Linksee tracks product intent across multiple AI coding agents

michielinksee opens by scoping the problem: a single `CLAUDE.md` is adequate for one product, but the approach breaks down when running several products in parallel across multiple LLMs. The author experienced two simultaneous drift vectors — human drift (forgetting decisions made on one product while working on another) and LLM drift (a Cursor session unaware of a design locked in via Claude Code). The combined effect meant no human or AI session held the full picture, and one morning the author could no longer describe any of the products in a single sentence.

Three components are described — the Product Map (chart of declared anchors vs.

The post surveys names that have emerged for this failure mode: "intent drift" (underspecified prompts cause models to fill in reasonable-but-divergent defaults), "context decay" (codebases outgrow effective context windows, causing silent contradictions), and "intent debt" (framed as the successor to technical debt, with a paper cited). The author argues that AI accelerates drift rather than causing it — speed, local-optimum optimization per task, session resets, and the N×M multiplication of products and models all remove the natural brakes a solo human developer would apply.

Linksee is presented as the response: a tool built on a "declare-don't-mine" principle, where the developer explicitly declares the decisions and constraints that define a product as anchors rather than having the tool infer them from code. Three components are described — the Product Map (chart of declared anchors vs. committed reality), `where_am_i` (per-turn position check including blast radius of a proposed change), and `drift_status` (a live alignment report with four status levels). The tool can operate in warning mode or be hardened into a `PreToolUse` gate that blocks drifting actions. The post notes that Linksee began as a memory layer but the author concluded the real value was reshaping memory into a navigational chart.

Key facts

1. 01The author runs multiple products in parallel across Claude Code, Cursor, and ChatGPT, which caused intent drift across N×M session contexts.
2. 02Three named failure modes are cited: intent drift, context decay, and intent debt (the last described as having an actual paper behind it).
3. 03Linksee uses a 'declare-don't-mine' principle — developers explicitly declare product decisions as anchors rather than having the tool infer them from code.
4. 04The `where_am_i` function shows the agent its current position on the Product Map and the blast radius of a proposed change.
5. 05The `drift_status` function reports each declared decision as 🔴 drift / 🟡 review / ⚪ held / 🔵 aligned.
6. 06A `PreToolUse` gate can optionally block a drifting action before it is executed.
7. 07Linksee started as a memory layer but the author reframed its core value as providing a navigational chart rather than raw memory.

Topics