aligned human-ai

AHA_

/aha runs the full cycle and remembers.

Turn a fuzzy agentic ask into typed questions, an alignment ledger, a premortem, a cold-runnable prompt, a handoff contract, and a return-packet debrief. Use the standalone skills when you only need a quick check.

bash — install
# Claude Code
$ /plugin marketplace add omerakben/aha
$ /plugin install aha@aha
# Cursor (or manual)
$ git clone https://github.com/omerakben/aha && ./install.sh
Get started → GitHub

## cycle

1
/ahaOne command loads memory, detects resume or debrief intent, and starts the full cycle when the task is handoff-worthy.
↓
askTyped questions with why it matters, if A, and if B. The run writes before it stops.
↓
alignLedger with restate, decisions, assumptions, unknowns, confidence basis, and status.
↓
critiquePremortem on execution failure, missing evidence, unsafe assumptions, and tool risks.
↓
optimize + handoffStandalone prompt, eval rubric, human packet, handoff_contract YAML, and required return_packet block.
↻
debrief post-runPacket versus return packet, AHA delta, and proposed rules. Promotion still needs approval.

## why

▸
Strict stop gates. The model asks, aligns, and critiques before handoff. Each STOP writes state first.
▸
Private memory. .aha/state.yaml tracks sessions, contracts, return packets, and proposed rules. .aha/ is gitignored by default.
▸
Agent-to-agent contract. The generated prompt requires a return_packet so the debrief can compare acceptance checks against reality.
▸
Quick checks remain. Use /ask-me, /align-me, /critique-this, /optimize-prompt, or /debrief when one step is enough.

## quickstart

aha — usage 
/aha

Task: triage failed webhook deliveries and suggest retries.
  1. Run /aha. If the task is small, it redirects to the right quick check.
  2. Answer typed questions, then reply continue after the premortem STOP gate.
  3. Hand the packet to the executing agent and require the filled return_packet.
  4. Paste the return packet into /aha for debrief and rule proposals.