Achieved a result with your AI agent? Save, reuse, share the experience.

Closed a deal. Shipped an app. Set up a Google Analytics process? Any incremental win — write down how you got there:

$ git clone github.com/anthroos/openexp && ./setup.sh

View on GitHub →

FIG.01 / TRAJECTORY REUSE

DEAL #1 · CLOSED day +57 · grade 1.0

DEAL #2 · LIVE cursor at day +12

01
Record.

The AI session trace.

Prompts, reasoning, skills, scripts, tool calls. Day by day. Storage-agnostic.
02
Share.

Open-source. Or with your team.

Your exp drops into ~/.claude/skills/. Their Claude cites your day.
03
Replay.

Hand it to your agent.

Next time it hits a similar task, it picks the matching exp and follows the trajectory that led to the result. Your agent stops reasoning from scratch.

All open source, MIT.

FIG.02 / EXP CONTENTS

Four files. That's the whole exp.

meta.yaml

facts only — id, outcome, grade

trajectory.anonymized.yaml

raw timeline, day by day

README.md

for humans

SKILL.md

for Claude

~/.claude/skills/openexp:author:pack-name/

Benefits

Replay only what worked. Exps exist only for paths that reached a result with a grade. Proof, not advice.
Cheaper tokens. Replay an Opus arc on Sonnet or Haiku. Same decisions, fraction of the cost.
Open-source. MIT. Build on others' arcs. Your authorship sticks — like citation on a paper. No vendor lock-in, no cloud.
Local-first. Your data stays with you. Runtime on your machine. No telemetry, no third-party server, no API key required. Your repo, your terms.

Existing exps

Five exps already published.

Today You describe the situation and reference the exp — by name or by tag — and your Claude pulls from it. Roadmap Auto-firing on situation patterns alone, no naming required.

MIT·open source·install via CLI·github.com/anthroos/openexp

Achieved a result with your AI agent? Save, reuse, share the experience.

Record.

Share.

Replay.

Four files. That's the whole exp.

Five exps already published.