Tide — AI skill for Claude Code

TIDE

A pattern for letting an LLM agent get better at *noticing* over time.

This is an idea file. It is designed to be copy-pasted into your own LLM agent (e.g. Claude Code, OpenAI Codex, OpenCode, etc.). Its goal is to communicate the high-level idea; your agent will fill in the specifics as it works with you. It assumes nothing about the kind of work you do or the kind of files you keep.

The core idea

Most agents start every task from a blank slate. They re-derive the same hard-won insights again and again — that a "successful" API call can still return empty content, that an upload tool reporting `200 OK` doesn't mean the remote asset is actually correct, that a relative path breaks the moment you hand work to another process.

TIDE (**T**emplate-**I**nformed **D**etection **E**xtraction) is a small loop that fixes this. After a task succeeds, the agent extracts the *detection patterns* it used — not the solution, but the **sequence of signals it checked to find the problem**. These become reusable templates. On the next task, the agent matches its situation against the accumulated templates and runs their checks proactively.

The more work you do, the more templates accumulate, and the wider the agent's detection coverage on new tasks becomes.

past work  → [extraction prompt] → template store
                                          ↓
new task   ←  [application guide]  ←  matched templates
                                          ↓
task done  → new pattern noticed   →  template store  (loop)

A template is a detection, not a solution

This is the whole discipline. A template captures *how a problem was found*, never *how it was fixed*.

• ✅ "Check signal X first → then compare against Y → then verify Z"
• ❌ "Install tool X and run command Y"

Fixes are specific to a tool, a version, a moment. Detections generalize: the *shape* of "a success response that hides an empty body" recurs across wildly different domains. Store the shape.

Temp