Hindsight - Making AI Sessions More Efficient · Useful AI • AI For Business

Hindsight - Making AI Sessions More Efficient

I am starting to dedicate part of my Thursdays to R&D. I came up with this new skill today to help keep AI learning from all chat sessions to make it more effective with every new session. I call the skill Hindsight. At the end of your sessions, just /hindsight

## About Hindsight

A self-improvement skill for AI coding assistants. At the end of a session, it reviews everything that actually happened — the tasks, the corrections, the dead ends — pulls out the lessons that would genuinely change how future work gets handled, and saves the durable ones as persistent memory. Not a journal of what got done; a filter for what's worth remembering.

## What you get

- **One trigger, one pass** — "get some hindsight" (or equivalent phrasing) reviews the whole session end-to-end, not just the last exchange.

- **Root causes, not symptoms** — for anything that took multiple attempts, it asks what single upstream fix would have prevented the whole chain, instead of just noting that a mistake happened.

- **A real filter, not a transcript** — one-off content specifics get discarded; only lessons that would change how a *future* task gets handled get kept. A clean session gets told "nothing to report," not padded filler.

- **Updates in place** — before writing a new memory entry, it checks whether an existing one already covers the topic and extends that instead of creating a near-duplicate.

- **Never redoes work** — strictly read-only against the session's actual tasks. Its only writes are to memory (and whatever indexes that memory).

- **Portable** — no project-specific content baked in. It works the same way regardless of what the session was actually about.

- **Pays for itself in tokens, not just tidiness** — a lesson written down once means a future session doesn't re-spend tokens rediscovering it through the same trial and error, re-litigating a question that was already settled, or repeating a correction you already made. The saving compounds: the more sessions an assistant runs, the more that accumulated memory shortens the average one.

## Requirements

- An AI coding assistant with some form of persistent, file-based memory across sessions (this was built for and tested against **Claude Code**'s per-project memory folder).

- Nothing else. No API keys, no external services, no dependencies.

https://github.com/EfficientStreet/hindsight

0 comments

Hindsight - Making AI Sessions More Efficient