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.
Read GUIDE.md for the full walkthrough of how it thinks through a session and what it will and won't write down.
- 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.
- 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). It also degrades gracefully — see GUIDE.md — to a standalone written report if no such memory system is present.
- Nothing else. No API keys, no external services, no dependencies.
Copy SKILL.md into your assistant's skills folder.
Claude Code:
mkdir -p .claude/skills/hindsight
cp SKILL.md .claude/skills/hindsight/SKILL.mdOr install it once at the user level so it's available in every project:
mkdir -p ~/.claude/skills/hindsight
cp SKILL.md ~/.claude/skills/hindsight/SKILL.mdTrigger it in conversation any time you want a session reviewed:
"Get some hindsight." "What should we improve?" "Review today and notate lessons." "What did you learn today?" "Save what's worth remembering from this session."
It's deliberately on-demand only (disable-model-invocation: true in
the frontmatter) — a hindsight pass writes persistent memory, so timing
stays your call, never something the assistant decides to run on its own.
The "In this environment" section of SKILL.md names Claude Code's memory
layout specifically (~/.claude/projects/<project-slug>/memory/,
MEMORY.md as the index). If your assistant or harness stores persistent
memory somewhere else, or under a different format, replace that section
with the equivalent path/convention for your setup — the rest of the skill
(steps 1–6, the definition of done, the scope boundary) is assistant-agnostic
and shouldn't need to change.
MIT — see LICENSE.
Jeffrey Smith Email: jeffrey@efficientstreet.com YouTube: https://youtube.com/@JeffreyEntrepreneur Website: https://efficientstreet.com