`SKILL.md`

134 lines
---
name: axel-bitblaze-scout
description: Scout the Claude Code / AI-agent / local-AI tool and workflow posts that Axel Bitblaze 🪓 (@Axel_bitblaze69) publishes on X, filter out the account's heavy crypto/airdrop noise, VERIFY every claim (star counts, "free", capabilities) against the real repo, dedupe against Steve's existing stack, score fit/leverage/trust with a hype discount, and draft gated adopt/codify memos. Use when Steve says "axel bitblaze", "/axel-bitblaze-scout", "what is axel posting", "scout axel's feed", drops an x.com/axel_bitblaze69 link and asks what it is and whether to adopt it, or wants Axel's AI tool picks evaluated. Retrieval survives the X auth wall via mirror fan-out (unrollnow, rattibha, vanlett, vibepicked, syndication endpoint).
---

# Axel Bitblaze Scout

## Overview

Axel Bitblaze (@Axel_bitblaze69, ~128K followers, ~11 posts/day since 2015) is
a high-volume X account that historically posted crypto/airdrop alpha and now
posts a steady stream of **Claude Code workflows, MCP tools, local-AI stacks,
and agent side-hustles** — e.g. the Claude Code + NotebookLM + Obsidian
research pipeline, the `video-use` transcript-first video editor, Higgsfield
MCP real-estate packages, Mac-hardware-for-AI guides.

The signal is real but the delivery is engagement-farmed: round-number star
counts, "X matches a 685B model", "$8–15k more a year" — every number is a
CLAIM until verified. This skill harvests the AI posts, drops the crypto,
verifies the claims against the actual repos, and turns survivors into
**adopt / watch / skip** decisions for Steve's stack. Never a hype relay,
never an auto-install.

Sibling skills: `gurisingh-scout` (tool-drop educator), `tiktok-architect`
(rules/tips educator). Same gates, same rubric family — this one adds the
noise filter and hype discount that Axel's feed specifically needs.

## When to use

- Periodic sweep: "what has Axel posted lately that's worth adopting?"
- One-off: Steve drops an `x.com/axel_bitblaze69/status/<id>` link.
- Cross-check: another scout or the idea-loop surfaces a tool that Axel also
  covered — pull his write-up as one more (low-trust) source.

## Workflow

### 1. Retrieve (fan out — X is auth-walled)

Direct X reads fail unauthenticated. Follow `references/sources.md` for the
full mirror map. The proven order:

```bash
node scripts/hydrate-tweet.js <tweet-id-or-url>   # $0, syndication endpoint; exit 3 = tombstone
```

- Per-thread mirror: `https://www.unrollnow.com/status/<tweet-id>` (best —
  serves full thread text, verified working 2026-07).
- Profile archives: `en.rattibha.com/Axel_bitblaze69`,
  `vanlett.com/Axel_bitblaze69` (includes reposts), and
  `vibepicked.com/builder/@Axel_bitblaze69/` (his own built projects).
- Discovery of NEW posts: Exa search (`mcp__exa__web_search_exa`) scoped to
  the handle — metered, show the cost.
- `threadreaderapp.com` is login-walled as of 2026-07 — deprioritize.

### 2. Classify — the noise filter (Axel-specific, do this FIRST)

Bucket every post before spending any verification effort:

| Bucket | Examples | Action |
|---|---|---|
| `ai-claude` | Claude Code skills/workflows, MCP servers, subagent patterns | Proceed |
| `ai-general` | Ollama/local models, ChatGPT agent mode, AI hardware guides | Proceed |
| `hustle-promo` | "make $X with this AI side hustle" framing around a real tool | Proceed, but strip the income claims — evaluate only the tool |
| `crypto` | airdrops, trading bots, token threads, Telegram-channel funnels | SKIP silently — never surfaces in output |

Reposts: credit the ORIGINAL author, and note whether that account merits its
own scout entry in `references/sources.md`.

### 3. Extract + verify (claims are claims)

For each surviving post capture: tool name, repo URL (resolve `t.co`), what
it claims to do, and the specific verifiable claims (stars, license, "free",
capabilities). Then verify:

```bash
node scripts/gh-repo-stats.js <github-url-or-owner/repo>   # $0; GITHUB_TOKEN raises rate limit
```

Compare claimed vs. actual stars/license/freshness. A claim that misses by
>2x (e.g. "17,000 stars" on a 900-star repo) is itself a trust signal —
record the delta in the output.

### 4. Dedupe against Steve's existing kit

```bash
ls ~/.claude/skills/ | grep -i "<keyword>"
grep -ril "<capability>" ~/.claude/skills/*/SKILL.md 2>/dev/null
```

Steve has 400+ skills and a full agent cabinet. Video editing → `hyperframes`
family; research pipelines → `last30`, `deep-research`, `web-extract`,
`knowledge-layer`; local models → Ollama fleet + `prism`. If an equivalent
exists, the useful output is a **diff**, not a redundant recommend.

### 5. Score & decide

Apply `references/evaluation-rubric.md` — the six-dimension family rubric
(stack fit, novelty, leverage, install cost, trust/license, safety) with two
Axel modifiers: **Trust caps at 2/3 unless independently verified** (the
account's crypto-shill history and hype style earn a standing discount), and
**income/performance claims score 0 weight** (evaluate the tool, not the
pitch). Verdict bands: **≥14 ADOPT · 9–13 WATCH · ≤8 SKIP.**

### 6. Deliver — shortlist + gated memos

Ranked table: `Tool | Repo (claimed vs actual stars, license, last-push) |
What it does | Bucket | Dedupe | Score | Verdict | 1-line reason`. For each
**ADOPT**, draft a memo to `~/.claude/yolo-queue/pending-approval/` with what
it is, why it fits, exact install steps, **cost** (always show — scripts are
`$0 local`, Exa discovery is metered), risks, and an APPROVE/REVISE/BLOCK line.

## Hard rules

- **Never auto-install.** Cloning, `/plugin install`, `npx` on unfamiliar
  code, pasting into `~/.claude`, new paid keys — all gated to
  pending-approval. The deliverable is a decision-ready shortlist.
- **Crypto bucket never surfaces.** No airdrop/trading content in output,
  even as "interesting" — it's outside every mandate Steve has.
- **Every number is a claim until `gh-repo-stats.js` confirms it.** Report
  claimed-vs-actual deltas; never relay Axel's figures as fact.
- **A mirror failure ≠ "nothing new"** — fan out before concluding quiet.
- **Show costs** every run; **log real wins only** (an adopted pick, not a scan).

## Resources

- `scripts/hydrate-tweet.js` — one X post → verbatim text + links, no auth,
  `$0` (local syndication token). Exit 3 on tombstone.
- `scripts/gh-repo-stats.js` — repo → stars/license/freshness/Claude-skill
  detection; the claim-verification backbone.
- `references/sources.md` — handle profile, mirror map with verified status,
  noise profile, repost-source ledger, seed-post record.
- `references/evaluation-rubric.md` — the six dimensions, Axel trust
  modifiers, dedupe procedure, gating, output shape.