Influencer Discovery
Find creators who bring their own audience — and get their contact info
influencer-discovery is a creator-discovery pipeline that runs as a Claude Code skill. It searches 15 channels for people who bring their own audience — followers, readers, subscribers — pulls their public contact info, and appends them to a Google Sheet. People building their own product are filtered out: they want reach, they don't provide it.
The operating manual is skills/influencer-discovery/SKILL.md. Per-platform capability boundaries (what each API exposes, which credentials it needs, what it costs) live in reference/datalayer/; how contact info is obtained per channel lives in reference/methodology/, where the directory-name prefix is the cooperation priority.
Channels
| Tier | Channels |
|---|---|
| 1 · Social | X/Twitter, Instagram, TikTok, Threads, YouTube, Reddit, Mastodon |
| 2 · Blog platforms | DEV.to, Hashnode, WordPress.com, Micro.blog |
| 3 · Personal sites | Self-hosted blogs, newsletters, podcasts |
| 5 · Media | freeCodeCamp News, HackerNoon |
Fetching is polite and read-only: explicit User-Agent, throttled, no anti-bot circumvention. Some channels work with zero credentials; others need an API key or a logged-in session — the per-channel requirements are in datalayer.
Install
No third-party packages, no build step — the pipeline is pure standard library
(Python ≥ 3.11). The skill is fully self-contained under
skills/influencer-discovery/; clone and run, or copy that one directory into
your agent's skills folder:
git clone https://github.com/tigerless-labs/influencer-discovery && cd influencer-discovery
python3 skills/influencer-discovery/scripts/run.py --help
Credentials go in ~/.config/influencer-discovery/.env, never in the repo. Pipeline state lives in
~/.local/share/influencer-discovery/; both paths are overridable via INFLUENCER_DISCOVERY_CONFIG_DIR /
INFLUENCER_DISCOVERY_STATE_DIR. Knobs that travel with the code (channel list, throttle parameters,
header mapping) stay in skills/influencer-discovery/config/.
Run
python3 skills/influencer-discovery/scripts/run.py --tiers 1 --per-channel 10
--channels / --tiers select channels, --subject sets the topic gate,
--min-followers / --min-karma set audience floors. Each run prints a report: planned vs
actual, verdict distribution, and the contactable list.
Sheets access
Set INFLUENCER_DISCOVERY_SPREADSHEET_ID and authenticate with a short-lived token minted by impersonating
a service account — no long-lived key file on disk. Google blocks adding the spreadsheets
scope to plain gcloud ADC ("This app is blocked"), so mint via the iamcredentials API:
T=$(gcloud auth application-default print-access-token)
curl -s -X POST \
-H "Authorization: Bearer $T" -H "Content-Type: application/json" \
-d '{"scope":["https://www.googleapis.com/auth/spreadsheets"],"lifetime":"3600s"}' \
"https://iamcredentials.googleapis.com/v1/projects/-/serviceAccounts/YOUR_SA_EMAIL:generateAccessToken"
Prerequisites: the project has the sheets, iam, and iamcredentials APIs enabled; your ADC
identity holds roles/iam.serviceAccountTokenCreator on the service account; the service
account can edit the target Sheet. A fresh IAM binding takes ~30s to propagate — a first 403 is
normal, retry.
Design invariants
Dedup key is (person, platform) | Checked against a local log; URLs are never parsed for identity — same-site different-person collisions taught us that. |
| Sheet writes are append-only | The pipeline creates rows, never edits an existing cell, and aborts if the header row doesn't match its mapping. |
| Blast radius is one record | A malformed page or link costs that one person, never the channel or the run; intermediate results are flushed as they are produced. |
| Hostile-input posture | Page content is data, not commands — instruction-shaped text in bios and pages is never executed. Every external request leaves through a single read-only choke point. |
| Contact data never enters the repo | Real names, emails, and handles live only in the Sheet; data/ is gitignored end to end. |