Articulate

articulate is the conductor for writing: it matches ceremony to the piece, picks the platform up front, and drives the pipeline to a finished, platform-ready draft — never reimplementing a phase inline when a mode owns it. Its sibling materialize owns code; articulate owns prose.

Pick the platform first

A blog post and a LinkedIn post differ from the first fragment, not just at the end — length, hook, what's worth mining, voice. So choose the target platform before mining, and let it tune the whole run. Load its profile from reference/platforms/ (linkedin.md, blog.md, twitter.md, …) and keep it in context across every phase. If the user hasn't said, ask which platform before starting.

Adding a platform is a new profile file, not a new mode. Repurposing a finished piece for a different platform (a blog → a LinkedIn teaser) is a fresh, platform-committed run seeded from the first.

Setup

1. If the user invoked a mode (init, fragments, shape, beats), read reference/<mode>/<mode>.md next.
2. Voice profile. Load the author's voice, differentiators, hard rules, and cadence, and keep it in context across every phase; it's what makes the writing theirs. The profile is user-global, so one voice serves every repo: load repo-local docs/writing/author.md if it exists, else global ~/.claude/writing/author.md. If neither exists, offer to run init before the first real piece (a writing session works without it, just more generic).
3. Load the target platform's profile (above), so voice (author) and format (platform) are both in context.
4. Read a sample of the author's existing writing on that platform when available, to match their voice before producing anything.

The pipeline

Run it once; scale the ceremony to the piece. Every phase is platform-aware — it reads the profile.

1. fragments — mine raw material (from the user, and from the repo/git/PRs for a coding project). What's worth mining depends on the platform: a tweet needs one sharp nugget, a blog needs many.
2. shape (argument) or beats (narrative) — build the finished piece for the platform, to its length budget and formatting. Pick by whether the piece makes a case or tells a story; beats also fits a tweet thread (one beat per tweet). When it does both, pick by which failure hurts more — a story losing its thread → beats; an argument losing its logic → shape.

Scale down for a quick piece (a tweet may go fragments → a short beats); scale up for a researched blog. Entering with notes already in hand → start at shape/beats; the platform is still chosen first.

Grilling — the technique under every piece

fragments is a grilling session: interview the user relentlessly, one question at a time, capturing nuggets as they fall out — don't impose structure while mining. The shaping modes invert it: stop asking "what are you noticing?" and start asking "what is this piece arguing, and in what order does the reader need it?" Push back; refuse weak transitions.

Durability

Two files, paths the user picks once and you remember:

• The pile — a markdown file fragments appends to. Raw material; read-only to the shaping modes.
• The piece — where shape/beats write the platform-ready draft.

Re-read from disk before every write — the user edits between turns. Never overwrite blindly; append, or edit a named block in place. No scratch-dir or marker machinery — a writing session is one sitting.

Modes

Platform profiles (data, not modes — loaded up front and improved over time): reference/platforms/ — one file per target with its length budget, hook, formatting, and failure modes.