intent

Code is disposable. Intent is the asset.

AI generates code. You can regenerate it anytime. What you can't regenerate is the sequence of decisions that shaped the software — the problems you identified, the constraints you set, the directions you chose. intent captures that sequence from your AI coding sessions and turns it into a reproducible, reviewable history.

The problem

When you build with AI agents, your codebase changes fast. But the why behind each change gets lost in chat logs. A month later, you open the project and ask yourself: "Why did I build it this way?"

Git log shows what changed. Commit messages try to explain why. But the real story — the trial-and-error, the constraints, the rejected approaches — lives in your conversation with the AI. And that conversation disappears when the session ends.

What intent does

intent reads your Claude Code session logs and extracts a structured timeline of human decisions.

$ intent log
  001  2026-02-15  認証機能の追加
  002  2026-02-15  ログインバグの修正
  003  2026-02-16  /commit スキル実行

Each step captures:

• prompt — what you asked the AI to do
• reasoning — why you chose this approach
• outcome — what was achieved
• friction — what went wrong along the way

Why this matters

Reproducibility

A prompt history is a recipe. If you have the full sequence of intents that built a piece of software, you (or someone else) can rebuild it — or build a better version by tweaking specific steps.

Trust signal

Software built through structured flows (plans, tested skills, slash commands) is more reliable than software built through free-form conversation. intent makes that distinction visible:

  001  2026-02-15  認証機能の追加              [plan]
  002  2026-02-15  /commit スキル実行          [skill]
  003  2026-02-16  パフォーマンス改善          [feature]

Steps tagged [plan] or [skill] were executed through structured, repeatable processes. This is a quality signal.

Team context

When someone joins a project, they don't just need the code — they need to understand the decisions behind it. intent gives them a readable timeline of why the software evolved the way it did.

Quick start

npm install -g intent-log

cd your-project
intent init
intent collect
intent log

Requirements

• Node.js 22+
• Claude Code CLI installed and authenticated

Commands

| Command | Description | |---|---| | intent init | Initialize .intent/ in your project | | intent collect | Extract intent from Claude Code session logs | | intent log | Show the timeline of intent steps | | intent show <step> | View details of a specific step | | intent replay | Export prompt history in a reproducible format | | intent rm <step> | Remove a step | | intent reset | Clear all collected steps |

How it works

1. You work with Claude Code as usual
2. intent collect reads the session logs from ~/.claude/projects/
3. Each session is analyzed:
   • Plan executions are parsed directly from the plan markdown (no AI call needed)
   • Skill/slash command sessions are auto-detected and tagged
   • Regular conversations are summarized via Claude to extract the human intent
4. Privacy filter removes sensitive data (API keys, tokens, short/trivial prompts)
5. Results are stored as numbered markdown steps in .intent/steps/

Replay

Export your intent history as a readable, reproducible document:

# Output to stdout
intent replay

# Save to file
intent replay -o replay.md

# Specific range
intent replay --from 3 --to 10

The output contains each step's prompt, reasoning, and outcome — everything needed to understand or reproduce the development process.

Step format

Each step is stored as a markdown file with frontmatter:

---
step: 1
title: "認証機能の追加"
session: "e527b849-..."
timestamp: 2026-02-15T10:30:00Z
tags: [plan]
---

## prompt

ログイン機能を追加。OAuth2対応で、Google/GitHubをサポート。

## reasoning

既存のセッション管理を活かしつつ、OAuth2プロバイダを抽象化して追加しやすくした。

## outcome

Google/GitHubログインが動作。プロバイダ追加はadapter実装のみで可能。

## friction

GitHub OAuthのcallback URLがlocalhostで動かず、ngrokで回避。

Privacy

intent never sends raw session logs to external services. The privacy filter runs locally before any summarization:

• API keys, tokens, and secrets are stripped by regex patterns
• Short/trivial prompts (< 20 chars) are excluded
• Agent system instructions are filtered out
• Custom patterns can be added in .intent/config.toml

[filter]
exclude_patterns = [
  "(?i)(api.?key|token|secret|password)",
  "^(はい|ok|yes|y|n|no)$",
]
min_prompt_length = 20

Roadmap

• [ ] Friction auto-extraction — Detect errors, retries, and failed tool uses to populate friction automatically
• [ ] Step dependencies — Track relationships between steps (e.g., "Step 5 was a bugfix for Step 3")
• [ ] Trust scoring — Score each step based on how it was created (plan > skill > free conversation)
• [ ] intent diff — Compare intent history between two projects or branches
• [ ] intent fork — Import someone's intent history and replay it in your environment
• [ ] Multi-agent support — Cursor, Windsurf, and other AI coding tools

License

MIT