txt2csv

Extract structured content from noisy social-media pastes and write it as a clean CSV.

Why this skill exists

The user collects useful content (prompt collections, tip lists, framework breakdowns) from social-media feeds. Raw paste is a wall of noise: creator handles repeated between items, timestamps, "Reply/Share" buttons, image alt-text, engagement counters. Doing this cleanup by hand each time is tedious and error-prone — especially the RFC 4180 CSV escaping for prompts that contain commas, double quotes, and special characters.

This skill turns a 5-minute manual cleanup into a 20-second extraction.

When this skill fires

Fire when ALL of these are true:

1. The user has pasted text (or is about to) that looks like a social-media post — not a code snippet, not a config file, not their own writing.
2. The pasted text contains a list of items (usually numbered 1–N) that the user wants to extract.
3. The user wants the result as a CSV or structured table.

Trigger phrases (user may use any of these):

• "filter the noise out of this"
• "clean this up into a table" / "tabulate this"
• "extract these prompts" / "pull out the X from this"
• "make a CSV from this" / "convert this to CSV"
• "do the same with this one" (when a prior txt2csv-style task happened in the conversation)
• the literal slash command /txt2csv

If the user pastes a numbered list from a social post but doesn't say "CSV", still consider firing — most likely intent is the same cleanup task.

Workflow

Follow these steps in order. Do not skip steps.

Step 1: Read the paste and identify structure

• Find the numbered items (1. 2. 3. ... or 1) 2) 3) ... or just "first... second... third...")
• Count them — could be 3, 7, 10, 20+, any number
• Note where each item starts and ends
• Each item typically has a short title line ("1. The Pattern Interrupt Script Builder") and a body wrapped in "..."

If you cannot find clear numbering, ask the user how items are separated before proceeding. Do not guess.

Step 2: Recognize noise and strip it

These patterns are noise. Remove them silently:

Step 3: Preserve one piece of metadata

Identify the source creator's handle (usually the first non-noise line, or the line that repeats most between items). Mentally note it — the user may want it later for attribution. Do not add it to the output unless the user opts in to a Source column.

Step 4: Propose the output schema

Default schema: Number, Title, Prompt

Show a 1-row preview using the first extracted item, then ask:

"Using Number, Title, Prompt as columns — first row would render as: 1, [title], [prompt body] Use this, or want to adjust? Common alternatives: add Role (the 'Act as a X' opener split out), Placeholder (the [paste]/[niche] variable), or Source (the creator handle)."

Wait for confirmation before writing.

Step 5: Propose a filename

Read the post's headline — usually the first content line starting with 🚨, "BREAKING:", "Here are", "X prompts that...", etc. — and derive a kebab-case topic.

Examples:

• 🚨BREAKING: Claude can now replace your entire YouTube growth team → youtube-growth-prompts.csv
• 🚨BREAKING: Claude can now build The complete marketing strategy for Your business → marketing-strategy-prompts.csv
• 7 prompts that turn ChatGPT into a senior copywriter → senior-copywriter-prompts.csv

Show the proposed filename, let the user override.

Fallback when no headline is present. If the user pastes starting from item 1 (skipping the thread header), there is no headline to derive from. In that case, infer the topic from the unifying theme of the prompts themselves — what's the one thing all items help the user accomplish? Examples:

• 6 prompts about removing AI tells, varying sentence length, voice-matching → humanize-ai-text-prompts.csv
• 7 prompts spanning channel strategy, idea generation, scripts, SEO, community, monetization → youtube-channel-blueprint-prompts.csv

State your inferred topic + filename so the user can correct it before the write.

Step 6: Write the CSV

• Location: ./prompt-library/<topic>-prompts.csv, relative to the current working directory. If the user names a different folder, use theirs.
• Create the folder if missing. PowerShell: New-Item -ItemType Directory -Force ./prompt-library. Bash: mkdir -p ./prompt-library.
• Encoding: UTF-8 (default Write tool behavior)
• Escaping: RFC 4180 (see rules below)
• Collision handling: if <topic>-prompts.csv already exists, do NOT blindly append -v2. First check whether the two pastes actually cover the same angle:
  • Compare the existing file's items to the new paste's items. Are they from the same creator and same domain (true duplicate)? Or different creators / different angle on the same broad topic (semantic distinction)?
  • Same creator + same content → ask the user before doing anything; they may have re-pasted by accident.
  • Different creator, but genuinely overlapping items → append -v2, -v3. The number captures "another take on the same thing."
  • Different angle on the same broad topic → coin a new semantic name that captures the difference. Example: existing youtube-growth-prompts.csv (Jamie's identity/positioning angle) + new paste on tactical execution (Shruti's idea-generation/SEO/monetization angle) → name the new one youtube-channel-blueprint-prompts.csv, not youtube-growth-prompts-v2.csv. The folder is easier to browse by topic distinction than by version number.
  • Always state the collision and your chosen name in the report so the user can rename if they disagree.
  • Never overwrite without explicit user confirmation, regardless of which case applies.

Step 7: Report

After writing, output a short report:

• How many items extracted
• What noise patterns were stripped (so user can verify nothing important got cut)
• The final file path as a clickable markdown link
• An ★ Insight ───── callout with 1–2 observations about the content (recurring theme, structural pattern, comparison to prior CSVs in prompt-library/ if relevant — since the user runs in learning/explanatory output style)

Character normalization (CRITICAL — do this before CSV escaping)

Excel on Windows defaults to cp1252 when opening a CSV with no UTF-8 BOM. UTF-8 bytes for smart typography then display as mojibake — I'm renders as I’m, "passionate" as “passionateâ€. The user explicitly flagged this on 2026-05-18.

Solution: normalize all "smart" Unicode typography to ASCII equivalents BEFORE applying RFC 4180 escaping. Apply these substitutions in order:

Why this matters: the user opens these CSVs in Excel on Windows. cp1252 is Excel's default for BOM-less CSVs. Adding a BOM would also work, but normalization is more portable (every tool everywhere reads ASCII, BOM-aware or not).

Order of operations:

1. Strip noise (Step 2 above)
2. Strip outer prompt-wrapping quotes
3. Normalize smart typography → ASCII (this section)
4. Apply RFC 4180 escaping below (note: step 3 may produce new ASCII " chars that now need to be doubled to "")

RFC 4180 CSV escaping (critical)

This is where manual extraction usually breaks. Follow these rules exactly:

Example transformations:

Strip outer prompt-wrapping quotes. Social posts often wrap each prompt body in "..." as a visual cue. Those outer quotes are noise — strip them before writing the cell, then re-apply CSV-level wrapping per the table above. Don't keep both layers, that creates ugly ""prompt"" in the output.

Edge cases

• Truncated source content (sentence cuts off mid-word, post hit a character limit): mark the cell [TRUNCATED IN SOURCE] instead of inventing missing words. Tell the user in the report so they can paste the rest if they have it. Never guess at completion.
• Ambiguous noise: if you cannot tell whether a line is noise or content, ask. Example: "Is Pro tip: between items a section header you want kept, or repeating noise to drop?"
• No clear numbering: ask the user how items separate. Don't guess at bullet points vs paragraphs.
• Mixed languages (English + Spanish, common for this user): pass through as-is. UTF-8 preserves everything.
• Image-only items (item that's just an image alt-text with no real content): drop the entry and tell the user so they can decide if it mattered.
• User says "do the same with this one": same schema (Number/Title/Prompt), same folder, derive filename from the new paste's headline. Don't re-ask the schema question every time.

What this skill does NOT do

These are deliberate design decisions:

• No master/consolidated file. Each invocation writes one topic CSV. The user explicitly declined a prompt-library.csv rollup file when this skill was designed.
• No translation, rewriting, or "improving" of prompts. Pure extraction + cleanup. If a prompt has a typo in the source, the typo stays.
• No external API calls or bundled scripts. Pure Claude reasoning + Write tool. Keeps the skill portable and zero-maintenance.
• No automatic move of existing CSVs. Files already at workspace root (e.g., youtube-growth-prompts.csv) stay where they are unless the user asks to consolidate.

Reference examples

Two shapes cover almost every real paste. Match one of them.

A. No internal quotes — plain cells, no escaping needed. Note the truncation marker, used when the source paste itself cut a prompt off mid-sentence:

Number,Title,Prompt
1,Pick One Content Pillar,"Act as a channel strategist. Given my niche of {{niche}}, name the single content pillar I should own and explain why in 3 bullets."
2,Ninety Day Upload Plan,"Build a 90-day upload calendar for {{niche}}. One video per week. For each, give a title and a one-line hook. [TRUNCATED IN SOURCE]"

B. Internal double quotes — every embedded " doubles to "". This is the case people get wrong, so check your output against it:

Number,Title,Prompt
1,The So What Test,"Rewrite my headline so it passes the ""so what?"" test. Headline: {{headline}}"
2,Objection Reversal,"List the top 5 objections to {{product}}, then rewrite each as a ""yes, and"" benefit statement."

Rules both examples demonstrate: header row always present; a cell is quoted whenever it contains a comma, a double quote, or a newline; Number is bare because it never needs quoting.

Example session (full trace)

User: /txt2csv [pastes a Threads post with 5 productivity tips, "@productivity_pat" handle repeated, "3h Reply Share" lines, an "Image may contain text" alt-line]

You:

1. (Internal) Scan paste → find 5 numbered items, identify noise patterns: handle @productivity_pat, timestamp 3h, buttons Reply/Share, alt-text Image may contain text.
2. (To user) "Found 5 tips. Using Number, Title, Prompt as columns. First row preview: 1, Time-Block Your Mornings, Carve out the first 90 minutes.... Proceed?"
3. After user confirms, propose: "Headline says 'Five productivity tips that changed my year' → filename productivity-tips-prompts.csv. Sound good?"
4. After user confirms, ensure folder exists, write file.
5. Report: "Extracted 5 tips → ./prompt-library/productivity-tips-prompts.csv. Stripped: repeated @productivity_pat handle (×5), 3h timestamps (×5), Reply Share buttons (×5), Image may contain text alt-text (×1). ★ Insight ─────
   • All 5 tips center on time-batching as the core mechanic — that's the post's underlying thesis even though the title says "productivity" generically. ─────"