`references/sources.md`

73 lines
# Sources — Axel Bitblaze retrieval map

## The account

| Field | Value |
|---|---|
| Handle | `@Axel_bitblaze69` |
| Display | Axel Bitblaze 🪓 |
| Bio | "Low iq idiot. I know nothing. Telegram: @Axelbitblaze" (self-deprecating shtick; the Telegram is a crypto-alpha funnel — ignore it) |
| Followers | ~128K (2026-07) |
| Volume | ~46.7K lifetime posts, ~11/day — a firehose; sweep by recency window, never "read everything" |
| Joined | April 2015 |
| History | 2015–2025: crypto/airdrop/trading alpha threads. 2026→: heavy pivot to Claude Code, MCP, local AI, agent side-hustles. Both streams still run concurrently. |

## Mirror map (X itself is auth-walled — WebFetch/Exa on x.com fails)

Ordered by usefulness; "verified" = actually returned content on 2026-07-22.

1. **`scripts/hydrate-tweet.js <id-or-url>`** — $0, no auth, computes the
   syndication token locally. First stop for any specific tweet ID. Exit 3 =
   tombstone (deleted/suspended) → fall through to mirrors.
2. **unrollnow** — `https://www.unrollnow.com/status/<tweet-id>` — VERIFIED.
   Full thread text, works within minutes of posting. Best per-thread mirror.
3. **rattibha** — `https://en.rattibha.com/Axel_bitblaze69` — VERIFIED.
   Profile-level thread archive with engagement counts; good for sweeps.
4. **vanlett** — `https://vanlett.com/Axel_bitblaze69` — VERIFIED. Profile
   mirror that includes REPOSTS (rattibha mostly doesn't) — the only mirror
   that shows what he amplifies vs. authors.
5. **vibepicked** — `https://vibepicked.com/builder/@Axel_bitblaze69/` —
   VERIFIED. Catalogs projects he built (with tools used, e.g. "Claude Code ·
   NotebookLM · Obsidian · skill-creator · yt-dlp"). Highest-signal page per
   byte: these are things he actually shipped, not just promoted.
6. **Exa search** — `mcp__exa__web_search_exa` with the handle in the query —
   METERED (~$0.005/search, show cost). Use for discovering new posts when
   the profile mirrors lag.
7. ~~threadreaderapp~~ — login-walled as of 2026-07. Skip unless others fail.

## Noise profile (why the classifier exists)

Rough historical mix: crypto/airdrop ≥ 60%, AI/Claude ~30%, misc ~10%. The
crypto stream is monetized (Telegram funnel, "like & RT" engagement bait) and
entirely out of scope. Never let a sweep's output mention it.

## Hype calibration

Observed claim patterns to always verify, never relay:
- Round star counts ("17,000 stars on GitHub") — check `gh-repo-stats.js`.
- Model-equivalence claims ("a 26B model matching 685B models") — discard;
  benchmark claims from this account carry zero weight (see
  `ollama-model-eval` for how Steve actually evaluates models).
- Income projections ("$8–15k more a year") — strip; evaluate the tool only.
- "Setup takes 2 minutes" — treat as marketing; estimate install cost from
  the repo README, not the tweet.

## Repost-source ledger

Accounts Axel amplifies whose original content scored well — candidates for
their own scout coverage. Append as found:

- (none yet)

## Seed post (the one that created this skill)

- `https://x.com/axel_bitblaze69/status/2078551452201754980` (2026-07-23) —
  promotes **video-use**: open-source Claude Code video-editing skill;
  claims 17K GitHub stars, transcript-first editing (cuts filler words /
  dead space), ffmpeg color-grade chains, editor sub-agents, plan-approval
  before cutting. Repo resolved at skill creation: **browser-use/video-use**
  — actual 17,582★, MIT, pushed 2026-07-01 (claim CONFIRMED; source is the
  reputable browser-use org). Full rubric scoring still pending a first
  sweep. Dedupe note: Steve's `hyperframes` family owns video *creation*;
  transcript-driven *editing of raw footage* is a plausible real gap.