, or quality | Use the new version |
| Context Addition | One version had more project-specific knowledge | Useful if the context is accurate |
| Expressive | Different wording, same substance | Either version is fine |
| Variance | Just normal LLM randomness | Original was already good |
Classification is based on:
Both skill review and prompt analysis check against 8 common patterns:
| Priority | Issues |
|---|---|
| Critical | Negative instructions ("don't do X"), vague instructions, missing output format |
| High Impact | Unstructured prompts, missing or excess context, complex tasks without gates |
| Enhancement | Unnecessary or biased examples, missing uncertainty handling |
| ID | Pattern | Problem | Fix |
|---|---|---|---|
| BP-001 | Negative Instructions | A prohibition can prime the forbidden behavior without naming the target state | Lead with the desired behavior; retain narrow prohibitions for irreversible actions |
| BP-002 | Vague Instructions | An outcome-relevant decision has materially different plausible interpretations | Apply the least-restrictive sufficient criterion that preserves valid solutions |
| BP-003 | Missing Output Format | No format spec leads to inconsistent outputs | Define expected structure: JSON schema, section headers, etc. |
| ID | Pattern | Problem | Fix |
|---|---|---|---|
| BP-004 | Unstructured Prompt | Wall of text obscures priorities | Apply 4-block pattern: Context / Task / Constraints / Output Format |
| BP-005 | Missing or Excess Context | Missing facts force guesses; excess facts obscure operative instructions | Supply decision-sufficient context and condense the rest |
| BP-006 | Complex Task Without Gates | A model can jump past prerequisites to a later result | Break into states with completion evidence and transition conditions |
| ID | Pattern | Problem | Fix |
|---|---|---|---|
| BP-007 | Unnecessary or Biased Examples | Generic examples consume context and anchor unrelated details | Use examples only for non-obvious domain or organization-specific mappings |
| BP-008 | Missing Uncertainty Handling | Unknown inputs have no defined next action | Classify evidence and stop at a gate when required input is unresolved |
Prompt optimization keeps analysis, optimization, and balance state in one gated JSON response. It avoids intermediate analysis files, so the skill does not depend on a specific agent or filesystem layout.
rashomon learns from your project over time.
Location: .claude/.rashomon/prompt-knowledge.yaml
How it works:
Key principle: Old knowledge isn't automatically removed. Patterns that have worked for a long time are often the most valuable.
If rashomon exits unexpectedly, temporary directories might remain:
# Worktrees are stored in system temp directory
# Clean up manually if needed:
rm -rf ${TMPDIR:-/tmp}/worktree-rashomon-*
For complex prompts that need more time, mention it when invoking:
/recipe-eval-prompt Complex task here. This might take longer than usual.
rashomon requires a git repository. Initialize one with:
git init
MIT
.agents/
plugins/
marketplace.json
.claude-plugin/
marketplace.json
plugin.json
.github/
workflows/
test.yml
.gitignore
.pre-commit-config.yaml
agents/
knowledge-optimizer.md
prompt-analyzer.md
prompt-executor.md
report-generator.md
skill-creator.md
skill-eval-reporter.md
skill-reviewer.md
assets/
rashomon-banner.jpg
LICENSE
plugins/
rashomon/
.codex-plugin/
plugin.json
skills/
prompt-optimization/
agents/
openai.yaml
references/
execution-quality.yaml
patterns.yaml
skills.md
SKILL.md
pyproject.toml
README.md
scripts/
sync_plugin.py
skills/
knowledge-base/
SKILL.md
prompt-optimization/
agents/
openai.yaml
references/
execution-quality.yaml
patterns.yaml
skills.md
SKILL.md
recipe-eval-prompt/
SKILL.md
recipe-eval-skill/
references/
create.md
eval.md
update.md
scripts/
eval-executor.py
SKILL.md
worktree-execution/
scripts/
worktree-cleanup.sh
worktree-create.sh
SKILL.md
tests/
test_eval_executor.py
test_worktree_scripts.py© 2026 Flowy · Free and open source
Built for Claude Code · Not affiliated with Anthropic