a* (autostar)

If you can measure it, you can improve it.

Soft RLVR for the masses. a* turns any measurable goal into a structured optimisation loop. You define what "good" looks like. a* runs experiments, scores the results, learns from every attempt, and converges on the best version within your budget.

No reward model to train. No environment to build. No GPU cluster to provision. Just a goal, an evaluator, and an agent that knows how to search.

---

Fastest path

If you just want to install the skill and try it once in Claude Code:

npx skills add chrisvoncsefalvay/autostar

Then invoke it in Claude Code:

/skill autostar

The skill handles onboarding, confirms the mission with you, and runs experiments within your approved budget.

---

What this is

Most artifacts live where quality is real but hard to verify fully. Code can be type-checked but not beauty-checked. Prose can be spell-checked but not tone-checked.

Traditional RLVR (reinforcement learning from verifiable rewards) needs rewards you can compute with certainty: math proofs, unit tests, formal proofs. That covers a narrow slice of what people want to improve.

a* uses verifiable-ish rewards instead. It combines hard signals (type checkers, linters, test suites) with soft ones (LLM judges, human gates). Each track has its own verifier. The system runs enough steps per lap to build statistical confidence.

The result: an optimisation loop for anything you can split into measurable dimensions. Code quality, documentation, prompt engineering, writing style, API design, accessibility.

---

See it in action

Fixing this very documentation

---

How it works

a* runs in five phases:

1. Onboarding

An interactive dialogue — never skipped, never auto-inferred. The system breaks your goal into tracks (measurable dimensions), sets up verifiers, sets constraints, agrees a budget, and gets your approval before running.

2. Pre-run preparation

Baseline measurement, tool checks, disposition library query, and final mission confirmation.

3. Execution loop

The core cycle:

Step  → mutate artifact, evaluate all tracks, ratchet (keep/revert)
Lap   → N steps with same parameter family → statistical verdict
Round → set of laps → mandatory reflection (worth pursuing? ask user? pivot?)

Each round ends with a structured reflection: Worth pursuing? Ask the user? Pivot? The system escalates when scores plateau, tracks diverge, or budget pace is at risk — always with specific questions, not vague updates.

A built-in visualiser renders live progress as an inline HTML dashboard.

4. Memory and learning

The canonical source of truth is one local persistent backend for the whole system. Short-term run logs stay append-only inside a run, while long-term memory is stored in that backend and exported as human-readable JSON/JSONL mirrors.

Within a run:

• step_log.jsonl and reflections.jsonl are append-only sources of truth
• hypothesis_stack.json, track_trajectories.json, and momentum.json are derived snapshots
• if a derived file disagrees with a log, the log wins

Across runs:

• episodes and run summaries are persisted in the backend and mirrored to JSONL
• dispositions are versioned in the backend and mirrored to JSON
• Claude.ai fallback sessions use a text-first project memory pack exported from the backend

Memory access modes: