pm-claude-skills 61.1.0 → 61.2.2

package/README.md

@@ -1,4 +1,4 @@
-# 🧠 PM Skills — 750 Professional Agent Skills for Claude, ChatGPT, Gemini, Cursor, Codex & Hermes
+# 🧠 PM Skills — 751 Professional Agent Skills for Claude, ChatGPT, Gemini, Cursor, Codex & Hermes
 
 [![In the official Anthropic plugin directory](https://img.shields.io/badge/Anthropic%20Plugin%20Directory-Published-D97757?logo=anthropic&logoColor=white)](#-quick-start)
 [![Stars](https://img.shields.io/github/stars/mohitagw15856/pm-claude-skills?style=social)](https://github.com/mohitagw15856/pm-claude-skills/stargazers)
@@ -12,7 +12,7 @@
 [![License](https://img.shields.io/badge/license-MIT-lightgrey)](LICENSE)
 [![Sponsor](https://img.shields.io/badge/sponsor-❤️-ff69b4)](https://github.com/sponsors/mohitagw15856)
 
-**A library of 750 skills — each one a plain `SKILL.md` file that teaches your AI assistant to do one professional task properly.** Decode a lease before you sign it. Write a PRD your team can execute. Simulate the promotion committee before the real one meets. Check the weather with zero API keys. Generic AI gives you filler; these give you the structure a senior professional actually uses.
+**A library of 751 skills — each one a plain `SKILL.md` file that teaches your AI assistant to do one professional task properly.** Decode a lease before you sign it. Write a PRD your team can execute. Simulate the promotion committee before the real one meets. Check the weather with zero API keys. Generic AI gives you filler; these give you the structure a senior professional actually uses.
 
 Works natively in **Claude Code** and **Hermes Agent**, with ready-to-paste exports for **ChatGPT, Gemini, Cursor, Codex** and 8 more tools. *(PM stands for Professional, not just Product Management.)*
 
@@ -30,7 +30,7 @@ No `npm install` needed — `npx pm-claude-skills …` always runs the latest. `
 
 ## 📚 The skills
 
-Every skill follows the same discipline: what it produces, the inputs it needs, a real framework (severity scales, decision rules — not vibes), a concrete output template, quality checks, and anti-patterns. All 750 pass the [SkillSpec](SKILLSPEC.md) L3 gate and a security audit in CI.
+Every skill follows the same discipline: what it produces, the inputs it needs, a real framework (severity scales, decision rules — not vibes), a concrete output template, quality checks, and anti-patterns. All 751 pass the [SkillSpec](SKILLSPEC.md) L3 gate and a security audit in CI.
 
 ### For everyone — life's paperwork and decisions
 
@@ -155,7 +155,7 @@ The whole library on one poster — start path, standout features, and install o
 
 ## 🆕 Latest
 
-**v61.1.0 — the demand side:** **[find](https://mohitagw15856.github.io/pm-claude-skills/find.html)** — a natural-language task→skill router (`pm-claude-skills find "prep a QBR"`) with an honest per-skill **measured / no-claim** badge; **per-skill versioning** (`version:` frontmatter + a git-derived `changelog` command); and **[SkillSpec as an adoptable standard](conformance/REGISTRY.md)** — an earnable `--badge` any repo generates from its own CI, plus a public conformance registry. *Earlier — v61, the trust layer:* **[prove](CHANGELOG.md)** (A/B-verify any skill), **mcp-audit**, **[pm-seatbelt](plugins/pm-seatbelt)** (6 agent-safety skills), and the [Cowork](https://mohitagw15856.github.io/pm-claude-skills/cowork.html) + [AI Spend](https://mohitagw15856.github.io/pm-claude-skills/spend.html) pages. Full history: **[CHANGELOG](CHANGELOG.md)** · [releases](https://github.com/mohitagw15856/pm-claude-skills/releases)
+**v61.2.2 — long-form video:** **[youtube-script](skills/youtube-script/SKILL.md)** — a full YouTube script (explainer/tutorial/video essay) engineered for watch-time: title + thumbnail concepts, a timed cold open, retention beats with B-roll cues, and chapter timestamps. Sibling to [short-form-script](skills/short-form-script/SKILL.md); for other languages, [professional-translator](skills/professional-translator/SKILL.md). *Earlier — v61.2, the builder's toolkit:* prove quality against real references, red-team robustness, test skills before a PR, run them offline, and sign what they produce. *v61.1, the demand side:* **[find](https://mohitagw15856.github.io/pm-claude-skills/find.html)** (task→skill router) with an honest per-skill **measured / no-claim** badge, per-skill versioning, and **[SkillSpec as an adoptable standard](conformance/REGISTRY.md)** (earnable `--badge` + public registry). Full history: **[CHANGELOG](CHANGELOG.md)** · [releases](https://github.com/mohitagw15856/pm-claude-skills/releases)
 
 ## 🤝 Contributing
 
@@ -171,4 +171,4 @@ MIT — use them, fork them, ship them at work. Skills are judgment, and judgmen
 
 ---
 
-*Built by [Mohit](https://github.com/mohitagw15856) with Claude. 750 skills · 90 bundles · 31 professions · every commit gated. The long version of this README — every feature, wave, and frontier bet — lives in the **[Showcase](docs/SHOWCASE.md)**.*
+*Built by [Mohit](https://github.com/mohitagw15856) with Claude. 751 skills · 90 bundles · 31 professions · every commit gated. The long version of this README — every feature, wave, and frontier bet — lives in the **[Showcase](docs/SHOWCASE.md)**.*

package/bin/cli.mjs

@@ -197,6 +197,7 @@ function add(opts) {
       else copyFileSync(f, dest);
       count++;
     }
+    if (opts.private) console.log('  note: --private overlay applies to native agents (claude/hermes/codex/openclaw); for rule-file agents, use `serve --private` or export your private skills first.');
   } else {
     for (const name of readdirSync(skillsDir)) {
       const src = join(skillsDir, name);
@@ -204,6 +205,21 @@ function add(opts) {
       placeDir(src, join(target, name), opts);
       count++;
     }
+    // Private overlay: a company's own skills merged on top of the public library
+    // WITHOUT forking. Same-named private skills override the public one; the
+    // rest are additive. (Mirrors `serve --private`; see docs/private-overlay.md.)
+    if (opts.private) {
+      const priv = resolve(opts.private);
+      if (!existsSync(priv)) { console.error(`Error: --private dir not found: ${priv}`); process.exit(1); }
+      let pc = 0;
+      for (const name of readdirSync(priv)) {
+        const src = join(priv, name);
+        if (!statSync(src).isDirectory() || !existsSync(join(src, 'SKILL.md'))) continue;
+        placeDir(src, join(target, name), opts);   // same name → overrides the public skill
+        pc++; count++;
+      }
+      console.log(`  ${opts.dryRun ? 'would overlay' : 'overlaid'} ${pc} private skill(s) from ${priv}`);
+    }
     // Claude Code also gets subagents, slash commands, and output-styles.
     if (agent === 'claude') {
       // Extras are siblings of `skills/` (…/.claude/{skills,agents,commands}),
@@ -387,6 +403,7 @@ Examples:
   npx pm-claude-skills search launch --json    # machine-readable (Raycast/Alfred/scripts)
 
   npx pm-claude-skills run prd-template --text "a referral program for B2B users"   # run a skill (needs ANTHROPIC_API_KEY)
+  npx pm-claude-skills test prd-template --runs 3   # author's test bench: score a skill vs its eval case + variance
   cat notes.txt | npx pm-claude-skills run meeting-notes --out summary.md           # pipe input, write the artifact
   cat notes.txt | npx pm-claude-skills chain run-discovery --deck                   # notes → 4 artifacts + a real .pptx
   npx pm-claude-skills install acme/their-skills --dry-run                          # audit a third-party skill repo
@@ -404,6 +421,11 @@ else if (cmd === 'search') search(opts);
 else if (cmd === 'find') find(opts);
 else if (cmd === 'changelog') changelog(opts);
 else if (cmd === 'add') add(opts);
+else if (cmd === 'test') {
+  const { run } = await import('./test.mjs');
+  try { process.exit(await run(process.argv.slice(3))); }
+  catch (e) { console.error(`Error: ${e.message}`); process.exit(1); }
+}
 else if (cmd === 'migrate') {
   const { run } = await import('./migrate.mjs');
   try { process.exit(await run(process.argv.slice(3))); }

package/bin/test.mjs

@@ -0,0 +1,110 @@
+// `pm-claude-skills test <skill>` — the author's test bench.
+// Run a skill against its eval case (evals/cases.json) with the library's judge
+// and see the score, the rubric breakdown, and variance across repeated runs —
+// locally, before you open a PR. Lowers the barrier to contributing a skill that
+// actually holds up.
+//
+//   npx pm-claude-skills test executive-update            # 1 run, scored
+//   npx pm-claude-skills test executive-update --runs 3   # variance across runs
+//   npx pm-claude-skills test executive-update --input "your own test input"
+//   npx pm-claude-skills test executive-update --dry-run  # no API; show the case
+// Needs ANTHROPIC_API_KEY (unless --dry-run).
+import { readFileSync, existsSync } from 'node:fs';
+import { join, dirname } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { complete } from './lib/anthropic.mjs';
+
+const ROOT = dirname(dirname(fileURLToPath(import.meta.url)));
+
+const RUBRIC = [
+  ['structure', 'Would a senior professional recognise the artifact\'s shape? Right sections, right order.'],
+  ['completeness', 'Are the load-bearing sections present and specific (not placeholder)?'],
+  ['usefulness', 'Could the reader act on it without a rewrite?'],
+  ['grounding', 'Does it use the input\'s facts, label assumptions, and avoid fabrication?'],
+];
+
+function parseArgs(argv) {
+  const o = { _: [] };
+  for (let i = 0; i < argv.length; i++) {
+    const a = argv[i];
+    if (a === '--dry-run') o.dryRun = true;
+    else if (a === '--json') o.json = true;
+    else if (a === '--runs') o.runs = Number(argv[++i]);
+    else if (a === '--input') o.input = argv[++i];
+    else if (a === '--model') o.model = argv[++i];
+    else if (!a.startsWith('--')) o._.push(a);
+  }
+  return o;
+}
+
+function loadSkill(name) {
+  const sj = join(ROOT, 'web', 'skills.json');
+  if (existsSync(sj)) {
+    const s = (JSON.parse(readFileSync(sj, 'utf8')).skills || []).find((x) => x.name === name);
+    if (s) return { title: s.title, body: s.body || s.instructions || s.description };
+  }
+  const p = join(ROOT, 'skills', name, 'SKILL.md');
+  if (existsSync(p)) { const raw = readFileSync(p, 'utf8'); return { title: name, body: raw.replace(/^---[\s\S]*?---\n/, '') }; }
+  return null;
+}
+
+function evalInput(name) {
+  const p = join(ROOT, 'evals', 'cases.json');
+  if (!existsSync(p)) return null;
+  const cases = JSON.parse(readFileSync(p, 'utf8')).cases || [];
+  const c = cases.find((x) => x.skill === name);
+  return c ? c.input : null;
+}
+
+const judgePrompt = (title, input, output) => `You are grading the output of a professional skill ("${title}") on a 1-5 rubric.
+
+RUBRIC:
+${RUBRIC.map(([k, d], i) => `${i + 1}. ${k} — ${d}`).join('\n')}
+
+THE INPUT THE SKILL WAS GIVEN:
+"""${input}"""
+
+THE SKILL'S OUTPUT:
+"""${String(output).slice(0, 8000)}"""
+
+Score each dimension 1-5. Return STRICT JSON only:
+{"structure": n, "completeness": n, "usefulness": n, "grounding": n, "overall": <mean, one decimal>, "fix": "the single highest-impact improvement, one sentence"}`;
+
+export async function run(argv) {
+  const o = parseArgs(argv);
+  const name = o._[0];
+  if (!name) { console.error('Usage: pm-claude-skills test <skill> [--runs N] [--input "..."] [--dry-run]'); return 1; }
+  const skill = loadSkill(name);
+  if (!skill) { console.error(`No skill named "${name}". Try: pm-claude-skills search ${name}`); return 1; }
+  const input = o.input || evalInput(name);
+  if (!input) { console.error(`No eval case for "${name}" in evals/cases.json — pass --input "..." to test it.`); return 1; }
+
+  if (o.dryRun) {
+    console.log(`Skill:  ${skill.title} (${name})`);
+    console.log(`Input:  ${String(input).slice(0, 200)}${input.length > 200 ? '…' : ''}`);
+    console.log(`Rubric: ${RUBRIC.map((r) => r[0]).join(', ')}`);
+    console.log(`Would run ${o.runs || 1}× and judge (needs ANTHROPIC_API_KEY).`);
+    return 0;
+  }
+  const key = process.env.ANTHROPIC_API_KEY;
+  if (!key) { console.error('Set ANTHROPIC_API_KEY (or use --dry-run).'); return 1; }
+
+  const runs = Math.max(1, o.runs || 1);
+  const scores = [];
+  for (let i = 0; i < runs; i++) {
+    const output = await complete({ apiKey: key, model: o.model, system: skill.body, messages: [{ role: 'user', content: input }], maxTokens: 2000 });
+    const raw = await complete({ apiKey: key, model: 'claude-sonnet-4-6', system: 'You are a strict, fair evaluator. Output only the requested JSON.', messages: [{ role: 'user', content: judgePrompt(skill.title, input, output) }], maxTokens: 400 });
+    let v; try { v = JSON.parse(String(raw).match(/\{[\s\S]*\}/)[0]); } catch { v = null; }
+    if (v) { scores.push(v); if (!o.json) console.log(`run ${i + 1}: ${v.overall}/5  (S${v.structure} C${v.completeness} U${v.usefulness} G${v.grounding})  — ${v.fix}`); }
+  }
+  if (!scores.length) { console.error('Judge returned no parseable scores.'); return 1; }
+  const mean = scores.reduce((a, s) => a + Number(s.overall), 0) / scores.length;
+  const overalls = scores.map((s) => Number(s.overall));
+  const spread = Math.max(...overalls) - Math.min(...overalls);
+  const verdict = mean >= 4 ? 'PASS' : mean >= 3 ? 'WEAK' : 'FAIL';
+
+  if (o.json) { console.log(JSON.stringify({ skill: name, runs: scores, mean: Number(mean.toFixed(2)), spread, verdict }, null, 2)); return mean >= 3 ? 0 : 1; }
+  console.log(`\n${verdict} — mean ${mean.toFixed(2)}/5${runs > 1 ? ` · spread ${spread.toFixed(1)} across ${runs} runs${spread >= 1 ? ' (high variance ⚠ — tighten the skill)' : ''}` : ''}`);
+  console.log(mean >= 4 ? 'Ship it.' : 'Below the L3 bar (4.0). Address the "fix" note above and re-test.');
+  return mean >= 3 ? 0 : 1;
+}

package/exports/README.md

@@ -8,7 +8,7 @@ by hand; edit the source skill and run:
 node scripts/build-exports.mjs
 ```
 
-Currently exporting **750 skills** to:
+Currently exporting **751 skills** to:
 
 - **ChatGPT — Custom GPT instructions** → `exports/chatgpt/`
 - **Google Gemini — Gem instructions** → `exports/gemini/`

package/exports/aider/README.md

@@ -3,7 +3,7 @@
 > Auto-generated from `skills/*/SKILL.md` by `scripts/build-exports.mjs`.
 > **Do not edit these files by hand** — edit the source skill and regenerate.
 
-750 skills exported. Copy a `.md rule` into the tool to use it.
+751 skills exported. Copy a `.md rule` into the tool to use it.
 
 | Skill | Bundle | Path |
 |---|---|---|
@@ -756,4 +756,5 @@
 | World Clock | `pm-live` | `pm-live/world-clock/world-clock.md` |
 | Writing Great Skills | `pm-engineering` | `pm-engineering/writing-great-skills/writing-great-skills.md` |
 | Writing Plans | `pm-method` | `pm-method/writing-plans/writing-plans.md` |
+| YouTube Script | `pm-creator` | `pm-creator/youtube-script/youtube-script.md` |
 | YouTube Script Writer | `pm-writers` | `pm-writers/youtube-script-writer/youtube-script-writer.md` |

package/exports/aider/pm-creator/youtube-script/youtube-script.md

@@ -0,0 +1,74 @@
+# YouTube Script Skill
+
+Long-form video is won twice: first the **packaging** earns the click, then the **script** keeps the promise the packaging made. A YouTube video does not fail in the middle — it fails at predictable cliffs (the first 30 seconds, the ~2-minute mark, the midpoint). This skill writes a script engineered to survive those cliffs: a cold open that pays off fast, value delivered early and escalated, and an open loop pulling the viewer across every segment.
+
+## Working from a brief
+
+Given a topic, a blog post, a talk transcript, or a rough outline, **write the full script anyway** — infer the single promise and the audience. Target the requested runtime (default ~8–12 min ≈ 1,200–1,800 spoken words; ~130–150 words/min). Never pad to hit a length; a tight 7 minutes beats a bloated 12. If the topic is genuinely short, say so and recommend [[short-form-script]] instead.
+
+## Required Inputs
+
+Ask for (if not already provided, else infer and label the assumption):
+- **Topic / the idea** (or a source: blog post, transcript, docs) and the **one promise** the video delivers
+- **Audience** and their level (complete beginner → practitioner)
+- **Target runtime** (~5 / ~10 / ~20 min) and **format** (explainer, tutorial, video essay, review, talking-head/vlog)
+- **Creator voice** (or pull from a [[creator-brand-kit]]) and the **primary CTA** (subscribe, a lead magnet, a product, the next video)
+
+## Output Format
+
+### The promise
+One sentence: what the viewer can do or understand by the end. Every segment serves this; anything that doesn't, cut.
+
+### Packaging (title + thumbnail)
+- **3–5 title options** — curiosity gap or clear payoff, front-load the keyword, ≤ ~60 chars.
+- **2 thumbnail concepts** — the visual + ≤ 3–4 words of thumbnail text. The title and thumbnail must not say the same thing (they should combine).
+
+### Cold open (0:00–0:30)
+Written word-for-word — it's the most important 30 seconds:
+- **Hook** — the tension, result, or claim that stops the click from bouncing. No "Hey guys, welcome back."
+- **The promise** — what they'll get and *why it's worth their next 10 minutes*.
+- **The stay reason** — an open loop or stake ("by the end you'll… but first the mistake almost everyone makes").
+
+### Script (segmented)
+
+For each segment (usually 3–6):
+
+| Segment | Spoken (VO / on-cam) | Visual / B-roll / graphics | Retention device |
+|---|---|---|---|
+| # — segment title | the actual words, conversational | screen-record, cutaway, chart, location change | open loop / callback / pattern interrupt / re-hook |
+
+- **Deliver value early** — front-load the best insight before the ~2-min cliff; don't save everything for the end.
+- **Escalate** — each segment raises stakes or specificity, and ends by opening the next ("that fixes X — but it creates a new problem…").
+- **Re-hook at the cliffs** — a pattern interrupt (b-roll, graphic, tone shift) near 0:30, ~2:00, and the midpoint.
+
+### CTAs
+- **Soft, integrated** — one early ask placed *after* the first real value lands (not before it).
+- **Primary CTA** — at the payoff, when goodwill peaks.
+
+### Outro & end-screen
+- A one-line payoff recap (deliver on the promise, explicitly).
+- End-screen: the **one** next video to watch (relevant, not random) + subscribe.
+
+### Description + chapters
+A 2–3 sentence description with the keyword, the primary link/CTA, and **chapter timestamps** (`0:00 Intro`, …) matching the segments.
+
+End with **▶ Automate:** a one-line note that [ContentGoldMine](https://github.com/mohitagw15856/ContentGoldMine) can generate this script (and a short-form cutdown) from a source URL.
+
+## Quality Checks
+
+- [ ] The cold open pays off within ~30s and opens a loop; no throat-clearing intro
+- [ ] The best value lands before the ~2-minute drop-off, not saved for the end
+- [ ] Every segment ends by opening the next (a visible chain of open loops)
+- [ ] A pattern interrupt / re-hook sits at each cliff (~0:30, ~2:00, midpoint)
+- [ ] Title and thumbnail combine (don't repeat) and match the promise the script keeps
+- [ ] One primary CTA at the goodwill peak; the early ask follows real value
+- [ ] Chapter timestamps match the segments; spoken word count fits the runtime
+
+## Anti-Patterns
+
+- A slow intro before the hook ("Hey guys, welcome back to the channel, so today…")
+- Saving all the value for the end so the first two minutes are setup
+- A wall of narration with no B-roll, graphics, or visual cues — it's a *video* script, not an essay
+- Clickbait packaging the script never pays off (kills trust and session time)
+- Asking to subscribe before delivering any value
+- Short-form pacing stretched thin, or long-form crammed — if it's 15–60s vertical, use [[short-form-script]]

package/exports/aider/pm-writers/notes-humanizer/notes-humanizer.md

@@ -59,8 +59,8 @@ CHANGES MADE
    → "X and Y. Z is different — [expanded thought]"
    Why: all three items had identical rhythm; broke the pattern
 
-4. Added short sentence: "That's the problem."
-   Why: needed a sub-8-word sentence to vary rhythm
+4. Split a 30-word sentence where the point turned: "…and it failed. That's the problem."
+   Why: the meaning lands on a short beat here — the brevity is earned, not inserted for variance
 
 5. Added sentence starting with "But"
    Why: human writers do this; AI avoids it as a statistical default
@@ -84,6 +84,16 @@ The full rewritten text, ready to copy and paste — no annotations, no formatti
 
 ## Instructions for Claude
 
+### Phase 0: Calibrate to genre (do this first)
+
+Before touching anything, decide what kind of text this is — the genre sets how far Phase 2 goes:
+
+- **Neutral / informational** (documentation, reference, summaries, status reports, official notices, most business comms): run **Phase 1 only**. Strip the tells and stop. Do **not** inject opinion, asides, or personality — a clean doc is the goal, not a voice. Injected "humanity" here reads as unprofessional.
+- **Personal / persuasive / social** (opinion pieces, blog posts, founder notes, LinkedIn, marketing, personal emails): run **Phase 1 + Phase 2**. Here earned voice is the point.
+- **Unsure or mixed:** default to the lighter touch (Phase 1, plus only the Phase 2 moves the content genuinely earns), and say which genre you assumed.
+
+The single rule underneath: **calibrate voice to the job of the text.** Never dismantle useful structure (a scannable doc, a real list) just to look less templated.
+
 ### Phase 1: Audit
 
 Read the full text before making any changes. Identify and count every instance of these patterns:
@@ -108,19 +118,19 @@ Read the full text before making any changes. Identify and count every instance
 
 Do not remove every em dash — only the ones used as parenthetical substitutes. Do not remove all hedging — only empty hedging that adds no information.
 
-### Phase 2: Inject
+### Phase 2: Inject (personal / persuasive genres only — see Phase 0)
 
-After stripping patterns, add the following signals. Each one should emerge from the actual content — don't add generic filler:
+Only for text where earned voice is the point. These are **options the content earns, not a checklist to complete** — apply the ones the material genuinely supports and **skip any that would be forced**. A move inserted to hit a quota is itself an AI tell. Never fake humanity: no invented typos, no forced slang, no staged messiness, and no chopping or padding sentences just to manufacture rhythm variance.
 
-1. **One genuine opinion or take.** The author appears to actually believe something specific. State it without hedging. ("This approach works, and I think most people underestimate how rarely the alternative does.")
+1. **A genuine opinion or take** *(if the author clearly holds one).* State the belief the text already implies, without hedging. Don't manufacture an opinion the author never expressed.
 
-2. **One specific detail, example, or number.** Ground the most abstract claim in the text with something concrete. If the text says "this happens frequently," add a real or illustrative number. If it says "many companies do this," name the type of company.
+2. **A specific detail or example** *(only if a real one exists).* Ground the most abstract claim in something concrete the author actually knows. **Never invent a number, quote, date, name, or fact to sound specific** — that's fake specificity, a worse tell than vagueness. If no real detail is available, ask the author for one or keep the honest abstraction; do not fabricate.
 
-3. **One aside or parenthetical thought that breaks the fourth wall slightly.** This is the signal most synthetic text lacks — the writer momentarily steps out of the formal argument to say something human. ("(I've seen this specific mistake made by people who absolutely should have known better.)")
+3. **An aside that steps out of the formal argument** *(where it fits the register).* The signal most synthetic text lacks — but only in genres that welcome it, and only when it connects to a real point. Forced or cutesy asides are worse than none.
 
-4. **At least one sentence under 8 words.** Make it land on a point, not a transition.
+4. **Let a sentence be short where the point lands on a short beat.** Follow the meaning, not a word count. Never shorten a sentence just to add "variance."
 
-5. **One sentence that starts with "And" or "But."** Place it where the rhythm earns it, not randomly.
+5. **Start a sentence with "And" or "But" only where the rhythm truly earns it.** If no place does, don't add one — a random "But" is a tell, not a fix.
 
 ### Phase 3: Report
 
@@ -136,20 +146,25 @@ Present the output in the four-section structure defined above. The change log m
 
 ## Quality Checks
 
+- [ ] Genre was assessed first (Phase 0); neutral/informational text got Phase 1 only, with no injected opinion, aside, or personality
 - [ ] Audit was completed before rewriting (patterns counted, not just detected)
 - [ ] Every removed pattern is listed in the change log with a specific reason
 - [ ] Em dashes were assessed individually — only parenthetical-substitute uses were removed
 - [ ] Rule-of-three lists: the rhythm was actually checked, not just the fact that there were three items
-- [ ] At least one sentence under 8 words was added (or was already present)
-- [ ] At least one sentence starts with "And" or "But" in the final text
+- [ ] No fact, number, quote, date, or name was invented to add specificity
+- [ ] Any Phase 2 moves applied were earned by the content — none inserted to hit a quota (short sentence / "And"/"But" opener / aside added only where the material genuinely called for it, or not at all)
 - [ ] The specific detail or example added connects to an actual claim in the text, not floated in generically
-- [ ] The aside breaks the fourth wall slightly without being forced or cutesy
+- [ ] The aside (if any) breaks the fourth wall slightly without being forced or cutesy
 - [ ] The change log lists specific instances, not categories
 - [ ] The clean output section has no annotations or formatting artifacts — ready to paste
 - [ ] If the original was already clean, that was stated explicitly rather than changes invented
 
 ## Anti-Patterns
 
+- [ ] Do not fake humanity: no invented typos, no forced slang, no staged messiness, and no programmatic sentence-length variation added to hit a target — faked signals are themselves AI tells
+- [ ] Do not inject voice into neutral genres — documentation, summaries, reports, and official comms get the tells stripped and nothing added; personality there reads as unprofessional
+- [ ] Do not invent specificity — never add a number, quote, date, name, or fact that isn't real to make prose sound concrete; ask for the detail or keep the honest abstraction
+- [ ] Do not dismantle useful structure (a scannable layout, a genuine list) just to look less templated
 - [ ] Do not remove all em dashes — only the ones functioning as parenthetical substitutes should be removed; genuine dramatic pauses are valid
 - [ ] Do not invent problems to justify changes when the original is already clean — report what was found honestly, even if the answer is "this text is mostly fine"
 - [ ] Do not add the aside or opinion generically — the injected human signals must connect to an actual claim or argument in the text, not float in as decoration

package/exports/chatgpt/README.md

@@ -3,7 +3,7 @@
 > Auto-generated from `skills/*/SKILL.md` by `scripts/build-exports.mjs`.
 > **Do not edit these files by hand** — edit the source skill and regenerate.
 
-750 skills exported. Copy a `SYSTEM_PROMPT.md` into the tool to use it.
+751 skills exported. Copy a `SYSTEM_PROMPT.md` into the tool to use it.
 
 | Skill | Bundle | Path |
 |---|---|---|
@@ -756,4 +756,5 @@
 | World Clock | `pm-live` | `pm-live/world-clock/SYSTEM_PROMPT.md` |
 | Writing Great Skills | `pm-engineering` | `pm-engineering/writing-great-skills/SYSTEM_PROMPT.md` |
 | Writing Plans | `pm-method` | `pm-method/writing-plans/SYSTEM_PROMPT.md` |
+| YouTube Script | `pm-creator` | `pm-creator/youtube-script/SYSTEM_PROMPT.md` |
 | YouTube Script Writer | `pm-writers` | `pm-writers/youtube-script-writer/SYSTEM_PROMPT.md` |

package/exports/chatgpt/pm-creator/youtube-script/SYSTEM_PROMPT.md

@@ -0,0 +1,74 @@
+# YouTube Script Skill
+
+Long-form video is won twice: first the **packaging** earns the click, then the **script** keeps the promise the packaging made. A YouTube video does not fail in the middle — it fails at predictable cliffs (the first 30 seconds, the ~2-minute mark, the midpoint). This skill writes a script engineered to survive those cliffs: a cold open that pays off fast, value delivered early and escalated, and an open loop pulling the viewer across every segment.
+
+## Working from a brief
+
+Given a topic, a blog post, a talk transcript, or a rough outline, **write the full script anyway** — infer the single promise and the audience. Target the requested runtime (default ~8–12 min ≈ 1,200–1,800 spoken words; ~130–150 words/min). Never pad to hit a length; a tight 7 minutes beats a bloated 12. If the topic is genuinely short, say so and recommend [[short-form-script]] instead.
+
+## Required Inputs
+
+Ask for (if not already provided, else infer and label the assumption):
+- **Topic / the idea** (or a source: blog post, transcript, docs) and the **one promise** the video delivers
+- **Audience** and their level (complete beginner → practitioner)
+- **Target runtime** (~5 / ~10 / ~20 min) and **format** (explainer, tutorial, video essay, review, talking-head/vlog)
+- **Creator voice** (or pull from a [[creator-brand-kit]]) and the **primary CTA** (subscribe, a lead magnet, a product, the next video)
+
+## Output Format
+
+### The promise
+One sentence: what the viewer can do or understand by the end. Every segment serves this; anything that doesn't, cut.
+
+### Packaging (title + thumbnail)
+- **3–5 title options** — curiosity gap or clear payoff, front-load the keyword, ≤ ~60 chars.
+- **2 thumbnail concepts** — the visual + ≤ 3–4 words of thumbnail text. The title and thumbnail must not say the same thing (they should combine).
+
+### Cold open (0:00–0:30)
+Written word-for-word — it's the most important 30 seconds:
+- **Hook** — the tension, result, or claim that stops the click from bouncing. No "Hey guys, welcome back."
+- **The promise** — what they'll get and *why it's worth their next 10 minutes*.
+- **The stay reason** — an open loop or stake ("by the end you'll… but first the mistake almost everyone makes").
+
+### Script (segmented)
+
+For each segment (usually 3–6):
+
+| Segment | Spoken (VO / on-cam) | Visual / B-roll / graphics | Retention device |
+|---|---|---|---|
+| # — segment title | the actual words, conversational | screen-record, cutaway, chart, location change | open loop / callback / pattern interrupt / re-hook |
+
+- **Deliver value early** — front-load the best insight before the ~2-min cliff; don't save everything for the end.
+- **Escalate** — each segment raises stakes or specificity, and ends by opening the next ("that fixes X — but it creates a new problem…").
+- **Re-hook at the cliffs** — a pattern interrupt (b-roll, graphic, tone shift) near 0:30, ~2:00, and the midpoint.
+
+### CTAs
+- **Soft, integrated** — one early ask placed *after* the first real value lands (not before it).
+- **Primary CTA** — at the payoff, when goodwill peaks.
+
+### Outro & end-screen
+- A one-line payoff recap (deliver on the promise, explicitly).
+- End-screen: the **one** next video to watch (relevant, not random) + subscribe.
+
+### Description + chapters
+A 2–3 sentence description with the keyword, the primary link/CTA, and **chapter timestamps** (`0:00 Intro`, …) matching the segments.
+
+End with **▶ Automate:** a one-line note that [ContentGoldMine](https://github.com/mohitagw15856/ContentGoldMine) can generate this script (and a short-form cutdown) from a source URL.
+
+## Quality Checks
+
+- [ ] The cold open pays off within ~30s and opens a loop; no throat-clearing intro
+- [ ] The best value lands before the ~2-minute drop-off, not saved for the end
+- [ ] Every segment ends by opening the next (a visible chain of open loops)
+- [ ] A pattern interrupt / re-hook sits at each cliff (~0:30, ~2:00, midpoint)
+- [ ] Title and thumbnail combine (don't repeat) and match the promise the script keeps
+- [ ] One primary CTA at the goodwill peak; the early ask follows real value
+- [ ] Chapter timestamps match the segments; spoken word count fits the runtime
+
+## Anti-Patterns
+
+- A slow intro before the hook ("Hey guys, welcome back to the channel, so today…")
+- Saving all the value for the end so the first two minutes are setup
+- A wall of narration with no B-roll, graphics, or visual cues — it's a *video* script, not an essay
+- Clickbait packaging the script never pays off (kills trust and session time)
+- Asking to subscribe before delivering any value
+- Short-form pacing stretched thin, or long-form crammed — if it's 15–60s vertical, use [[short-form-script]]

package/exports/chatgpt/pm-writers/notes-humanizer/SYSTEM_PROMPT.md

@@ -59,8 +59,8 @@ CHANGES MADE
    → "X and Y. Z is different — [expanded thought]"
    Why: all three items had identical rhythm; broke the pattern
 
-4. Added short sentence: "That's the problem."
-   Why: needed a sub-8-word sentence to vary rhythm
+4. Split a 30-word sentence where the point turned: "…and it failed. That's the problem."
+   Why: the meaning lands on a short beat here — the brevity is earned, not inserted for variance
 
 5. Added sentence starting with "But"
    Why: human writers do this; AI avoids it as a statistical default
@@ -84,6 +84,16 @@ The full rewritten text, ready to copy and paste — no annotations, no formatti
 
 ## Instructions for Claude
 
+### Phase 0: Calibrate to genre (do this first)
+
+Before touching anything, decide what kind of text this is — the genre sets how far Phase 2 goes:
+
+- **Neutral / informational** (documentation, reference, summaries, status reports, official notices, most business comms): run **Phase 1 only**. Strip the tells and stop. Do **not** inject opinion, asides, or personality — a clean doc is the goal, not a voice. Injected "humanity" here reads as unprofessional.
+- **Personal / persuasive / social** (opinion pieces, blog posts, founder notes, LinkedIn, marketing, personal emails): run **Phase 1 + Phase 2**. Here earned voice is the point.
+- **Unsure or mixed:** default to the lighter touch (Phase 1, plus only the Phase 2 moves the content genuinely earns), and say which genre you assumed.
+
+The single rule underneath: **calibrate voice to the job of the text.** Never dismantle useful structure (a scannable doc, a real list) just to look less templated.
+
 ### Phase 1: Audit
 
 Read the full text before making any changes. Identify and count every instance of these patterns:
@@ -108,19 +118,19 @@ Read the full text before making any changes. Identify and count every instance
 
 Do not remove every em dash — only the ones used as parenthetical substitutes. Do not remove all hedging — only empty hedging that adds no information.
 
-### Phase 2: Inject
+### Phase 2: Inject (personal / persuasive genres only — see Phase 0)
 
-After stripping patterns, add the following signals. Each one should emerge from the actual content — don't add generic filler:
+Only for text where earned voice is the point. These are **options the content earns, not a checklist to complete** — apply the ones the material genuinely supports and **skip any that would be forced**. A move inserted to hit a quota is itself an AI tell. Never fake humanity: no invented typos, no forced slang, no staged messiness, and no chopping or padding sentences just to manufacture rhythm variance.
 
-1. **One genuine opinion or take.** The author appears to actually believe something specific. State it without hedging. ("This approach works, and I think most people underestimate how rarely the alternative does.")
+1. **A genuine opinion or take** *(if the author clearly holds one).* State the belief the text already implies, without hedging. Don't manufacture an opinion the author never expressed.
 
-2. **One specific detail, example, or number.** Ground the most abstract claim in the text with something concrete. If the text says "this happens frequently," add a real or illustrative number. If it says "many companies do this," name the type of company.
+2. **A specific detail or example** *(only if a real one exists).* Ground the most abstract claim in something concrete the author actually knows. **Never invent a number, quote, date, name, or fact to sound specific** — that's fake specificity, a worse tell than vagueness. If no real detail is available, ask the author for one or keep the honest abstraction; do not fabricate.
 
-3. **One aside or parenthetical thought that breaks the fourth wall slightly.** This is the signal most synthetic text lacks — the writer momentarily steps out of the formal argument to say something human. ("(I've seen this specific mistake made by people who absolutely should have known better.)")
+3. **An aside that steps out of the formal argument** *(where it fits the register).* The signal most synthetic text lacks — but only in genres that welcome it, and only when it connects to a real point. Forced or cutesy asides are worse than none.
 
-4. **At least one sentence under 8 words.** Make it land on a point, not a transition.
+4. **Let a sentence be short where the point lands on a short beat.** Follow the meaning, not a word count. Never shorten a sentence just to add "variance."
 
-5. **One sentence that starts with "And" or "But."** Place it where the rhythm earns it, not randomly.
+5. **Start a sentence with "And" or "But" only where the rhythm truly earns it.** If no place does, don't add one — a random "But" is a tell, not a fix.
 
 ### Phase 3: Report
 
@@ -136,20 +146,25 @@ Present the output in the four-section structure defined above. The change log m
 
 ## Quality Checks
 
+- [ ] Genre was assessed first (Phase 0); neutral/informational text got Phase 1 only, with no injected opinion, aside, or personality
 - [ ] Audit was completed before rewriting (patterns counted, not just detected)
 - [ ] Every removed pattern is listed in the change log with a specific reason
 - [ ] Em dashes were assessed individually — only parenthetical-substitute uses were removed
 - [ ] Rule-of-three lists: the rhythm was actually checked, not just the fact that there were three items
-- [ ] At least one sentence under 8 words was added (or was already present)
-- [ ] At least one sentence starts with "And" or "But" in the final text
+- [ ] No fact, number, quote, date, or name was invented to add specificity
+- [ ] Any Phase 2 moves applied were earned by the content — none inserted to hit a quota (short sentence / "And"/"But" opener / aside added only where the material genuinely called for it, or not at all)
 - [ ] The specific detail or example added connects to an actual claim in the text, not floated in generically
-- [ ] The aside breaks the fourth wall slightly without being forced or cutesy
+- [ ] The aside (if any) breaks the fourth wall slightly without being forced or cutesy
 - [ ] The change log lists specific instances, not categories
 - [ ] The clean output section has no annotations or formatting artifacts — ready to paste
 - [ ] If the original was already clean, that was stated explicitly rather than changes invented
 
 ## Anti-Patterns
 
+- [ ] Do not fake humanity: no invented typos, no forced slang, no staged messiness, and no programmatic sentence-length variation added to hit a target — faked signals are themselves AI tells
+- [ ] Do not inject voice into neutral genres — documentation, summaries, reports, and official comms get the tells stripped and nothing added; personality there reads as unprofessional
+- [ ] Do not invent specificity — never add a number, quote, date, name, or fact that isn't real to make prose sound concrete; ask for the detail or keep the honest abstraction
+- [ ] Do not dismantle useful structure (a scannable layout, a genuine list) just to look less templated
 - [ ] Do not remove all em dashes — only the ones functioning as parenthetical substitutes should be removed; genuine dramatic pauses are valid
 - [ ] Do not invent problems to justify changes when the original is already clean — report what was found honestly, even if the answer is "this text is mostly fine"
 - [ ] Do not add the aside or opinion generically — the injected human signals must connect to an actual claim or argument in the text, not float in as decoration

package/exports/cline/README.md

@@ -3,7 +3,7 @@
 > Auto-generated from `skills/*/SKILL.md` by `scripts/build-exports.mjs`.
 > **Do not edit these files by hand** — edit the source skill and regenerate.
 
-750 skills exported. Copy a `.md rule` into the tool to use it.
+751 skills exported. Copy a `.md rule` into the tool to use it.
 
 | Skill | Bundle | Path |
 |---|---|---|
@@ -756,4 +756,5 @@
 | World Clock | `pm-live` | `pm-live/world-clock/world-clock.md` |
 | Writing Great Skills | `pm-engineering` | `pm-engineering/writing-great-skills/writing-great-skills.md` |
 | Writing Plans | `pm-method` | `pm-method/writing-plans/writing-plans.md` |
+| YouTube Script | `pm-creator` | `pm-creator/youtube-script/youtube-script.md` |
 | YouTube Script Writer | `pm-writers` | `pm-writers/youtube-script-writer/youtube-script-writer.md` |