workinggenius: Instructions file for Claude Code

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

What this repository is

The Working Genius plugin itself: a Claude Code plugin — 24 skills under skills/ and five subagents inside the skills that spawn them, distributed through the plugin marketplace (.claude-plugin/). The skills are prose, plus three things only Claude Code reads that ride inside the same folders (below); nothing here is held back for another client (user ruling, 2026-09-05). There is no runtime, no build and no test suite — what the skills claim is measured by an eval suite, evals/ (below); the one script is the format's instrument, skills/genius-file/measure.py, which counts and never decides, and whose one writing command, append, writes only the log entry it is handed, where the caller named — a fork must be able to carry the whole thing as plain folders, and still can. Dogfood work files (.genius/) are gitignored and never shipped.

Commands

There is nothing to build or run. Four checks matter when editing skills:

• Count characters, never bytes. python3 skills/genius-file/measure.py count <file> is the instrument and this one-liner is what it does; any number that enters a skill or a record comes from one of these and not from wc -m (which counts bytes under most non-interactive locales) or awk length():

  python3 -c "import sys;print(len(open(sys.argv[1],encoding='utf-8').read()))" <file>
  

• Relative links between skill files must resolve (agents/inventor.md, ../tenacity/agents/reviewer.md, FILE-FORMAT.md, ../genius-file/BACKLOG-FORMAT.md; a work file's own links in a template, log/… and CONTRACT.md, are skipped). A one-off check:

  python3 -c "import re,os,glob;print([(f,t) for f in glob.glob('skills/**/*.md',recursive=True) for t in re.findall(r'\]\(([^)#]+)(?:#[^)]*)?\)',open(f,encoding='utf-8').read()) if not t.startswith('http') and '<' not in t and t!='CONTRACT.md' and not t.endswith('.log.md') and not t.startswith('log/') and '/log/' not in t and not os.path.exists(os.path.join(os.path.dirname(f),t))])"
  

• Frontmatter must parse, and the instrument must not fail. The preloads and the judge's command hooks live in YAML frontmatter, and a stray indent drops the whole block without a word; measure.py is injected into four commands and a skill whose injected command fails is aborted whole, so it exits 0 on an empty directory and on a malformed snapshot; stop-judge.py blocks a stop only on a clear verdict, so it exits 0 on garbage and on an absent claude. One check for all three:

  python3 -c "import yaml,glob;[yaml.safe_load(open(f,encoding='utf-8').read().split('\n---\n')[0][4:]) for f in glob.glob('skills/*/SKILL.md')+glob.glob('skills/*/agents/*.md')];print('frontmatter ok')" && python3 skills/genius-file/measure.py status >/dev/null && printf 'x' | python3 skills/genius-file/stop-judge.py && echo 'instruments ok'
  

• The skills have an eval suite, evals/, run by claude plugin eval (2.1.269 and later): one case per claim the skills make that a fresh session can be caught breaking — an ordinary request never enters the flow (no-hijack), a question about work in flight is answered from the instrument and invents nothing (nothing-in-flight), an idea handed to the flow is interviewed and not built (genius-starts-wonder), discernment on a work with no paths runs Invention first — more than one inventor, the paths in the log — and presents one commitment with its kills, not a menu (discern-invents-first, scaffolded), one slice is built red before green with its evidence in the log (enable-red-before-green, scaffolded), and a close-out whose log records a criterion as passed that fails on the code runs it fresh and does not mark the work done (tenacity-reruns, scaffolded). Each case is a prompt as a person would type it, graders that cost nothing (tool_used, tool_order, regex, file_exists) and one judge rubric; results land in evals/results/, gitignored, and --keep-temp keeps a run's trace. Two facts about the runner, measured 2026-09-23: a prompt that is a slash command (/genius, /enable demo, slice 2) is not run — zero turns, no error — so the user-typed commands are outside an eval's reach and the cases say what a person would say; and a case that grants Bash, Write or Edit runs inside the runner's sandbox, where a machine that cannot make user namespaces kills every shell call (a root cloud container did, and scored the enable case 0.57 for it). The smoke case runs on haiku and the flow cases on sonnet, because haiku did not fire genius-file on the in-flight question at all (measured), a fact about triggering on a small model and not a reason to loosen the case:

  claude plugin eval . --tag smoke --runs 1 --ablation none --model haiku --allow-tools Bash Write Edit --trust-plugin --no-publish
  claude plugin eval . --tag flow --runs 1 --ablation none --model sonnet --scaffold --allow-tools Bash Write Edit --trust-plugin --no-publish --max-cost-usd 10
  

  Where each case stands (the run-by-run history, every miss and what fixed it, is in the commit messages, not a results file (user ruling, 2026-10-03): git log -- CLAUDE.md evals/):
  • no-hijack 1.0 on haiku, every run since 2026-09-23.
  • nothing-in-flight and genius-starts-wonder 1.0 on sonnet at v0.33.0; genius-starts-wonder once scored 0.67 at v0.31.0, a first round whose questions carried no recommendation.
  • discern-invents-first 0.83, 0.88, then 1.0 at v0.33.0 on sonnet, $1.96–2.42 a run; the misses fixed the attacker and inventor briefs, the graders, and invent, which now names log/invention.md.
  • enable-red-before-green 1.0 in the runner on the owner's Mac, 2026-09-30, sandbox working, seven graders of seven; one run, not yet a rate.
  • tenacity-reruns by hand only: the criterion re-run fresh, errata loaded before the correction, the work closed on a fix shown fresh; the case allows 80 turns because a full close-out needs them.
  • After the prose pass of 2026-10-03 (every sentence in skills/ over 60 words split, no rule changed): no-hijack 1.0 (haiku), nothing-in-flight 1.0 and genius-starts-wonder 1.0 (sonnet) in the runner; by hand, enable-red-before-green three runs, each red before green, ## slice-2 with its command lines, S2 checked, the reviewer, one commit, a close naming its evidence ($0.28–0.35), and tenacity-reruns two runs, each re-running the criterion, loading errata, fixing the code and showing hello, Ada fresh before done ($0.60–0.61).