init

# Init The one command that takes a repository from any starting state — nothing, a partial setup, or a drifted one — to a complete, wired set of context primitives: a root `CLAUDE.md` (nested ones where ownership demands), path-scoped rules in `.claude/rules/`, a decision-record archive with its index, and a maintenance section telling every future session which skill keeps each artifact true. **The failure this skill exists to prevent is a half-initialized repository.** A `CLAUDE.md` with no decision index, an ADR directory nothing references, a setup done once with no note saying how it is maintained — each looks finished and quietly stops paying off. > **Scope boundary.** This skill owns the **sequence and the wiring** — which > skill runs when, and the maintenance section in `CLAUDE.md`. Every rule about > the artifacts themselves belongs to the owning skill: > [`managing-claude-context`](../managing-claude-context/SKILL.md) for the memory > layer, [`managing-decision-records`](../managing-decision-records/SKILL.md) for > record format and lifecycle. This skill loads their rules; it never restates > them. ## When to Use - A user runs `/tsh-core:init` — this skill is user-invoked only and never loads by description match - Setting up a repository for the first time, or re-running after a template or fork left context files describing the wrong project - A repository whose setup is partial or drifted — running init again is repair, not a rewrite ## Applicability and Precedence Everything `managing-claude-context` and `managing-decision-records` state about precedence applies unchanged here: an existing `CLAUDE.md` is revised in place, an existing ADR convention and location win over the defaults, and Claude Code's own documentation outranks all of it. Init adds one thing — it is safe to run repeatedly, and a second run must converge to "nothing to do", never re-create what the first run built. ## Explicit Exclusions This skill does not: - restate, soften or override any rule owned by the skills it orchestrates — with one addition of its own, marked in the rules table: the decision index is always left in place - write or edit product code, configuration logic, tests, or infrastructure - invent a convention or a decision — it records what exists and asks about the rest - write a decision record's body, or set a status — it proposes candidates and hands the writing to [`managing-decision-records`](../managing-decision-records/SKILL.md) - install plugins, configure permissions, or write outside the repository ## Non-negotiable Rules | Severity | Rule | | --- | --- | | MUST | Execute `managing-claude-context`'s procedure in full as Step 1, starting with its inventory. Init never writes before that audit has run. | | NEVER | Overwrite an existing `CLAUDE.md` wholesale. Revise in place and report what changed. (Owned by `managing-claude-context`; restated because it blocks.) | | MUST | Read the owning skill's `SKILL.md` before touching its artifact — never work its area from memory. | | MUST | Leave the repository with a decision index — `docs/decisions/README.md`, or the existing archive location — even when no decision has been recorded yet. An empty indexed archive is the correct starting state. (Init's own rule, not the owner's: `managing-claude-context` wires an index only where records exist or are wanted; a bootstrap counts as wanting one.) | | NEVER | Invent a decision record to fill an empty archive. Propose candidates in the report; a human confirms and `managing-decision-records` governs the writing. | | MUST | Wire the maintenance section into root `CLAUDE.md` exactly per `./references/wiring-tsh-core-skills.md` — conditional phrasing, every path in backticks, never a leading `@`. | | MUST | End with a report of what was created, what was repaired and what was left alone, and flag the result as a proposal for human review. | ## Reference Loading | Reference | Load when | Covers | | --- | --- | --- | | [managing-claude-context](../managing-claude-context/SKILL.md) | Step 1, before writing anything | The memory layer: inventory, bootstrap vs. drift audit, layer choice, `CLAUDE.md` and `.claude/rules/` rules, decision-index wiring | | [managing-decision-records](../managing-decision-records/SKILL.md) | Step 2, when a record is proposed or the archive's shape is in question | Record format, numbering, status vocabulary and lifecycle | | [Wiring tsh-core skills](./references/wiring-tsh-core-skills.md) | Step 3, before writing or repairing the maintenance section | The canonical maintenance block, why its phrasing is conditional, repairing a stale variant, adapting paths | ## Procedure **Step 1 — Establish the memory layer.** Read [`managing-claude-context`](../managing-claude-context/SKILL.md) and execute its procedure in full. Its own first step inventories what loads and branches — nothing found → bootstrap, something found → drift audit — and its Step 5 wires the decision index. Run it through its self-check even when the repository looks already set up: repair is this command's job as much as creation. **Step 2 — Ensure the decision archive exists.** If Step 1 left the repository without one, create the index — default `docs/decisions/README.md`; an existing archive location wins — with the binding note and an empty four-column table, per the indexing reference `managing-claude-context` loads. Do not backfill records: when the codebase clearly embodies undocumented decisions, list them as candidates in the final report and point at [`managing-decision-records`](../managing-decision-records/SKILL.md) instead of writing them. **Step 3 — Wire the maintenance section.** Read [`wiring-tsh-core-skills.md`](./references/wiring-tsh-core-skills.md), then add the section to root `CLAUDE.md` — or repair the existing variant in place. It counts against the 200-line budget like every other line. **Step 4 — Verify and report.** Run `/context` and confirm everything written or repaired loads, or is scoped to load on read. Report three lists — created, repaired, left alone — plus any decision candidates from Step 2, and close by flagging the setup as a proposal for the team to review, not a finished artifact. ## Self-check Before Handoff ```text - [ ] managing-claude-context's procedure ran in full, inventory first, self-check included - [ ] No existing file was overwritten wholesale; every revision is reported - [ ] Every artifact was written under its owning skill's rules, loaded not recalled - [ ] A decision index exists, at the default or the repository's own location - [ ] No decision record was invented; candidates are listed for a human - [ ] The maintenance section is in root CLAUDE.md — conditional, backticked, no bare @ - [ ] /context confirms the result loads - [ ] The report lists created / repaired / left alone and calls the result a proposal ```