Spek New

Version check first. Before running any other command, run spektacular version check.

• On status: "match", continue with the skill and produce no version-related output.
• On "mismatch" or "missing", the installed Spektacular files are out of date: relay the response's action message to the user, ask them to re-run spektacular init <agent>, and wait for their decision before continuing.
• Never modify or re-install any installed files yourself — refreshing the installation is always an explicit, user-initiated re-run of init.

STOP. Read this before running any command below. A single successful CLI call — including the very first spec new — is NOT task completion. It is not a milestone to report back to the user. It is one step out of many in a workflow that you must keep driving, turn after turn, without stopping, until the CLI itself tells you the workflow is finished. If you find yourself about to say "successfully completed" or summarize results after calling spec new or spec goto even once, you are wrong — go back and read the instruction field you just received, do what it says, and call goto again.

What this skill does

This skill drives a multi-step interactive workflow that produces a complete specification file at the spec_path returned by the CLI. The workflow is owned by the spektacular CLI, not by you — the CLI is the state machine and you are the executor.

On each turn, the CLI returns JSON containing an instruction field. That instruction describes exactly one step (e.g. overview, requirements, acceptance criteria, …). You must:

1. Read the instruction carefully.
2. Perform the step — usually this means interviewing the user and capturing their answers. Some steps tell you to commit the gathered content to the spec file.
3. When the step is complete, run the goto command named at the bottom of the instruction to advance the state machine.
4. Read the next instruction from the new JSON response and repeat.

This is a loop. Do not stop after the first step. Keep looping — step → goto → next instruction → step — until a returned instruction tells you the workflow is finished. Only then should you report completion to the user.

Concretely: do not stop after spec new. That command only starts the workflow — it returns the first instruction (the interview step), not a finished spec. Seeing a clean JSON response with no error is not a signal to stop; it is the signal to keep going. Reporting success, summarizing "spec initialized," or handing control back to the user at this point is the single most common way this skill is executed incorrectly — do not do it.

The interview step

Before any section is drafted, the workflow opens with an interview step: a single open-ended conversation, not a fixed script. You ask adaptive questions about what's being built, who it's for, and what constraints apply, following up on what the user has already said rather than working through a predetermined list. You have the project's full registered-repo roster available during this step — if the project spans more than one repo and the feature reads as focused on one of them, ask whether it also needs changes in another registered repo, shaped by what that other repo actually is (for example, a documentation repo invites asking whether docs need updating). Stop the interview once further questions wouldn't materially change the draft, not once every conceivable detail has been asked about — this should take a small number of exchanges, not an exhaustive back-and-forth. Save your synthesized understanding (not a transcript) to .spektacular/work/<spec_name>/interview.md with your own Write tool before advancing; every later section step drafts from this file and presents its draft back for confirmation, rather than asking its own scripted question from a blank prompt. A session interrupted mid-interview resumes on the interview step itself — read back .spektacular/work/<spec_name>/interview.md (if partially written) and .spektacular/context.md before continuing the conversation.

Reading and writing the spec file

The CLI owns the spec file. Never read or write it with the Write, Edit, or Read tools — those bypass Spektacular and the configured spec directory. All spec file access goes through spektacular spec file:

• spektacular spec file read <name>.md — read a spec file from the spec store.
• spektacular spec file write <name>.md --from <source-path> — write a spec file into the spec store from a source file on disk. Stage the body under .spektacular/tmp/ first, then rm the scratch file after a successful write.
• spektacular spec file list — list spec files in the spec store.

Path arguments are spec file names; spec file resolves them against the configured spec directory itself.

Working files vs. the store document

While you gather each section, write that section's agreed content directly to its own git-tracked working file under .spektacular/work/<spec_name>/<section>.md using your own Write tool. These working files are not store documents — writing them directly with Write is correct and expected, and is the one deliberate exception to the "never use Write/Edit" rule above. That rule protects only the final assembled spec, which is written solely through spektacular spec file write. The per-section working files are scratch-but-durable: the verification step reads them back to assemble the final spec, and then the working directory is removed once the store write succeeds.

.spektacular/context.md has a narrower role: it holds only your cross-cutting learnings and the answers the user gave to your questions — never a copy of section content (that lives in the per-section working files). On resume, read back both the section working files in .spektacular/work/<spec_name>/ and .spektacular/context.md, so you continue from the interrupted step without re-asking for sections already completed.