# EvanFlow--一个用于 Claude Code 的基于 TDD 的反馈循环

- 来源：Hacker News 热门（buzzing.cc 中文翻译）
- 作者：evanklem2004
- 发布时间：2026-04-28 05:43
- AIHOT 分数：61
- AIHOT 链接：https://aihot.virxact.com/items/cmoi4smzq00g9sle9rw1k09fr
- 原文链接：https://github.com/evanklem/evanflow

## AI 摘要

EvanFlow是一个专为Claude Code设计的、基于测试驱动开发（TDD）的反馈循环工具，已在GitHub开源。该工具旨在通过TDD流程优化开发反馈循环，提升代码质量与开发效率。项目在Hacker News上获得了100点热度，显示出社区的关注。其核心是构建一个系统化的自动化测试与编码迭代流程，帮助开发者更高效地利用Claude Code进行编程。

## 正文

EvanFlow

A TDD-driven iterative feedback loop for software development with Claude Code.

16 cohesive skills + 2 custom subagents walk an idea from brainstorm through implementation, with checkpoints throughout where you stay in control. One entry point: say "let's evanflow this" and the orchestrator runs the loop.

brainstorm → plan → execute (vertical-slice TDD per task) → iterate → STOP └─ sequential, or parallel coder/overseer

brainstorm → plan → execute (vertical-slice TDD per task) → iterate → STOP └─ sequential, or parallel coder/overseer

TDD is not a separate phase after execute — it's the discipline inside each code-writing task. Execute is the harness (task tracking, blockers, quality checks); evanflow-tdd is what runs inside any task that produces production code.

evanflow-tdd

The loop is conductor, not autopilot: real checkpoints at design approval, plan approval, and after iteration. The agent stops short of every git operation and waits for your direction. No auto-commits. No forced ceremony. No "must invoke a skill" tax.

Quick Install

The recommended path — Claude Code's plugin marketplace:

/plugin marketplace add evanklem/evanflow /plugin install evanflow@evanflow

/plugin marketplace add evanklem/evanflow /plugin install evanflow@evanflow

Restart, then try:

"Let's evanflow this — I want to add a small feature that does X."

"Let's evanflow this — I want to add a small feature that does X."

evanflow-go fires and walks the loop. The git-guardrails hook auto-activates with the plugin (no settings.json edit needed). Skills appear under the evanflow: namespace (e.g., /evanflow:evanflow-go).

evanflow-go

evanflow:

/evanflow:evanflow-go

See Installation below for two alternative paths.

What Makes It a Feedback Loop

The loop is built around discipline that compounds across iterations, not single-shot generation. Every step has a checkpoint that gates the next:

Brainstorm clarifies intent, proposes 2–3 approaches with embedded grill (stress-test) → you approve the design

Plan maps file structure first (deep modules, deletion test) → you approve the plan

Execute runs task-by-task with inline verification → blockers stop the loop and surface to you. Inside each code-writing task, TDD is the discipline (not a separate phase that comes after).

TDD is vertical-slice only and per-cycle full RED → GREEN → REFACTOR: one failing test → minimal impl → refactor while the test you just wrote is still fresh as your safety net → next test. Refactor is not deferred to the end. Tests verify behavior through public interfaces, so they survive refactors

Iterate re-reads the diff with fresh eyes, runs quality checks, screenshots UI changes, and runs against a Five Failure Modes checklist (hallucinated actions, scope creep, cascading errors, context loss, tool misuse). Hard cap of 5 iterations

STOP. Report. Await your direction. The agent never auto-commits, never auto-stages, never proposes a PR

For plans with 3+ truly independent units, the loop forks into a parallel coder/overseer orchestration: one coder per unit (using vertical-slice TDD with a RED checkpoint), one overseer per coder (read-only review subagent that can't modify code), plus an integration overseer that runs named integration tests at every touchpoint. The integration tests are the executable contract — interfaces can't drift if both sides have to satisfy the same passing test.

Hard Rules Baked Into the Loop

Each rule below cites the source it came from. If a citation is missing, the rule is opinion from running the loop on real projects, not research — labeled as such.

Never invent values — file paths, env vars, IDs, function names, library APIs. If unsure, the agent stops and asks. Source: action-hallucination is the top failure mode in DAPLab/Columbia "9 Critical Failure Patterns of Coding Agents".

Assertion-correctness warning — over 62% of LLM-generated test assertions were incorrect across HumanEval evaluation of four LLMs. Source: "Test-Driven Development for Code Generation" (arXiv 2402.13521), §3.2. Both evanflow-tdd and the overseer review explicitly check whether a one-character bug in the implementation would still let the assertion pass.

evanflow-tdd

Five Failure Modes pass in iterate + overseer review — hallucinated actions, scope creep, cascading errors, context loss, tool misuse. Source: synthesized from the DAPLab failure patterns paper above.

Context drift watch — evanflow-compact triggers at clean phase boundaries and on drift symptoms (re-asking settled questions, contradicting earlier decisions). Source: nearly 65% of enterprise AI failures in 2025 were attributed to context drift or memory loss during multi-step reasoning, not raw context exhaustion — see Alex Merced, "Context Management Strategies for OpenCode" (March 2026).

evanflow-compact

Never auto-commit, never auto-stage — opinion, not research. Came from running the loop on real projects: every time the agent decided to integrate, it integrated wrong.

No skill tax — opinion. Skills are tools, not a tollbooth.

The Skill Set

Default Loop (5 skills)

Skill Purpose evanflow-brainstorming Clarify intent, propose 2–3 approaches with embedded grill (stress-test). Mockup quick-mode for visual-only requests. evanflow-writing-plans File structure first, bite-sized tasks, embedded grill. Step 2.5 offers evanflow-coder-overseer if the plan is parallelizable. evanflow-executing-plans Task-by-task with inline verification. Step 0 re-offers parallel path. Hands off to iterate, then STOPS. evanflow-tdd Vertical-slice TDD. One test → one impl → repeat. Behavior through public interface. Assertion-correctness warning. evanflow-iterate Self-review loop after implementation. Re-read diff, fix issues, run quality checks, screenshot UI (via headless Chromium). Five Failure Modes checklist. Hard cap of 5 iterations.

evanflow-brainstorming

evanflow-writing-plans

evanflow-coder-overseer

evanflow-executing-plans

evanflow-tdd

evanflow-iterate

Special-Purpose (9 skills)

Skill Purpose evanflow-go Single entry point. Say "let's evanflow this" and it walks the whole loop. evanflow-coder-overseer Parallel implementation: coder/overseer pairs + integration overseer + executable cohesion contract. For plans with 3+ independent tasks. evanflow-glossary Extract canonical domain terms into CONTEXT.md. Flag ambiguities and synonyms. evanflow-improve-architecture Surface refactor opportunities via the deletion test + deep-modules vocabulary. evanflow-design-interface "Design it twice" — spawn 3+ parallel sub-agents with radically different constraints, compare on depth/simplicity/efficiency. evanflow-debug Root-cause discipline. Hypothesis stated explicitly, embedded grill before fixing, failing test first. evanflow-review Both halves of code review (giving + receiving). Don't capitulate to feedback you can't justify. evanflow-prd Synthesize a PRD from existing context. For substantial new features. evanflow-qa Conversational bug discovery → issue draft. Asks before filing.

evanflow-go

evanflow-coder-overseer

evanflow-glossary

CONTEXT.md

evanflow-improve-architecture

evanflow-design-interface

evanflow-debug

evanflow-review

evanflow-prd

evanflow-qa

Cross-Cutting (1 skill)

Skill Purpose evanflow-compact Long-session context management. Strategies for proactive summarization at clean boundaries. Drift symptoms checklist.

evanflow-compact

Meta (1 skill)

Skill Purpose evanflow The index. Shared vocabulary + when to invoke each evanflow-* skill.

evanflow

evanflow-*

Custom Subagents (2)

In agents/ — invoked via Agent tool with subagent_type: parameter:

agents/

Agent

subagent_type:

Subagent Tool restrictions Purpose evanflow-coder Read, Edit, Write, Glob, Grep, Bash, TodoWrite Implementation subagent for evanflow-coder-overseer. Tools + system prompt prevent git ops, out-of-scope edits, value hallucination. evanflow-overseer Read, Grep, Glob (no Edit/Write/Bash) Read-only review subagent. Tools physically enforce "report findings, never fix."

evanflow-coder

evanflow-coder-overseer

evanflow-overseer

Bundled Hook

hooks/block-dangerous-git.sh — PreToolUse hook that blocks destructive git ops (git push, git reset --hard, git clean -f, git branch -D, git checkout ., git restore .). Auto-activates with the plugin install path.

hooks/block-dangerous-git.sh

git push

git reset --hard

git clean -f

git branch -D

git checkout .

git restore .

Hard Rules (apply to every skill)

Never auto-commit, never auto-stage, never auto-finish. Every git write op requires you to explicitly ask in the current turn.

Never invent values. File paths, env vars, IDs, function names, library APIs — if unsure, the agent stops and asks.

No skill tax. Ad-hoc questions don't require a skill invocation. Skills are tools, not a tollbooth.

No forced spec/plan paths. Files live where you want them.

Verify before claiming done. Quality checks (typecheck, lint, test) run before any "done" report.

Requirements

Claude Code (any recent version)

Bash — for the bundled hook script (Linux, macOS, or Windows + WSL)

jq — used by the hook script to parse Claude's JSON tool input. Install via apt install jq, brew install jq, or your platform's package manager. If jq is missing, the guardrail hook fails silently and dangerous git ops are NOT blocked.

jq

apt install jq

brew install jq

jq

Optional but recommended:

chromium or google-chrome — for evanflow-iterate's visual verification of UI changes (chromium --headless --screenshot=...). Falls back gracefully if missing — the skill flags it and asks you to verify visually.

chromium

google-chrome

evanflow-iterate

chromium --headless --screenshot=...

Installation

Three paths, in priority order. All three end with the same skill set in your .claude/skills/. The plugin path additionally auto-wires the guardrail hook.

.claude/skills/

Path 1 — Claude Code Plugin Marketplace (recommended)

This is the cleanest install. Skills, agents, AND the guardrail hook all activate automatically.

/plugin marketplace add evanklem/evanflow /plugin install evanflow@evanflow

/plugin marketplace add evanklem/evanflow /plugin install evanflow@evanflow

Restart Claude Code (or /reload-plugins). Skills appear namespaced as /evanflow:evanflow-go, /evanflow:evanflow-tdd, etc. Auto-invocation via "let's evanflow this" still works regardless of namespace.

/reload-plugins

/evanflow:evanflow-go

/evanflow:evanflow-tdd

To uninstall: /plugin uninstall evanflow@evanflow.

/plugin uninstall evanflow@evanflow

Path 2 — npx skills@latest add CLI

npx skills@latest add

Works against any GitHub repo with SKILL.md-shaped folders. Installs skills only — does not install the guardrail hook or custom subagents (you'd add those manually if you want them).

SKILL.md

# Install all 16 skills at once npx skills@latest add evanklem/evanflow -s '*' -y # Or install individual skills npx skills@latest add evanklem/evanflow/evanflow-go npx skills@latest add evanklem/evanflow/evanflow-tdd # ...

This places skills under ~/.claude/skills/ (global) or .claude/skills/ (project, auto-detected).

~/.claude/skills/

.claude/skills/

Path 3 — Manual Copy

For users who want full control, no CLI dependencies.

git clone https://github.com/evanklem/evanflow.git cd evanflow # Skills (project-level — adjust to ~/.claude/skills/ for global) mkdir -p .claude/skills cp -r skills/* .claude/skills/ # Agents (custom subagents used by evanflow-coder-overseer) mkdir -p .claude/agents cp agents/*.md .claude/agents/ # Git guardrails hook (optional but recommended) mkdir -p .claude/hooks cp hooks/block-dangerous-git.sh .claude/hooks/ chmod +x .claude/hooks/block-dangerous-git.sh

Then register the hook in your .claude/settings.json:

.claude/settings.json

{ "hooks": { "PreToolUse": [ { "matcher": "Bash", "hooks": [ { "type": "command", "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/block-dangerous-git.sh" } ] } ] } }

Optionally, paste examples/CLAUDE.md.snippet into your project's CLAUDE.md to brief Claude about EvanFlow's conventions.

examples/CLAUDE.md.snippet

CLAUDE.md

Verify Any Install Path

Restart Claude Code. Try saying:

"Let's evanflow this — I want to add a small feature that does X."

"Let's evanflow this — I want to add a small feature that does X."

evanflow-go should fire and walk you through the loop. To verify the guardrail hook (paths 1 and 3 only): try git reset --hard HEAD from the Bash tool — it should be blocked with "BLOCKED: ... matches dangerous pattern".

evanflow-go

git reset --hard HEAD

Customization

Every skill has a clear structure with a ## Hard Rules section. To adapt to your project:

## Hard Rules

Replace and placeholders in skills like evanflow-writing-plans with your actual paths if you find yourself answering the same question repeatedly.

evanflow-writing-plans

Document your project's quality checks in your CLAUDE.md — exact typecheck, lint, and test commands. The skills reference these abstractly.

CLAUDE.md

typecheck

lint

test

Adapt the visual verification step in evanflow-iterate if you don't have chromium available — substitute google-chrome --headless or another tool.

evanflow-iterate

chromium

google-chrome --headless

Edit the cohesion contract template in evanflow-coder-overseer to match your project's conventions (your authentication middleware name, your DB write helper, etc.).

evanflow-coder-overseer

The skills are designed to be edited. Treat them as starting points, not gospel.

If you fork to make a vendor-specific variant (your-name-flow), great — that's the spirit.

How EvanFlow Works End-to-End

You say: "let's evanflow this — I want to add a feature that does X" │ ▼ evanflow-go (the conductor) │ ├─ Phase 0: Restate idea, scope check ├─ Phase 1: evanflow-brainstorming (CHECKPOINT: design approval) ├─ Phase 2: evanflow-writing-plans (CHECKPOINT: plan approval) │ └─ Step 2.5: parallelization check ├─ Phase 3: evanflow-executing-plans (sequential) │ OR │ evanflow-coder-overseer (parallel) │ ├─ contract with named tests + integration tests │ ├─ RED checkpoint (all coders write failing tests, orchestrator verifies) │ ├─ GREEN phase (vertical-slice TDD per coder) │ ├─ per-coder overseers (review, never fix) │ └─ integration overseer (runs touchpoint tests) ├─ Phase 4: evanflow-iterate (5x cap, Five Failure Modes pass) └─ Phase 5: STOP. Report what was done. Await your direction.

You say: "let's evanflow this — I want to add a feature that does X" │ ▼ evanflow-go (the conductor) │ ├─ Phase 0: Restate idea, scope check ├─ Phase 1: evanflow-brainstorming (CHECKPOINT: design approval) ├─ Phase 2: evanflow-writing-plans (CHECKPOINT: plan approval) │ └─ Step 2.5: parallelization check ├─ Phase 3: evanflow-executing-plans (sequential) │ OR │ evanflow-coder-overseer (parallel) │ ├─ contract with named tests + integration tests │ ├─ RED checkpoint (all coders write failing tests, orchestrator verifies) │ ├─ GREEN phase (vertical-slice TDD per coder) │ ├─ per-coder overseers (review, never fix) │ └─ integration overseer (runs touchpoint tests) ├─ Phase 4: evanflow-iterate (5x cap, Five Failure Modes pass) └─ Phase 5: STOP. Report what was done. Await your direction.

Cross-cutting: evanflow-compact runs at clean boundaries when context gets heavy.

evanflow-compact

Special-purpose skills (evanflow-debug, evanflow-improve-architecture, evanflow-design-interface, evanflow-glossary, evanflow-prd, evanflow-qa, evanflow-review) are pulled in mid-flow when relevant.

evanflow-debug

evanflow-improve-architecture

evanflow-design-interface

evanflow-glossary

evanflow-prd

evanflow-qa

evanflow-review

Repository Structure

. ├── .claude-plugin/ │ ├── plugin.json — plugin identity (name, description, version) │ └── marketplace.json — marketplace manifest (lists EvanFlow as one bundled plugin) ├── skills/ — 16 SKILL.md folders │ ├── evanflow/ │ ├── evanflow-go/ │ ├── evanflow-brainstorming/ │ ... (etc) ├── agents/ — 2 custom subagent definitions │ ├── evanflow-coder.md │ └── evanflow-overseer.md ├── hooks/ │ ├── hooks.json — auto-activated when plugin installs │ └── block-dangerous-git.sh ├── examples/ │ └── CLAUDE.md.snippet — for the manual-copy install path ├── docs/ │ └── skills-audit.md — verdict on all 38 candidate skills considered ├── README.md └── LICENSE — MIT

. ├── .claude-plugin/ │ ├── plugin.json — plugin identity (name, description, version) │ └── marketplace.json — marketplace manifest (lists EvanFlow as one bundled plugin) ├── skills/ — 16 SKILL.md folders │ ├── evanflow/ │ ├── evanflow-go/ │ ├── evanflow-brainstorming/ │ ... (etc) ├── agents/ — 2 custom subagent definitions │ ├── evanflow-coder.md │ └── evanflow-overseer.md ├── hooks/ │ ├── hooks.json — auto-activated when plugin installs │ └── block-dangerous-git.sh ├── examples/ │ └── CLAUDE.md.snippet — for the manual-copy install path ├── docs/ │ └── skills-audit.md — verdict on all 38 candidate skills considered ├── README.md └── LICENSE — MIT

Credits

EvanFlow synthesizes ideas from:

mattpocock/skills by Matt Pocock — vertical-slice TDD, deep modules, deletion test, design-it-twice, ubiquitous language, grill-me, caveman.

superpowers by Jesse Vincent — verification-before-completion, code review patterns, parallel agent dispatch, finishing-a-development-branch (the 4-option presentation).

git-guardrails-claude-code — bundled in hooks/ (script copied verbatim). Original by Matt Pocock.

hooks/

Industry research informing the design:

Anthropic's 2026 Agentic Coding Trends Report

9 Critical Failure Patterns of Coding Agents (DAPLab, Columbia)

Test-Driven Development for Code Generation (arXiv 2402.13521) — assertion-correctness findings

License

MIT. See LICENSE.

Contributing

Issues and pull requests welcome. EvanFlow is opinionated by design — proposals to add ceremony or auto-actions will be politely declined. Proposals to further reduce ceremony, sharpen rules, or add evidence-backed improvements are very welcome.