`SKILL.md`

139 lines
---
name: tiktok-architect
description: Scout the Claude Code / vibecoding tips and anti-patterns that TikTok creator @tiktok_architect posts (e.g. "Never ask Claude Code to do this!"), VERIFY each claim against how Claude Code actually works, dedupe against Steve's existing rules, score it, and draft gated memos to codify the good ones into MEMORY.md / CLAUDE.md. Use when Steve says "tiktok architect", "/tiktok-architect", "what is tiktok_architect posting", "scout the claude tiktok tips", "any new Claude Code tips worth adopting", "check @tiktok_architect", or drops a tiktok.com/@tiktok_architect link (or a /t/ short link) and asks what the tip is and whether to adopt it. The feed is MIXED (motorcycles, hobbies, AND Claude tips) so it filters to the on-topic videos; tips are treated as CLAIMS to verify, never as facts to relay.
---

# TikTok Architect Scout

## Overview

@tiktok_architect is a TikTok creator whose Claude-Code content is a stream of
short tips and anti-patterns — *"Never ask Claude Code to do this!"*, *"Never
ever ask Claude for this"*, *"How to get Claude Code for free"*. This skill
harvests that stream, **verifies each claim against how Claude Code actually
works**, and turns the true, useful ones into **codify / watch / skip** decisions
for Steve's own standing rules — never a hype relay, and never an auto-edit of
his instructions. Retrieval and scoring run freely; changing MEMORY.md or
CLAUDE.md is gated.

This is the **rule-scout** counterpart to `gurisingh-scout` (the tool-scout).
Where that one adopts open-source repos, this one adopts *behavioral rules* — so
its first gate is **truth**, not stars.

## When to use

- A periodic sweep: "what's @tiktok_architect posted lately that's worth
  adopting as a rule?"
- A one-off: Steve drops a `tiktok.com/@tiktok_architect/...` or `/t/<short>`
  link and asks "what's the tip and should we do it?"
- A wider sweep of TikTok/Shorts Claude-Code-tip creators, using
  @tiktok_architect as the anchor plus any similar handles in
  `references/sources.md`.

## Workflow

### 1. Retrieve (fan out — TikTok is JS-rendered and hostile)
The video page HTML is useless (JS-rendered), so use the no-JS primitives.
Follow `references/retrieval-playbook.md`:

- **One dropped link** → clean caption + hashtags + author, `$0`:
  ```bash
  node scripts/tt-oembed.js "<tiktok-url-or-/t/short-link>"
  ```
- **"What's new" sweep** → newest videos, each `claude_relevant`-flagged, `$0`:
  ```bash
  node scripts/tt-list.js tiktok_architect 12
  ```
- **Caption too vague** → optional local transcript (`$0`, degrades honestly):
  ```bash
  bash scripts/tt-transcript.sh "<tiktok-video-url>"
  ```

### 2. Filter to the Claude/AI videos
The feed is **mixed** — motorcycles, hobbies, privacy, AND Claude tips. Scout
**only** the videos where `claude_relevant` is true. Do not treat every post as
a tip. (See the mixed-feed note in `references/sources.md`.)

### 3. Extract the candidate tip — DECODE THE IRONY FIRST
@tiktok_architect's *"Never ask Claude to do X"* is **reverse psychology**: it
means *"here's an X you can build."* Flip every "never/don't" to "here's how to"
and evaluate the REAL payload — never the literal surface. (Verified live
2026-07-19; the decoded payloads trend TOS-violating / harmful, so most land at
SKIP. See `references/sources.md`.) For each relevant video capture: **decoded
claim** (the real thing being taught, as an imperative), **kind** (anti-pattern /
best-practice / workflow / tool-repo / prompt-pack / setting), and the
**specifics** (the *why* + the *how*) from caption/on-screen/transcript. If the
tip names a repo, resolve + enrich it:
```bash
node scripts/gh-repo-stats.js owner/repo
```

### 4. VERIFY the claim (the correctness gate — do this before anything else)
A short-form tip is a CLAIM, and short-form claims are often outdated or wrong.
Decide **is this true for CURRENT Claude Code?** using, in authority order: the
`claude-code-guide` subagent, the `claude-docs` / `claude-api` skills (both `$0`),
official docs (docs.claude.com / docs.anthropic.com), then a second web source
via Exa/WebSearch if still unsure (metered — show cost). A claim that is false,
version-stale, or unverifiable is **SKIP**, with a one-line "why it's wrong".
Only verified-true claims proceed. (Full gate in `references/adoption-rubric.md`.)

### 5. Dedupe against what Steve already codified
Before crediting novelty, check for an existing equivalent rule:
```bash
grep -in "<keyword>" ~/.claude/CLAUDE.md 2>/dev/null
grep -rin "<keyword>" ~/.claude/projects/*/memory/ 2>/dev/null | head
grep -ril "<capability>" ~/.claude/skills/*/SKILL.md 2>/dev/null | head
```
If he already has it, the useful output is a **diff**, not a redundant re-add.

### 6. Score & decide
Apply the six-dimension rubric in `references/adoption-rubric.md` (correctness
confidence, stack fit, novelty, leverage, adoption cost, safety-of-acting).
Verdict bands: **≥14 CODIFY · 9–13 WATCH · ≤8 SKIP.**

### 7. Deliver — shortlist + gated codify-memos
Output a ranked table:
`Tip (claim) | Video (id, date) | Kind | Verified? (source) | Dedupe | Score | Verdict | reason`.
For each **CODIFY**, draft a memo to `~/.claude/yolo-queue/pending-approval/`
with: the claim restated as a rule in Steve's voice, **why it's true + the
source**, the exact proposed diff (a MEMORY.md feedback entry with frontmatter +
**Why:** / **How to apply:** lines, OR a specific CLAUDE.md line), adoption cost
(`$0` to codify), risks/caveats (loudly if the tip is dangerous), and an
APPROVE/REVISE/BLOCK line.

## Hard rules

- **Verify before you codify.** A tip is a claim, not a fact. An unverified or
  false "never do X" never becomes a rule — that's the whole point of this skill.
- **Never auto-edit `CLAUDE.md` or write a rule-memory unprompted.** Adopting an
  outside creator's claim as a standing rule is gated → pending-approval. The
  deliverable is a decision-ready shortlist; adoption is Steve's call.
- **Filter the mixed feed.** Only scout `claude_relevant` videos; ignore the
  motorcycles/hobbies posts.
- **A TikTok/throttle failure is not "nothing new."** Fall back to mirrors
  (`references/retrieval-playbook.md`); never conclude "no new tips" from one
  failed call.
- **Flag dangerous tips loudly.** Some tips ("disable permission prompts",
  "let it auto-push") are harmful — surface them as anti-recommendations, never
  quietly codify them.
- **Show costs** on every run (the scripts are `$0` local; Exa discovery is
  metered). **Log real wins only** — a tip Steve actually adopts, not the scan.

## Resources

- `scripts/tt-oembed.js` — one TikTok URL (incl. `/t/` short link) → caption +
  hashtags + author + `claude_relevant` flag. No JS, no auth, `$0`.
- `scripts/tt-list.js` — a creator's newest N videos, each `claude_relevant`-
  flagged, via yt-dlp flat playlist. `$0`. Exit 3 = throttled → use mirrors.
- `scripts/tt-transcript.sh` — a tip's spoken text: yt-dlp subs → local whisper
  on the audio → honest "no transcript" fallback. `$0`, optional.
- `scripts/gh-repo-stats.js` — GitHub repo → stars/license/freshness +
  Claude-skill detection, for tips that name an open-source tool. Honors
  `GITHUB_TOKEN`.
- `references/sources.md` — the anchor handle, the mixed-feed filter, the
  verify-the-claim principle, retrieval primitives, similar creators.
- `references/retrieval-playbook.md` — the ordered, JS-wall-resistant retrieval
  procedure and failure-mode reads.
- `references/adoption-rubric.md` — the correctness gate, the six-dimension
  scoring model, the dedupe procedure, gating rules, and output shape.