intentracker

Track project intent across AI agent sessions.

The Problem

When working on a project across multiple agent sessions — whether in Claude Code, Cursor, Copilot, or any other AI coding tool — the agent starts each session with no understanding of what you're trying to build or why you made past decisions. You end up repeating yourself:

• "Remember, we decided not to use a config file — everything should be CLI flags."
• "The goal is to keep this lightweight, no Electron."
• "We already discussed this — we chose Puppeteer over WeasyPrint for PDF rendering."

Existing memory tools (CLAUDE.md, .cursorrules, Memory Bank, etc.) track operational context — build commands, code patterns, conventions. None of them systematically track intent — what should happen next and why.

| | Operational memory | Intent tracking | |---|---|---| | Records | What happened | What should happen | | Source of truth | Code | User's words | | Goes stale when | Code changes | User changes their mind | | Used for | Resuming work | Resolving ambiguity, guiding decisions |

intentracker fills this gap. It maintains a ## Intent Tracking section in your AGENTS.md with inline maintenance rules and four subsections:

• Current Goals — what you're building (snapshot, rewritten on change)
• Constraints — technical boundaries and requirements (snapshot)
• Key Decisions — why the project looks the way it does (append-only log with dates)
• Open Questions — unresolved decisions (maintained actively)

The maintenance rules are written as visible markdown — not HTML comments — so any agent that reads the file knows how to keep intent up to date.

Usage

Quick Start (any agent)

# Create AGENTS.md with intent tracking + symlink CLAUDE.md
pnpx intentracker init

# Also inject rules into your agent's config
pnpx intentracker init --agent cursor
pnpx intentracker init --agent cursor,windsurf,copilot

# Display current intent
pnpx intentracker show

Supported agents: cline, codex, copilot, cursor, windsurf

The --agent flag injects intent-awareness rules into the agent's config file (.cursorrules, .windsurfrules, .github/copilot-instructions.md, .clinerules). The AGENTS.md file itself is always created — most agents read it natively.

init also handles CLAUDE.md automatically:

• If CLAUDE.md doesn't exist, it creates a symlink pointing to AGENTS.md — so Claude Code loads the same intent without duplication.
• If CLAUDE.md already exists as a real file, the intent section is injected into both files.

Without any CLI

You don't need this package at all. Just add this to your AGENTS.md:

## Intent Tracking

This section tracks project intent. When the user changes project
direction — new goals, constraints, or decisions about approach —
update the subsections below following these rules:

- **Current Goals** and **Constraints** are snapshots: rewrite to reflect current state
- **Key Decisions** is append-only: add `[YYYY-MM-DD] description` entries, never remove
- **Open Questions**: add when raised, remove when resolved (move to Key Decisions)

### Current Goals
- Build a CLI that converts markdown to PDF with custom templates

### Constraints
- No Electron — keep it lean, pure Node
- Must work on Node 18+

### Key Decisions
- [2026-04-14] Chose Puppeteer over WeasyPrint for PDF rendering
- [2026-04-15] Decided against config file — all options via CLI flags

### Open Questions
- Should --watch use chokidar or fs.watch?

The inline rules tell any agent how to maintain the section. The CLI just scaffolds this for you.

Claude Code Plugin

For Claude Code users, the plugin provides automatic hooks that load intent on session start and reconcile it before commits — no manual maintenance needed.

/plugin install intentracker

Commands:

• /intent:show — Display current intent
• /intent:update — Guided intent update
• /intent:init — Initialize intent tracking

Hooks:

• Session start — Loads intent silently, uses it as authoritative context
• Pre-commit — Detects direction changes in the conversation and updates the intent section

How It Works with Other Tools

intentracker complements existing memory systems — it doesn't replace them:

• CLAUDE.md / Auto Memory handles operational knowledge (build commands, code patterns)
• .cursorrules / .windsurfrules handles behavioral instructions
• intentracker handles directional context (goals, constraints, decisions)

The AGENTS.md file is plain markdown with visible instructions. Any agent that reads project files will see the rules and know how to maintain intent. The Claude Code plugin adds automation on top; the file works without it.

License

MIT License © Anthony Fu