SLC CLI

Spec-driven generation for AI builders. SLC reads your requirement.md, uses your own LLM, and generates a complete, validated SLC spec tree — for 100% faithful, decision-accurate specs. It generates specs, never application code.

Terminal UI built on Ink (React for terminals). Requires Node ≥22.

Run it

npx @wewiselabs/slc    # or: pnpm dlx @wewiselabs/slc   ·   bunx @wewiselabs/slc

Want the bare slc command? Install it globally once:

npm install -g @wewiselabs/slc
slc --help

Heads-up: always use the full scoped name with npx. npx slc (unscoped) is a different, unrelated package on npm.

New here? GROUNDWORK.md is the field guide — step-by-step scenarios (bridge mode with Claude Code, local Ollama, resuming after Ctrl+C, CI validation) and a troubleshooting table for every known "it won't run" situation.

What it does

1. Uses your LLM. Pick Anthropic (Claude) or any OpenAI-compatible endpoint (OpenAI, Google, OpenRouter, Ollama, …). Your model, your key, your decisions.
2. Finds your requirement, any name, any format. Common names first, a content-scored scan of the folder if not, slc --file <path> if you'd rather just say where it is — or takes a raw idea typed into the terminal and drafts one. Your file's structure is never judged against SLC's own template, only whether it covers the essentials.
3. Scrutinizes before it builds — reviews, never rewrites. Your LLM audits the idea like a principal engineer — contradictions, feasibility, stack compatibility, security footguns, scale/cost for your ambition level (hobby · growth · business) — then asks about every decision that genuinely matters (no artificial cap). If the brief is already solid, nothing changes — no reflow, no menu. If there are real fixes, you see the exact diff before anything is written. Skipped decisions become explicit assumptions, never silent inventions. Nothing generates until you confirm.
4. Splits scope — a single full-app requirement becomes backend + frontend work.
5. Backend first. Generates the backend spec tree, validates it, and stops at a review gate.
6. Frontend second. Only after you confirm the backend, it generates the frontend specs with their CONTRACT derived from the backend (FE-derives-from-BE).
7. Validates everything with slc doctor.

Companion files count too: if your requirement says "authoritative data model: schema.sql", SLC finds that file and sends it alongside the brief — so the specs derive from the real column names, not from a prose paraphrase of them.

Without the CLI (legacy mode)

The format does not depend on the tool. Drop SLC.md + slc_universal_structure.md in your project root, write requirement.md, and hand all three to any capable LLM using the prompts in SLC_GETTING_STARTED.md. That path is fully supported and unchanged. npx @wewiselabs/slc doctor still validates the result — no key, no config, no terminal needed.

Commands

| Command | Does | |---|---| | slc | The guided generator (the flow above). | | slc feature "<desc>" | Add one feature to an existing spec tree — additive only, never regenerates or touches frozen files. | | slc estimate | Token/cost report for an existing spec tree — real numbers labelled measured vs. derived vs. estimated, no built-in pricing. | | slc db | Data model view from your backend ARCH — entities, fields, inferred relationships. No LLM call. | | slc audit | Security/architecture review of the generated specs — read-only report. Run it, fix what it finds, then run it again: a fix can introduce a finding of its own. | | slc doctor | Validate an existing spec/ tree — structure, references, secrets, and post-merge damage (conflict markers, duplicate ids/blocks, stale counts). Run it after every merge. --json for CI, --fix to recompute derived counters. | | slc bridge | Write AGENTS.md + CLAUDE.md so any agent opening the repo auto-loads SPEC.md and its read order. Needs no terminal. | | slc --no-open | The guided flow, but never launch a browser for the taste demo. | | slc --help / --version | Help / version. |

Driving SLC from an agent

The guided flow renders interactive prompts, so it needs a real terminal — which an agent's shell usually isn't. slc doctor, slc db and slc bridge work anywhere, and the two commands you run repeatedly have a headless file-in / file-out pair:

slc audit --emit                 # writes .slc/run/audit.prompt.md, exits
# …your agent follows that file and saves .slc/run/audit.output.json…
slc audit --consume              # prints the findings, exits non-zero on a blocker

slc feature "<desc>" --emit
slc feature --consume --yes      # --yes is required before anything is written

slc audit --reconfigure re-asks which AI to use; otherwise audit reuses the saved one silently, because it is a command you run repeatedly.

Config

Your provider + key are saved to .slc/config.json (auto-gitignored). ANTHROPIC_API_KEY or OPENAI_API_KEY in the environment are picked up automatically.

Develop

npm install
npm run build
node dist/index.js --help