Skill: config

The settings file has many creators and one editor: this one. Six pipeline skills (interview / survey / roadmap / scaffold / specify / implement) — and this skill too — write .claude/sdd.local.md with documented defaults, so the pipeline always finds a real file. None of the six offers to change a value. That offer lives here, and only here.

Everything about the file itself — the template, the create procedure, the .gitignore patch, the key semantics, the edit rules — is canonical in ../_shared/settings-file.md. This skill is the conversation over it.

Owner

Whoever owns the repo's setup — usually the lead who runs the pipeline first, once per project.

Inputs

• Nothing is required. No gate, no upstream artifact: a bare repo is a valid starting point.
• (Optional) an existing .claude/sdd.local.md — read, diffed against the defaults, patched in place. Never regenerated.

Protocol

1. Ensure the settings file (first thing, after this skill's own gate). If .claude/sdd.local.md is absent, create it now from the canonical template — documented defaults + the self-documenting body — and patch .gitignore; if it exists, read it and never overwrite. The one procedure lives in ../_shared/settings-file.md. Creating is unconditional; changing values is only ever offered by config. Say one line: «.claude/sdd.local.md created with documented defaults — /sdd:config to tune it». (Here that line is the opening move, not a footnote: the user has a working config before the first question.)

2. Read the current state. Report in three lines: which keys differ from the documented defaults (with both values), which unknown keys and inline comments are present (they are preserved verbatim — say so), and whether the frontmatter parsed. A malformed frontmatter is never repaired silently — print the lines that failed to parse and offer exactly two ways forward: rewrite from the template (naming what will be lost) or stop so the user fixes it by hand.

3. One forking question: defaults or tune? Ask it per ../_shared/ask-style.md. «Stay on defaults» → skip to step 7 and print the handoff; the file is already on disk and working, so this is a legitimate exit, not an abort.

4. Guess the host, with the evidence. Which tool is running this session decides whether three keys are even usable. Rank the signals and name them out loud — the guess is never written silently, and contradictory traces are reported as contradictory. Signals + the always-confirm rule → ./references/detection.md.

5. Ask the six grouped questions in three AskUserQuestion calls. Questions are grouped by decision, not by key — a person tunes «how strict are the gates», not gate_vet. The six: host confirmation · model tier · strictness and gates · execution and commits · interview depth · document language. Together they cover 21 of the 25 keys. The four cmd_* keys are not asked: empty means the detection cascade reads the repo's own Makefile / package scripts / manifests, which is the better answer than a pinned command that rots. Verbatim wording, the options, and the question → keys map → ./references/interview.md. Model tier is asked directly rather than probed, because the session cannot enumerate the models an account actually has.

6. Write the confirmed keys, key by key. Patch only the lines whose values were confirmed; preserve comments, key order and unknown keys exactly. Print a old → new line per changed key, and separately list what was derived rather than asked — a non-Claude host clamping team_mode: false / workflow_mode: off / max_parallel_agents: 1 (TeamCreate and Workflow exist only in Claude Code), and the cmd_* keys left empty for autodetect. A value the user did not choose must never read as if they had. Nothing else on disk is touched.

7. Structural self-check + handoff. Run the structural self-check per ../_shared/self-check.md — re-read the file from disk and verify 6 items: (1) .claude/sdd.local.md exists (test -f); (2) its frontmatter parses and every confirmed key holds the confirmed value; (3) every key present before the run is still present — none dropped, including unknown ones; (4) every inline comment that was there is still there; (5) .gitignore contains .claude/*.local.md and .worktrees/, each exactly once; (6) no file outside .claude/sdd.local.md + .gitignore was modified (git status --short). Fix + re-check ≤2 cycles; surface anything unresolved. Then propose commit chore: gitignore sdd settings only if .gitignore actually changed — the settings file is gitignored and has nothing to commit. Then emit the stage-handoff block per ../_shared/handoff.md — What I did (the changed keys + the self-check line) + Review (.claude/sdd.local.md) + Run next (utility variant: /clear optional, resume your backbone stage — name the likely one).

Definition of Done

• .claude/sdd.local.md exists and parses, whichever branch the run took (defaults or tuned).
• Every key the user confirmed holds the confirmed value; every key they didn't confirm is byte-identical to what was there before.
• Unknown keys, key order and inline comments survived the run.
• Every derived value (host clamp, autodetect) was named out loud, separately from the asked ones.
• .gitignore covers .claude/*.local.md and .worktrees/.
• The 6-item check in step 7 is this skill's structural self-check; its result is one line in the handoff.

Anti-patterns

• Asking before creating. The file lands with documented defaults first; a user who answers nothing still ends up better off than before the run.
• Regenerating the file from the template because a key looked stale. That is how a hand-written comment or a hand-added key disappears. Patch key by key.
• Writing a detected host silently. Detection is evidence for a question, never an answer.
• Offering team_mode / workflow_mode on a non-Claude host. They clamp, and the clamp is stated — an option that cannot work is worse than no option.
• Asking about cmd_*. Empty = autodetect, which is the recommended answer; asking invites a pinned command that goes stale.
• A settings question inside another skill. The six pipeline skills create and read. Only config tunes.
• Touching pipeline artifacts. This skill writes two files and no others.

References

• ../_shared/settings-file.md — the canonical template, procedure and key semantics.
• ./references/detection.md — host signals, ranked, and the always-confirm rule.
• ./references/interview.md — the six questions verbatim + the question → keys map.
• ../_shared/ask-style.md — how every question is phrased.
• ../_shared/tool-adapters.md — what each Claude mechanism maps to under Codex CLI / Cursor.
• ../_shared/agent-roster.md — what the model tier actually changes.
• ../implement/references/settings.md — the engine-side delta.