| name | build-style-guide |
| description | Use to create or update the per-author Ghost style guide from all available signals — interview answers, reference docs, and the published corpus — and to fold a newly approved post into it. Triggers on "build my style guide", "update the style guide", "learn my voice from my posts", and is invoked by push-draft after a successful push. |
Build the Ghost style guide
You are synthesizing a living, per-author style guide from all available signals: interview answers, reference material, and the published corpus. The guide lives at the path set in style_guide_path (default .claude/ghost-style-guide.md).
Work through each phase in order. Do not skip a phase — each feeds the next.
Phase 1: Load config and inputs
1a. Read project config
Read .claude/ghost.local.md and extract the frontmatter fields:
Read(".claude/ghost.local.md")
The fields you need:
corpus_filter — NQL filter for corpus pull (default status:published).corpus_limit — max posts to pull (default 25).style_guide_path — output path for the style guide (default .claude/ghost-style-guide.md).
If ghost.local.md does not exist, use the defaults above. Note it in your working notes.
1b. Read the voice-inputs handoff
Read the interview answers and reference notes written by /ghost:define-voice:
Read(".claude/ghost-voice-inputs.md")
This file is the authoritative source for Stated voice. Parse it into:
- Interview answers (Q1–Q6)
- Reference notes (one block per source, with tone/structure/quotes)
If the file does not exist, continue — this is the cold-start-without-interview path. The Stated voice section will be sparse; note that explicitly in the guide.
1c. Check whether a style guide already exists
Read("<style_guide_path>")
If it exists: you are in merge mode — preserve the existing Stated voice section verbatim. Continue to Phase 2.
If it does not exist: you are in initial build mode — you will create it from the template. Continue to Phase 2.
Phase 2: Pull the corpus
2a. Verify the Ghost connection
ghost_site_info()
If this fails (auth error, connection refused): skip corpus pull, note "corpus unavailable — Ghost not connected", and proceed with interview-only inputs. The guide will still be valid; the Observed patterns section will be empty.
2b. List published posts
ghost_post_list(filter="<corpus_filter>", limit=<corpus_limit>, fields="id,title,published_at,url")
Capture the list. If no posts are returned, note "no published posts found" and proceed.
2c. Fetch post content
For each post in the list (up to corpus_limit), fetch the full content:
ghost_post_get(id="<post_id>")
Analyze the returned html (this tool returns html + lexical; there is no plaintext field).
Work through all posts before moving to Phase 3. This is the most time-consuming step — process them in order, extracting style signals as you go.
Cold start without corpus: If ghost_site_info failed or returned no posts, skip Phase 2b–2c and proceed directly to Phase 3.
Phase 3: Synthesize the four sections
Build or update each section. Keep the rules for each section strictly separate — they serve different purposes and different update policies.
Section 1 — Stated voice (authoritative; never auto-overwritten)
In initial build mode: Derive this section from the interview answers in .claude/ghost-voice-inputs.md. Map each field directly:
- Audience → Q1 answer (verbatim, lightly paraphrased if the author was verbose)
- Reader outcome → Q2 answer
- Tone → Q3 answer — preserve the author's exact adjectives or phrase
- Admired references → Q4 answer (names + what the author said they admire)
- Avoid → Q5 answer — record every specific item the author mentioned
- Structural preferences → Q6 answer
If no interview exists, write placeholder values and note that /ghost:define-voice should be run first.
In merge mode: Read the existing Stated voice section from the style guide and carry it forward unchanged. Do not re-derive it from the interview. The author's edits are intentional and must be preserved. If there is new interview material (a re-run of define-voice), surface the delta for explicit confirmation before updating.
Section 2 — Observed patterns (refreshable; re-derived each run)
Analyze all fetched posts together. Derive:
- Typical post length — word count range (min, median, max across the corpus)
- Sentence rhythm — average sentence length, variation, use of short punchy sentences vs. long flowing ones
- Paragraph rhythm — typical paragraph length in sentences; how often one-sentence paragraphs appear
- Opening moves — how posts typically begin (anecdote, question, assertion, code example, context-setting)
- Closing moves — how posts typically end (summary, CTA, open question, sign-off)
- Structural moves — recurring transitions, section patterns, use of subheadings
- Vocabulary register — technical density (jargon, acronyms), formality level
- Formatting habits — frequency and style of code blocks, bullet lists, numbered lists, callouts, images
- PoV and tense — first/second/third person, past/present
Write these as concise observations, not prescriptions. Each observation should cite at least one example post by title.
If no corpus was available, write: "No corpus folded in. Run /ghost:build-style-guide after publishing posts to populate this section."
Section 3 — Anti-patterns (refreshable; re-derived each run)
Combine two sources:
- From the interview — Q5 (what to avoid). Translate each stated avoidance into a concrete pattern description. Example: "no jargon" → "Avoid insider acronyms without expansion on first use."
- From corpus analysis — flag any patterns from the Observed section that conflict with the Stated voice. Also flag any of these house anti-patterns if you observe them:
- Forbidden heading styles: ALL-CAPS headings, question-mark-only headings
- Editorial restatements: "In other words…", "To put it simply…", "What I mean is…"
- Hedge words: "just", "simply", "basically", "obviously", "of course"
- AI-flavored transitions: "In conclusion", "It's worth noting that", "It's important to remember", "Delve into", "Certainly"
- Cleft / focus-frames: "What this means is…", "The thing about X is…", "What's interesting here is…"
- Cute closers: "Happy coding!", "Until next time!", "Stay curious!", "Hope this helps!"
- Code-fence width: code blocks without a language tag, or lines exceeding ~70 columns (Ghost renders narrow)
Also include the reading-pass tells — the ones a grep cannot catch. These are rhythm and framing rather than fixed strings, so the guide must flag them for a read (with the read-aloud test: "if you wouldn't say it, it's a tell"). Carry each with a concrete before/after so the drafting and revising skills can match against real examples:
- Choppy rhythm — plain is not short; a run of short declaratives reads as mechanical. Connect them into substantive sentences that carry cause and effect; vary length.
- Cleft / focus-frame — "X is what Y", "the X is that"; lead with the subject.
- Manufactured contrast — "X is easy; Y is the hard part"; describe what actually happened.
- Em-dash reveal — "[claim] — and that turned out to be [payoff]"; end on the fact.
- Retcon — a fix framed as intentional design; describe what the thing does and why, not a story of having meant it all along.
- Overselling — "I set one rule at the start", "the part I lean on"; understate.
- Negative-first — lead with the reason, not the absence.
List each anti-pattern as a short bullet with a concrete example of what to avoid.
If no corpus and no interview: write the house defaults list above as baseline anti-patterns.
Section 4 — Provenance log (idempotent append-only)
In initial build mode: Create the table with one row per source:
| Date | Source | Type | Notes |
|---|
<today> | .claude/ghost-voice-inputs.md | interview | Initial build |
<today> | <post title> | corpus-post | Folded in on initial build |
Add one row per post fetched. If no corpus was available, note that.
In merge mode: Read the existing provenance log. Append new rows for any posts not already listed. Never delete existing rows. The log is the audit trail of what shaped the guide.
Phase 4: Write the style guide
Write the complete four-section document to <style_guide_path>.
Use the following skeleton as the file structure. Update the frontmatter fields:
last_updated: today's date (ISO 8601)sources_count: total number of posts folded in across all runs (from the provenance log)
Write the file with this exact structure:
---
# Ghost style guide — generated by /ghost:build-style-guide
# DO NOT hand-edit the Observed patterns or Provenance log sections — they are re-derived on each merge.
# Edit only the Stated voice section to add or refine authoritative voice guidance.
generated_by: ghost:build-style-guide
last_updated: "<YYYY-MM-DD>"
sources_count: <n>
---
# Ghost Style Guide
_Generated by `/ghost:build-style-guide`. The Stated voice section is authoritative and human-controlled; all other sections are derived and refreshable._
---
## Stated voice
<!-- SOURCE: Interview answers from .claude/ghost-voice-inputs.md — NEVER auto-overwritten. -->
<!-- Edit this section directly to add or revise authoritative voice rules. -->
**Audience:** <derived from interview Q1 or placeholder>
**Reader outcome:** <derived from interview Q2 or placeholder>
**Tone:** <derived from interview Q3 or placeholder>
**Admired references:** <derived from interview Q4 or placeholder>
**Avoid:** <derived from interview Q5 or placeholder>
**Structural preferences:** <derived from interview Q6 or placeholder>
---
## Observed patterns
<!-- SOURCE: Derived from the published corpus via ghost_post_list / ghost_post_get. -->
<!-- Re-derived on each /ghost:build-style-guide merge run — safe to overwrite this block. -->
<derived observations, or placeholder if no corpus>
---
## Anti-patterns
<!-- SOURCE: Synthesized from interview (avoid list) + cross-checked against corpus + house defaults. -->
<!-- Re-derived on each merge run — safe to overwrite this block. -->
<derived anti-patterns, or house defaults if no corpus/interview>
---
## Provenance log
<!-- SOURCE: Idempotent append-only log of every post and reference folded into this guide. -->
<!-- Append new entries; never delete existing ones. -->
| Date | Source | Type | Notes |
|------|--------|------|-------|
<one row per source>
After writing, confirm the file was saved and report:
- Path where the guide was written
- How many posts were folded in
- Whether this was an initial build or a merge
- Whether the Stated voice section is interview-derived, carried forward from a prior run, or placeholder-only
Phase 5: Continuous merge (post-push path)
This phase applies when build-style-guide is invoked by /ghost:push-draft after a successful push of a newly approved post, or when the author manually runs it after publishing a new post.
5a. Identify the new post
The caller (push-draft or the author) provides the post ID or title. Fetch it:
ghost_post_get(id="<new_post_id>")
5b. Check provenance log
Read the existing style guide's provenance log. If this post is already listed, this is a no-op — confirm "already folded in" and stop.
5c. Re-derive Observed patterns
Fetch the current corpus (same filter/limit as Phase 2) plus the new post. Re-derive the Observed patterns section with the expanded dataset.
5d. Surface guidance changes for confirm
Before writing, diff the new Observed patterns against the existing ones. If any observation changed meaningfully (e.g., average post length shifted, a new structural pattern emerged), surface those changes explicitly:
"The Observed patterns section would change in these ways:
- Average post length: 850 words → 920 words
- Opening moves: adds 'problem statement' pattern (3 of 10 posts now use it)
The Stated voice section will not be changed. Approve these updates? [y/n]"
Wait for explicit confirmation before writing. If the author says no, do not update the file.
5e. Never override Stated voice
Do not touch the Stated voice section under any circumstances during continuous merge. The author's intentional edits are permanent until they explicitly re-run /ghost:define-voice and then confirm an update.
5f. Append to provenance log
Add one row for the new post. Write the updated file.
Notes
- Cold start (no corpus, no interview): The guide is created with house-default anti-patterns and placeholder Stated voice. It is valid and usable; it just needs more signal. Run
/ghost:define-voice first, then re-run this skill.
- Cold start (interview only, no corpus): Stated voice is fully populated from the interview. Observed patterns and anti-patterns are partially populated (interview-derived avoidances + house defaults). Add corpus once posts are published.
- Cold start (corpus only, no interview): Observed patterns and anti-patterns are fully derived. Stated voice will be placeholder. Run
/ghost:define-voice to fill it in, then re-run this skill.
- The Stated voice section is the author's voice. It must never be overwritten silently. Only the author can change it — either by editing the file directly or by re-running
/ghost:define-voice and explicitly confirming an update.
- The style guide is the source of truth for all writing skills in the ghost plugin. If the guide and the handoff file disagree, the guide wins (it may contain post-interview edits).
- MCP tools used:
ghost_site_info, ghost_post_list, ghost_post_get. Config keys read: corpus_filter, corpus_limit, style_guide_path from .claude/ghost.local.md.
