Setting up a repository

A repository is not finished at git init. The complete unit is: the local
scaffold and a first commit, the GitHub repository created under
alternative-intelligence-cp and pushed, and a description and topics set. Do
all three in one pass. Repository creation is never delegated (W-26): the
orchestrator or the author runs this skill, because it is outward-facing and
it edits the registry.

1. Claim the names first

In the same commit that starts the work, add a row to LIBRARIES.md (a
library) or APPS.md (an application) claiming:

• the package name — the ecosystem's n-prefix convention: ntui,
  nparse, nregex, nsockets, ntime
• the decision prefix — two letters, distinct from every other, and not
  colliding with the single letters used as rule prefixes inside
  specification documents, nor with D- (the compiler's, permanently)

2. The scaffold

src/            the source; a README per subdirectory naming what lives there,
                which specification governs it, and which cycle builds it
tests/          probe, conformance, unit, fixtures, plus this project's own
                oracle stage — named for the property it checks, not for
                another project's "golden"
harness/        the Python build and test runner, until npkg can
tools/          generators; everything they emit is committed and
                regeneration-checked
examples/       built AND run by the harness, so a broken example is a red run
docs/           written at 1.0
meta/           specs/, DECISIONS.md, OPEN_QUESTIONS.md, roadmap/, research/
                (with CURRENCY.md — skills/research/SKILL.md §7), scratch/

.gitignore — build output, *.o, *.ll (negating any committed
fixture), __pycache__, *.log, generator inputs that are large and
reproducible, editor residue, .internal, .claude/settings.local.json. Plus
an explicit "NOT ignored, deliberately" block naming the generated tables
and fixtures with the reason for each.

In a workbench repository the ignore is /*/, which ignores every
top-level directory including dot-directories. Anything the workbench
itself owns must be un-ignored by name or it vanishes silently.

nitpick.toml — the compiler's schema exactly: [project] with target,
[build], [toolchain] pinned to the exact patch release with the four flag
lists, an empty [dependencies], and a [[test]] table that starts empty
with a comment saying each cycle adds its own. An entry naming an empty
directory is a suite that reports green while checking nothing.

CLAUDE.md — the read order, the non-negotiables, and a pointer to the
board (../BOARD.md from a library, ../../nitpick-libs/BOARD.md from an
app). Count the levels: CLAUDE.md sits at the repository root.

3. GitHub, in the same pass

gh repo create alternative-intelligence-cp/<name> --public --source=. \
  --remote=origin --push --description "<what it is, in one sentence>"
gh repo edit alternative-intelligence-cp/<name> --add-topic … (up to 20)

Topics are for discoverability, not decoration: the domain terms somebody
with this problem would actually search, plus the ecosystem set (nitpick,
nitpick-lang, safety-critical, zero-dependency, formal-verification)
so the family is findable together.

Scan before pushing. These repositories are public:

python3 ${CLAUDE_SKILL_DIR}/../check/scripts/check_refs.py .

which reports any absolute home path or credential in a tracked file.

4. Then plan, before any code

The order is specs → decisions → open questions → cycle map → cycle 0.0
execution-grade → code. ../PLAYBOOK.md §1 is the full statement, and §12 is
what a finished plan looks like.

Cycle 0.0 is always the language probes, because a construct that parses
is not a construct that works — and a probe that fails changes the design,
which costs a day now and a cycle later.