# Write > Master writing skill. Orchestrates the full pipeline (seed → draft → edit → slop-check → optional repurpose) for any post type — long-form (long-form-weekly-newsletter, long-form-monthly-devlog, long-form-hot-take, long-form-how-to, long-form-launch-retro, long-form-manifesto) or short-form (short-form-x-vs-y, short-form-listicle, short-form-observation, short-form-motivational, short-form-present-vs-future, short-form-thread-from-longform, short-form-carousel-li). Drafts live inside the active week's calendar directory. Invoke with `/write ""` or just `/write` to be routed via interview. Interactive by default, autonomous with --auto. - Skill: `r-bart/write` (Agent Skill) - Install (CLI): `npx skillmds@latest add r-bart/write` - Raw SKILL.md: https://api.skillmd.com/api/skills/r-bart/write/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: DevOps & Infra - Author: r-bart (https://skillmd.com/u/r-bart) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/r-bart/write --- # /write Single entry point for the writing system. Loads the right templates, runs the pipeline, saves drafts inside the active week's calendar directory, and (for long-form) derives short-form repurposes. ## Invocation ``` /write → router: 2-3 question interview, picks type /write → directly to that type's draft flow /write "" → with topic seed, skips topic interview /write --from → seed from a notes/idea file /write --repurpose → skip everything, only run thread-from-longform /write --edit → skip drafting, only run editing-pass + ai-slop-check /write --list → show the 13 types with one-line guidance ``` ## Flags | Flag | Effect | |---|---| | (none) | Default: interactive. Asks before each phase transition. | | `--auto` | Run all phases end-to-end without confirmation prompts. Output the final draft. | | `--step` | Alias for default (interactive). Explicit version. | | `--draft-only` | Stop after Phase 2 (Draft). Skip edit + slop check + repurpose. | | `--no-repurpose` | Run all phases except Phase 5 (Repurpose). | | `--from ` | Use the file's content as input/notes for the draft. | | `--repurpose ` | Jump to Phase 5 only. `` must be an existing long-form draft. | | `--edit ` | Jump to Phases 3+4 only. `` must be an existing draft. | | `--ignore-agenda` | Skip the agenda proposal in Phase 0; go directly to the type interview. | | `--list` | Print the 13 types with one-line "use when" guidance. Exit. | ## Types (positional) **Long-form:** - `long-form-weekly-newsletter` — Sunday long-form weekly recap/newsletter (the recurring Sunday default) - `long-form-monthly-devlog` — reflective monthly recap, ~1200-2000 words - `long-form-hot-take` — contrarian opinion, ~800-1500 words - `long-form-how-to` — actionable guide, ~1500-3500 words - `long-form-launch-retro` — post-launch breakdown with numbers, ~800-1500 words - `long-form-manifesto` — vision/strategy piece, ~1500-3000 words **Short-form:** - `short-form-x-vs-y` — comparison post - `short-form-listicle` — numbered list post - `short-form-observation` — single noticed thing - `short-form-motivational` — motivational post (use sparingly, max 1-2/mo) - `short-form-present-vs-future` — temporal contrast - `short-form-thread-from-longform` — derives short-form pieces from existing long-form (alias for `--repurpose`-style flow when you want to start from a long-form file) - `short-form-carousel-li` — LinkedIn carousel content ## Pipeline ### Phase 0 — Routing If no type provided AND no `--repurpose`/`--edit` flag: First, unless `--ignore-agenda` was passed, consult today's agenda via the active week's plan: 1. Compute `current_week` = ISO week of today using `date +%G-W%V` (note: `%G`, not `%Y` — required around year boundaries). Format is `YYYY-WNN` (e.g. `2026-W18`). 2. Look for `scribetronic/calendar//plan.md`. - If file missing: resolver returns null (no week scaffolded). The fallback interview (below) should suggest `/agenda plan-week today` before continuing. 3. Find rows in plan.md where `Date` == today AND `Status` == `queued`. - 0 matches → resolver returns null. Fall through to rules-only resolution per `/agenda`'s algorithm, then to the interview if still null. - 1 match → propose to the user: "Today's slot is `{type}` — '{seed}' (slug: `{slug}`). Proceed? (y / pick another / cancel)". - `y` → skip the interview; go to Phase 1 with that `type` + `seed` + `slug`. - `pick another` → fall through to the interview. - `cancel` → exit. - >1 matches → surface the conflict, list all rows, ask which one (or `pick another` / `cancel`). 4. The resolver is **resilient** — if `plan.md` is missing or malformed, it returns null. It will not throw. Interview (fallback when no agenda slot, agenda declined, or `--ignore-agenda`): Ask 2-3 questions, max: 1. "Long-form or short-form?" 2. (Long) "Is this about something you DID (devlog/retro/how-to) or something you THINK (hot-take/manifesto)?" (Short) "Comparing two things, listing items, sharing an observation, or repurposing a long-form piece?" 3. (If still ambiguous) one final disambiguating question. Then proceed to Phase 1 with the chosen type. If type IS provided: skip Phase 0. ### Phase 1 — Seed Load context (mandatory order): 1. **Voice base** (always). Resolve in this order: a. `scribetronic/style/writing-style.md` in the project root, if it exists — user's personalized override. b. Otherwise, the bundled `writing-style/SKILL.md` template. 2. The type's template file (e.g., `long-form-hot-take/SKILL.md`, `long-form-weekly-newsletter/SKILL.md`). 3. If short-form: also `short-form-voice-adjustments/SKILL.md`. Then: - If `--from ` was passed: read it as raw notes/seed. - If a `""` was passed positionally: treat it as the topic. - If neither: ask 2-4 type-specific questions per the template's "Workflow" section. - Example for `long-form-hot-take`: "What's the conventional wisdom you're disagreeing with? What's your alternative? Concrete failure modes?" - Example for `long-form-weekly-newsletter`: "What did you ship this week? Any conversations/people moments? Insights from this specific week?" - Example for `long-form-launch-retro`: "What launched? When? What were the numbers? Pre-launch expectation?" Stop after seed is captured. Confirm with user: "Ready to draft? (y / change seed / cancel)" — unless `--auto`. ### Phase 2 — Draft Generate draft following the template's structure exactly. Apply the voice rules loaded in Phase 1 (project override or bundled template). **Save path** — drafts live inside the active week's calendar directory: 1. Compute `current_week` = `date +%G-W%V` (format `YYYY-WNN`). 2. If `scribetronic/calendar//` does NOT exist, refuse with the message: > "No week scaffolded for ``. Run `/agenda plan-week today` first." Do NOT auto-scaffold the week — that requires a newsletter seed and is owned by `/agenda plan-week`. 3. Determine the path: - For `type == long-form-weekly-newsletter`: ``` scribetronic/calendar//newsletter.md ``` - For all other types: ``` scribetronic/calendar//derivatives/--.md ``` where `` is today's weekday as a lowercase 3-letter abbreviation (`mon`/`tue`/`wed`/`thu`/`fri`/`sat`/`sun`). - Lazily create `derivatives/` if it doesn't exist yet. - **NO** `-x`/`-li`/`-threads`/`-carousel` suffix in the filename. ONE file per derivative — platform lives in the file's frontmatter. **Frontmatter** — two shapes per Contract 6. For `long-form-weekly-newsletter` at `/newsletter.md`: ```yaml --- type: long-form-weekly-newsletter status: draft created: YYYY-MM-DD week: YYYY-WNN seed: "" length_words: --- ``` For derivatives at `/derivatives/--.md`: ```yaml --- type: status: draft created: YYYY-MM-DD week: YYYY-WNN day: YYYY-MM-DD # date this derivative is targeting platform: x | linkedin | threads | carousel parent: newsletter # always seed: "" length_chars: --- ``` > `published_date: YYYY-MM-DD` is appended later by `/write-publish` on successful publish. The `published` status is owned by `/write-publish` — `/write` never writes it. Show the full draft to the user. If interactive: ask "Continue to editing? (y / iterate draft / save & stop)". If `--auto`: continue. If `--draft-only`: stop here. Output draft path. ### Phase 3 — Edit Run `editing-pass/SKILL.md` against the draft. Apply changes inline. Output: same file, edited. The pre-edit version is preserved as `{filename}.draft.md` (suffix `.draft` before `.md`) **in the same directory** so the user can recover the unedited version. Update frontmatter `status` to `edited`. Show the edits as a diff or as the new full version (user choice if interactive). If interactive: "Continue to slop check? (y / iterate edit / save & stop)". If `--auto`: continue. ### Phase 4 — Slop check Run `ai-slop-check/SKILL.md` against the edited draft. Output: a report with issues by severity (HIGH / MEDIUM / LOW). For HIGH issues: propose fixes inline. If interactive, ask before applying. If `--auto`, apply automatically and log. For MEDIUM: surface, ask for decision (interactive) or apply if confidence is high (auto). For LOW: just surface. After fixes, re-run if any HIGH was fixed. Re-run cap: 2 passes total. If still HIGH after 2 passes, flag for manual review. Update frontmatter `status` to `ready` (passes Phase 4 cleanly) or leave as `edited` (HIGH still unresolved after cap). If interactive: "Mark as ready to publish? (y / iterate / save & stop)". If `--auto`: continue. ### Phase 5 — Repurpose (long-form only) Skip if: - Type is short-form. - `--no-repurpose` flag. - User declines in interactive mode. Otherwise, derive short-form pieces from the just-finished long-form parent (the newsletter): 1. Read `scribetronic/calendar//plan.md`. 2. Find rows whose `Source` is `newsletter` (or `newsletter (thread)`) AND whose `Status` is `queued`. These are the planned derivatives for this week. 3. For each such row, generate one derivative draft using `short-form-thread-from-longform/SKILL.md`. Save to: ``` scribetronic/calendar//derivatives/--.md ``` where: - `` is the row's `Day` (lowercase 3-letter abbreviation), - `` is the row's `Type`, - `` is the row's `Slug`. 4. Frontmatter follows Contract 6's derivative shape. The `platform` field is taken from the plan.md row's `Platform` column. `parent: newsletter`. 5. If a derivative type proposed by Phase 5 has NO matching plan.md row: - Interactive: ask the user for the platform (and confirm slug). - `--auto`: skip that derivative with a warning logged in the final output. 6. Each derivative goes through Phase 4 (slop check) before save. Same severity handling, same re-run cap. There is ONE file per derivative. The channel (X, LinkedIn, Threads, carousel) is recorded in the file's `platform` frontmatter — never in the filename suffix. ### Phase 6 — Final output Print: ``` ✓ Long-form: scribetronic/calendar//newsletter.md ✓ Edited: yes ✓ Slop check: clean (or: 1 MEDIUM unresolved) Derivatives: - thu-thread-niches-are-dead (platform: x) - fri-carousel-niches-are-dead (platform: carousel) Ready to publish: [ ] Final read-through aloud [ ] Cross-platform cadence check (don't ship X+LI+Threads same day) [ ] Schedule or post ``` Update frontmatter `status` to `ready` on the main draft (and on each derivative that passed Phase 4 cleanly). Then update calendar bookkeeping (both files use markdown tables — see `agenda/SKILL.md` for the full schema): 1. **history.md** — append a row to the markdown table at `scribetronic/calendar/history.md`: ``` | YYYY-MM-DD | {type} | {slug} | ready | /write | ``` Use today's date. Schema: `| Date | Type | Slug | Status | Source |`. If the file or table is missing (first run), create it with the header documented in `agenda/SKILL.md` § File contracts, then append. Append one row per finalized draft (the newsletter and each derivative). 2. **plan.md** — for each finalized draft, find the row in `scribetronic/calendar//plan.md` whose `Slug` matches the draft's slug, and set its `Status` from `queued` to `drafted`. Plan column schema: ``` | Date | Day | Type | Slug | Platform | Source | Status | ``` The only transition `/write` is allowed to make is `queued → drafted`. ONE write per matched row. If no row matches: surface a warning (don't error — the user may have written off-plan). **Status-write ownership (do not violate):** - `/write` writes `drafted` (to plan.md) and `ready` (to history.md). Nothing else. - `/write-publish` owns `published` rows in history.md and the `published` transition on plan.md. - `/agenda skip` owns the `skipped` transition on plan.md. This avoids race conditions when multiple flows touch the same files. ## Interactive prompts — phrasing rules - Always offer 3 choices: continue (`y`), iterate (`change seed` / `iterate draft` / etc.), abort (`save & stop`). - Never ask open-ended "what would you like to do?" — too vague. - After a phase produces output, show the output IN the chat (don't just say "draft saved"). User must see what was written. - If user says "iterate", ask ONE specific question: "What should change?" — don't run a 5-question questionnaire mid-flow. ## Slug rules `{slug}` derivation, in priority order: 1. Slug from the matched plan.md row (if Phase 0 routed via agenda). 2. Explicit slug in seed (if user wrote one). 3. First 4-6 meaningful words from seed → kebab-case. 4. Auto from headline → kebab-case, max 60 chars. ## Examples ``` /write → Phase 0: reads scribetronic/calendar/2026-W18/plan.md. Finds today's queued slot. → "Today's slot is `long-form-weekly-newsletter` — 'shipping fast is overrated...'. Proceed? (y / pick another / cancel)" → y → loads long-form-weekly-newsletter template + writing-style. → Drafts to scribetronic/calendar/2026-W18/newsletter.md. Edits. Slop-checks. → Phase 5 reads plan.md, derives the queued thread + carousel rows. Done. /write long-form-hot-take "shipping fast is overrated when you have no audience" → Loads long-form-hot-take template + writing-style. → Asks: "Whose advice are you disagreeing with? What's your alternative?" → Drafts to scribetronic/calendar/2026-W18/derivatives/mon-long-form-hot-take-shipping-fast-overrated.md. → Edits. Slop-checks. Repurpose? → derives matching plan.md rows. /write --from scribetronic/ideas/idea-launch-retro.md → Reads file. Auto-detects type if possible (else asks). → Drafts using file as seed, into the active week's directory. /write --repurpose scribetronic/calendar/2026-W18/newsletter.md → Skips Phases 0-4. Reads same week's plan.md, generates queued derivatives. /write --edit scribetronic/calendar/2026-W18/derivatives/wed-short-form-thread-from-longform-niches-are-dead.md → Runs editing-pass + ai-slop-check on existing draft only. /write --auto long-form-launch-retro "Product #2 — Boilerplate kit" → Asks zero questions beyond seed details. Runs end-to-end. Outputs final. /write --list → Prints 13 types with one-line "use when" each. ``` ## Anti-patterns - Don't run any phase without loading the voice base first (Phase 1 step 1: project-local override `scribetronic/style/writing-style.md` if present, else bundled `writing-style/SKILL.md`). The voice base is non-negotiable. - Don't skip the seed interview for ambiguous types — the 2-4 quick questions save 10 minutes of bad draft. - Don't auto-fix MEDIUM slop issues without asking unless `--auto`. - Don't generate a draft + 4 derivatives in one phase. Repurpose is its own phase, optional, separate output. - Don't write outside `scribetronic/calendar//`. The week directory is the unit of organization. - Don't append `-x`/`-li`/`-threads`/`-carousel` to filenames. ONE file per derivative; platform is in frontmatter. - Don't auto-scaffold a missing week directory. Refuse and instruct the user to run `/agenda plan-week today`. ## Failure modes - **No week scaffolded:** refuse with a message pointing to `/agenda plan-week today`. Do not create the directory. - **User asks for a type that doesn't exist:** suggest the closest match from the 13. Don't invent. - **No seed and `--auto`:** abort with error. `--auto` requires a seed (positional or `--from`). - **Type is long-form but user wants only a thread:** suggest `/write short-form-thread-from-longform` or `/write --repurpose ` instead. - **Edit conflicts (user already modified the draft between phases):** detect via mtime; ask before overwriting. - **Slug not in plan.md at Phase 6:** warn ("off-plan draft — no plan.md row updated") but still complete the run and append to history.md.