A skill library that treats big features as fog of war

What it does

A personal library of ~25 domain-agnostic agent skills — prompt-level playbooks, not code — that chain into a full software pipeline: explore an idea, write a spec, implement it, review the result, and run autonomous experiments. Each skill is a folder of markdown you drop into any skills-compatible harness (Claude Code, Codex, Cursor, opencode, and 70+ others). Every skill also stands alone if you just want, say, a decent code review.

The interesting bit

The core move is slicing work into independently verifiable pieces rather than mere tasks — the README calls the unknown parts “fog of war” to be mapped and recursively re-sliced. The spec is a living document, re-planned mid-implementation when reality disagrees. And instead of reviewing diffs, you review a choices.md ledger of every decision the agent made where the spec was silent, ranked least-confident first. That’s a genuinely different review surface.

Key highlights

• Full loop: /explore-unknowns → /write-spec → /goal /implement-spec → choice-audit; the README claims one unattended Codex run lasted 1d 16h on a single goal
• auto-research runs progressive experiments and outputs a parameter-effect map, not just a “best” result
• Visual review is taken seriously: screenshot critique with an unprimed subagent, plus diff tooling that flags flat or misframed captures
• Meta-skills for authoring and evaluating the skills themselves, with blind runs and a separate judge
• Harness-agnostic: one copy in .agents/skills, symlinked so Claude Code and Codex read the same files

Caveats

• It’s one person’s personal library — the “1d 16h” run is a single anecdote, not a benchmark
• Some skills are environment-specific (preview-shots is macOS-only; renderer assumes WebGPU/three.js work)
• The value depends on your harness actually supporting skills and loop mode

Verdict

Worth a look if you’re trying to make long autonomous agent runs trustworthy — the choices-ledger idea alone is stealable. Skip it if you want a framework or code; this is opinionated prompt craft, and it works best if you buy the whole spec-driven loop.