popjam-io Writing Style Extraction
Reverse-engineer a brand's verbal DNA from the copy it already publishes, and codify it so a copywriter or a generation pipeline reproduces that voice instead of a generic one.
The core insight mirrors visual guideline extraction: a voice extracted from N texts is only as good as (a) how representative the corpus is across channels, and (b) how rigorously you separate rules (what the brand always does) from variations (what it does per channel, campaign type or language) and exceptions (one-offs). A brand that writes 40-character emoji-led captions and 1,500-word how-to articles has one voice and two systems — averaging them produces a voice nobody wrote. The whole workflow is built around that separation.
Work through five phases in order. Phase 3's extraction fan-out is the expensive step; everything else is cheap.
Phase 1 — Corpus inventory
Locate the texts. Usually the user provides a folder or an export (CSV/JSON of posts, a crawl, an ad-library dump); if they name sources (Instagram, blog, Meta Ad Library, website) gather what's accessible first, but never block on missing sources — work with what exists and record the gaps in the coverage section.
Build work/corpus.jsonl, one line per text unit, tagged by channel:
{"id": "ig-2026-04-12", "channel": "social | long_form | ads | web", "text": "verbatim", "language": "tr",
"date": "2026-04-12", "engagement": {"likes": 412, "comments": 9}, "source_hint": "instagram | blog/how-to | meta-ad | pricing page"}
- Channel is the primary axis. Social captions, long-form articles, ad copy and page copy follow different sub-systems even at brands with a strong voice; keep them distinguishable from the start or Phase 4 averages them away.
- Keep text verbatim — no cleanup, no trimming of hashtags or emoji; those are the evidence. Strip only boilerplate that isn't the brand's voice (cookie banners, nav menus, legal footers) and note that you did.
- Dates and engagement are optional but valuable: they let Phase 4 say which patterns are current and which perform.
- Corpus size: 40–120 texts with ≥10 per channel is the sweet spot. Under ~8 in a channel, that channel's block is provisional — say so. Over ~150, sample proportionally per channel and date rather than reading everything.
Phase 2 — Programmatic stats
Run scripts/text_stats.py work/corpus.jsonl -o work/stats.json (also accepts a folder of .txt/.md files in social/, long_form/, ads/, web/ subfolders). Pure standard library. It is the ground truth for lengths (chars, words, sentences, paragraphs), emoji/hashtag/mention/link counts, question and exclamation rates, first- and second-person rates, CTA presence, hook type, caps rate, reading grade, and the language of each text — per sample and aggregated per channel.
Numbers from this script beat impressions: LLM readers systematically overestimate emoji use and underestimate length variance. Anchor every quantitative claim in the guideline to stats.json, and let the reading pass own what the script can't see — why a hook works, what the offer framing is, which words are taboo.
Phase 3 — Per-channel pattern extraction
Fan out subagents by channel (and by language if the corpus is mixed), 15–25 texts per agent, each agent reading every text in full — never let an agent infer style from titles, IDs or the stats file; that defeats the entire purpose. Give each agent the schema from references/style-schema.md and have it write one JSON entry per text to its own output file.
Tell each agent explicitly:
- quote hooks, CTAs, sign-offs and recurring phrases verbatim in the source language — paraphrases destroy the evidence, and the prompt block needs the real words
- record structure as an ordered list of parts (hook → context → proof → offer → CTA → hashtags), not a summary
- record what is conspicuously absent (no emoji, no discounts, never "we", never an exclamation mark) — negative space is a rule too
- flag anything that looks like a deliberate departure (guest author, co-branded post, crisis notice, a different language) so Phase 4 can isolate it instead of averaging it in
Phase 4 — Synthesis
Aggregate the extraction JSON with stats.json and tally every dimension per channel and across the corpus (a small Python script over the JSONL beats eyeballing): hook type, structure template, person, formality, CTA type and wording, offer framing, urgency device, emoji and hashtag policy, sentence style, recurring phrases, language mix. Classify:
- ≥70% of texts → core rule. Codify it.
- 30–70% → variation. Find the trigger: split by channel, campaign type (launch / promo / educational / community), language or date and re-tally. A hook that is 45% overall but 90% inside ads is an ads rule, not noise. Document every trigger explicitly.
- <30% → exception. List as observed-but-not-a-rule; never codify.
Every rule carries an evidence count ("31/38 social captions") or names example IDs — unverifiable rules get challenged by writers and ignored by pipelines. Where engagement exists, note which patterns sit in the top tier, honestly (small samples, format and timing confounds).
Write the guideline following references/voice-template.md (10 sections, then the prompt-block template); read it before writing — it encodes the section-level expectations and the rule/variation/exception labeling. Then distill one voice prompt block per channel: compact, no prose, verbatim hook and CTA examples. This is what a generation pipeline consumes; treat it as a deliverable equal to the document. Rules and attributes are written in English; verbatim examples keep the source language; the block states which language(s) to write in.
Phase 5 — Validation
Hold out one corpus text per channel. From the voice prompt block alone (no other corpus access) rewrite its brief — same topic, same offer — then run original and rewrite through text_stats.py and diff: length, emoji/hashtag counts, person, hook type, CTA presence and reading grade should land inside the channel's observed range. A miss is either a generation slip or, just as often, a missing or ambiguous rule in the block — fix the block, not only the sample. Then read the pair side by side for what stats miss: vocabulary, warmth, taboo words.
Deliverables
Save to the user's workspace folder: the voice guideline (markdown unless the user asks for docx/pdf), the voice prompt block(s), and a data/ subfolder with corpus.jsonl, stats.json and the per-text extraction JSONs. Intermediate scratch goes in work/, not the deliverable folder. End the document with coverage: the channels, languages, campaign types and date range the corpus did NOT contain, so future users know which rules are extrapolations.