# Blog Write Post > Draft or revise a post for the Loud Trumpet blog in Ryan's voice, with the repo's file, image and SEO conventions. Use for any post, standalone or part of a series. - Skill: `loud-trumpet/blog-write-post` (Agent Skill) - Install (CLI): `npx skillmds@latest add loud-trumpet/blog-write-post` - Raw SKILL.md: https://api.skillmd.com/api/skills/loud-trumpet/blog-write-post/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Marketing & Growth - Author: loud-trumpet (https://skillmd.com/u/loud-trumpet) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/loud-trumpet/blog-write-post --- # Writing a post ## Interview before drafting The series' credibility comes from lived evidence, and the evidence lives with Ryan rather than in this repo. A scene, a number, a real example, an outcome, a mistake: none of these can be inferred from the plan or the earlier posts, and inventing one destroys the thing the whole blog is trading on. So the first move on any new post is an interview, not a draft. Ask, then write. Do not write a draft with gaps in it and ask afterward, because the answers reshape the argument around them and the draft gets rewritten anyway. **What the interview has to get.** At minimum, every one of these: - **The opening scene.** A specific failure or moment, with a number attached. Ask what actually went wrong, how long it stood, and how it was found. This is the hardest thing to invent and the most important thing to get. - **The concrete examples.** Two or three, with enough detail to be checkable: what the thing was, what it was built from, how many places it was used. An abstract example reads as an abstract post. - **The numbers.** How long, how many, how much faster. A claim with a number is a report; the same claim without one is an opinion. - **The mistake.** What was built wrong, or too early, or too late. A post with one of these in it is believed; a post without one reads as a sales pitch. - **The limits.** What the experience does not establish. Ask directly: what would you not be willing to claim from this? - **The evidence line.** For each section, lived or conviction. Ask rather than guess, and put the answer in the text. - **The audience.** Who is this for, and what are they assumed to already know? Ask for a sentence, not a category. "Technical" and "non-technical" are too coarse to decide anything: the real question is per-term, and a label gets it wrong in both directions. It cuts a term that carries the argument, and it keeps a term that quietly narrows the post. The answer goes in the `audience` frontmatter field and stays there, because it is obvious while you are talking and gone a week later. **When a series plan exists, its "what this post needs before drafting" list is the interview agenda.** Work straight down it. Anything still open in the plan's "open questions" that touches this post gets asked too. **How to ask: one question at a time.** Ask a single question, wait for the answer, then ask the next one. A numbered list of ten questions gets a list of ten short answers, and short answers are exactly what this interview cannot use. One question at a time gets a story, and the story is where the scene, the number and the mistake are hiding. It also lets each question be shaped by the last answer, which is most of the value. The follow-up is usually more productive than the question that prompted it: ask how it was noticed, what it cost, what was tried first, who pushed back. Keep pulling on an answer while it is still producing detail, and move on when it stops. **Expect prose, not a selection.** Almost nothing in this interview is a choice between options, so `AskUserQuestion` is the wrong instrument for it: recall has no options to offer, and offering some invites an answer shaped by the guesses rather than by what happened. Ask in plain text and say that a paragraph, a tangent or "that never came up" are all fine answers. Reserve `AskUserQuestion` for the few places where the answer really is a decision between known alternatives, such as a title, or whether to name a company. **Then confirm before writing.** Play the material back in a few lines: here is the scene, here is the argument it supports, here is what I will say the evidence does not cover. A wrong reading caught in three lines is cheaper than a wrong reading caught in a finished draft. **Never fill a gap with a plausible invention.** If an answer does not arrive, the section does not get written. A placeholder marked as a placeholder is acceptable in an intermediate state; a fabricated example is not, at any stage, including a draft nobody has agreed to publish yet. ## Voice Ryan's voice is plain, concrete, and unhedged. It earns authority by being specific and by admitting what it can't prove, never by sounding confident. **Open on a scene, not a thesis.** The strongest opening is a specific failure with a number attached. One post opens "For about a month, one of our buttons had white text on a white background." No setup, no framing, no "in this post I'll argue". The argument arrives after the reader is already in it. **Short declarative sentences.** Fragments are fine when deliberate: "Not defaulted. Not documented as 'usually this one.' The parameter wasn't there." **One idea per paragraph, and paragraphs stay short.** Two to four sentences. **Precision instead of hedging.** Don't reach for "maybe", "might", "probably", "sometimes". Either the claim is true as stated, or it gets narrowed until it is. "Most of the advice I run into is about the first one" is better than "almost all the industry's energy", not because it's softer but because it's defensible. **Name who did it.** The passive hides who acted, and in a post about working with AI that is the one thing the reader is trying to work out. "This post was written out of an interview" became "an interview helped me write this post" for exactly that reason: Ryan acts on the tool, the tool does not act on the post by itself. This is not a rule that Ryan did everything. Where the model did the thing, say so in the same plain way. "A model drafted this post and I took it apart" is accurate and specific; "AI helped write this post" is the phrase people reach for when they would rather not say what happened. Vagueness about agency reads as something being hidden, which costs more than the admission would. **Concede before the reader objects.** Name the counterargument yourself, in its strongest form, then answer it. A post arguing that shared building blocks keep a model from going wrong lists the four ways it can go wrong anyway, and that concession is what makes the surrounding claim credible. **Close by returning to the opening.** The last paragraph should make the first one mean something it didn't mean at the time. **No aside that costs the claim.** The wry half-sentence at the end of a paragraph is the most expensive thing in this voice, because it sounds like personality and pays for itself with the argument. "The rules as they stood when this post was written, rather than what I have since talked myself into" turns every later correction into a rationalization, in a post whose point is that the corrections are earned. Same with framing the drafting as "something to be dissatisfied with" when the actual claim is that a draft is a faster start. The test: read the clause on its own and ask what it concedes. If the post cannot afford the concession, the joke goes. Self-deprecation is not the same move as admitting a limit; a limit is stated plainly and strengthens what surrounds it, while an aside undercuts it on the way past. **Words to avoid:** "useful" more than twice in a piece, "leverage", "unlock", "delve", "it's worth noting". Anything that sounds like it was written to fill a paragraph. ## Structure that has worked Scene → what it actually cost → what we built → the mechanism, stated as one memorable line → the result with numbers → why it generalizes → what stays the reader's job → honest limits → callback close. When the argument is about specification, a before-and-after comparison is the centerpiece. Both versions must describe the *same* thing; if the short one omits something the long one carries, a reader will find it and conclude the comparison was rigged. Audit them line by line before publishing. ## Evidence discipline This is the series' whole credibility strategy. Hold it. - Label what is **lived** and what is **conviction**, explicitly, in the text. - State limits before a reader can raise them. A post reporting what two weeks of work produced says "two weeks says nothing about what that library costs to maintain" in Ryan's own voice, which is what makes the numbers above it believable. - Ground claims about other teams or the industry in personal observation: "in the codebases I've worked in", not "teams tend to". - Never state a product convention as a fact about the world. "We'd standardized on MM/DD/YYYY", not "there was exactly one correct date format". - Never convert a reduction in risk into a guarantee. See the `blog-review-draft` skill; this is the failure mode that recurs most. ## Style rules, non-negotiable - **No em dashes.** Use a comma, colon, period, or parentheses. An en dash is not a substitute. Em dashes read as AI-written and undermine authorship. - **Straight quotes only**, `"` and `'`, never curly. - **American English.** Color, not colour. -ize, not -ise. - Sentence case for headings. ## Files and conventions The repo README is the source of truth; read it rather than guessing. In short: - Standalone post: `posts/[date]-[name].md` - Series post: `posts/series/[series-name]/[NN]-[name].md` - `images/` and `social/` mirror `posts/`: one directory per post, named after the post's path with the leading `posts/` dropped. | Post | Images | Social | | --- | --- | --- | | `posts/2026-10-01-foo.md` | `images/2026-10-01-foo/` | `social/2026-10-01-foo/` | | `posts/series/bar/02-baz.md` | `images/series/bar/02-baz/` | `social/series/bar/02-baz/` | The hero is always `hero.png`. Reference it as `feature_image: hero.png`; paths resolve relative to the post's own image directory. **Only if the post belongs to a series:** read `_ideas/series/[series-name]/00-series-plan.md` first. It records the rules earlier posts settled, which later posts must not contradict or re-derive, and the evidence status of each planned post. A standalone post has no such file and needs no equivalent. ## Referring to other posts in the series **Never write "post 1", "post 3" or "post 4" in the post itself.** A reader arrives from a search result or a link with no map of the series, so a number identifies nothing. It also reads like filing-system prose rather than writing. For a post that hasn't been published yet, name the topic: "a loop that deserves a post of its own", "a later post turns on what happens when that list isn't empty". The reader learns what is coming rather than that something is. For a post already published, refer to it by phrase and vary the phrase. "The first post in this series" on first mention, then "last time", "that post", or the thing it was about, which is often the most informative option: "the design system I wrote about last time" does more work than any number, and disambiguates where several projects are in play. Where the idea can simply be restated, drop the reference and restate it. If a post refers back more than about three times, the references are doing work the prose should do. Cut some rather than restyling all of them. **Link twice, not every time.** Link the first mention and the last. A reader who wants to navigate back does it before committing to the argument or after seeing it pay off, not from the middle of a paragraph, and one URL repeated eight times reads as filler. ## Frontmatter ```yaml title: ... slug: ... meta_title: ... # the theme appends " - Loud Trumpet", so ~45 chars max meta_description: ... # 140-160 characters feature_image: hero.png audience: ... # who it is for and what they are assumed to know tags: [...] excerpt: ... ``` `audience` never reaches Ghost. The publisher builds its payload from an explicit allowlist, and a pull carries unknown keys forward, so the field lives with the post without appearing on the site. Every technical term in a draft has to survive one of two tests against that line: it carries the argument, or it is glossed in place. A term that does neither is set dressing, and set dressing narrows the audience for nothing. "Markdown" carries an argument about why plain text with structure works, so it stays and earns six words of explanation. "A library you considered and rejected" was one item in a list of three examples, so it became "a tool" and cost the technical reader nothing. `meta_title` and `meta_description` are the search result. Count the site suffix against the title budget, and keep `meta_title` starting with the post's own title: a search result that promises something different from the heading reads as a bait and switch, and Google often rewrites a title tag it judges unrepresentative of the page. `meta_description` also fills the link preview text on LinkedIn and X, so one field normally covers search and social both. Only set `og_description` or `twitter_description` when a card should read differently from a search result: around 100 to 120 characters, written to earn a click rather than summarize. That is the exception, not the habit. `og:title` has no equivalent and follows `meta_title`. ## Before handing a draft over Run the `blog-review-draft` skill against it. At minimum: no em dashes, no curly quotes, no unqualified claims, and the before/after examples describe the same thing. Also check that nothing survives from the interview stage: no placeholder markers, and no example, number or outcome that came from anywhere but an answer Ryan gave. There is no `status` field: publication state lives on each Ghost target, not in the file. Publishing to local always makes a post visible there, publishing to live always creates a draft, and `npm run ghost promote` is the separate command that puts something in front of readers. Never run it unless asked in the conversation.