collab-finder

---
name: collab-finder
description: Scores your niche's accounts and writes a pitch for the top ten. Use for partners, guest swaps, joint carousels or shout-outs.
---

# Find the accounts worth a collaboration and pitch them

Ten scored accounts, each with a reason and a pitch.

## Claude does

1. Reads `voice.md` and `about-me.md` if present, then asks only for what is missing: platform, niche, follower count, median engagement over your last 20 posts, audience, and what you offer a partner.
2. Bands by creator tier: nano 1k-10k, micro 10k-100k, mid 100k-500k, macro 500k-1M, mega 1M+. Targets 0.5x-3x your followers; below is charity, above rarely replies.
3. Builds 40 public candidates: who your audience follows, repeat commenters on your posts and on the niche's three biggest accounts, and anyone who collaborated publicly in the last 90 days; past collaborators say yes again.
4. Scores each 1-5 on five criteria, 25 maximum: audience overlap (same buyer, not the same content), size fit, engagement rate, collaboration evidence, offer fit. Engagement is reactions + (comments x 3) over followers, never follower count. Shortlist at 18+.
5. Refuses to guess. Any figure it has not read stays "unknown" and that candidate goes to the research list, not the pitch list.
6. Writes each pitch: a dated post of theirs it has read, one line on what you bring with a number, the format and dates, a yes or no close. DM 80-120 words, subject under 50 characters.
7. Proposes what the 2026 data supports: a joint or swapped carousel (LinkedIn 21.77% median engagement vs 7.35% video; Instagram carousels 9x the saves of single images), from personal profiles, +63% over company pages.

## Then Codex does

Codex takes the bulk: the 40-row scored table, the template applied per candidate, duplicate handles merged, the tracker written. Repetitive, no judgement.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Claude fills in first: {MY_PROFILE}, {CANDIDATE_LIST}, {OFFER}, {FORMAT}, {SCORING_RULES}, {PITCH_TEMPLATE}.

## Claude checks

1. Every row has a public source URL. No URL, strike the row.
2. Every pitch names a real dated post of theirs. Open it.
3. Every statistic traces to Buffer or Metricool 2026 or the user's export; "2.5x more shares" and "+27% completion" trace to nothing.
4. Recompute two engagement scores by hand. Follower order means the scoring never ran.
5. Names and handles read character by character.

Reject and rerun: more than two failures of ten sends the batch back, rows named. Never hand-patch.

## Rules

- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.
- No scraped or guessed emails. DM, or a publicly listed address.
- No hashtags: measured at -31.70% views and -33.89% interactions.
- One merge field is not personalisation. Never send two identical pitches.
- One follow-up after five working days, then stop.

## If Codex is not installed

Claude does the bulk half itself, more slowly, and says so.

## Built from

- coreyhaines31/marketingskills, https://github.com/coreyhaines31/marketingskills, 46,945 stars via api.github.com: partner scoring, tiers.
- charlie947/social-media-skills, https://github.com/charlie947/social-media-skills, 3,283 stars via api.github.com: engagement formula.
- Metricool LinkedIn 2026, https://metricool.com/press-release-linkedin-study-2026/, 673,658 posts.
- Metricool Instagram 2026, https://metricool.com/press-release-instagram-study-2026/, 24,364,803 posts.
- Buffer, State of Social Media Engagement 2026, https://buffer.com/resources/state-of-social-media-engagement-2026/, 52M+ posts.

# Find the accounts worth a collaboration and pitch them - Codex task

You are producing a collaboration shortlist and outreach drafts for one social media account owner. You have no memory of the conversation that produced these inputs. Work only from what is below.

## You are given

- {MY_PROFILE}: the owner's platform, niche, follower count, median engagement over their last 20 posts, and audience description.
- {CANDIDATE_LIST}: up to 40 candidate accounts, each with a handle and at least one public URL. Some rows have missing fields marked "unknown".
- {OFFER}: what the owner can give a partner.
- {FORMAT}: the collaboration format to propose, with dates.
- {SCORING_RULES}: the five criteria and how each 1-5 score is defined.
- {PITCH_TEMPLATE}: the approved message structure to fill per candidate.

## Produce

1. `candidates.md`: one table, every candidate from {CANDIDATE_LIST}, columns: handle, public URL, followers, engagement rate, five criteria scores, total out of 25, band (nano/micro/mid/macro/mega), verdict. Engagement rate = (reactions + comments x 3) / followers, shown to two decimals. Any row with an "unknown" in a scored field gets verdict "research", total left blank.
2. `shortlist.md`: the ten highest totals scoring 18 or above, ranked, each with a two-line reason naming the specific overlap with {MY_PROFILE}.
3. `pitches.md`: one pitch per shortlisted account, using {PITCH_TEMPLATE} and {OFFER} and {FORMAT}. DM version 80-120 words. Email version with a subject line under 50 characters. Each pitch names the dated post supplied for that candidate and links it.
4. `tracker.csv`: handle, url, total, band, channel, pitch_sent_date (blank), reply (blank), followup_due (blank).

## Rules

- Public information only. Never add an email address, phone number or fact that is not in the inputs.
- Never invent a number, a follower count, a quote or a post title. Missing stays "unknown".
- Every pitch ends with the owner's own name, exactly as given in {MY_PROFILE}.
- No hashtags anywhere in the drafts.
- No statistic in any pitch unless it appears in {MY_PROFILE} or {OFFER}.
- No two pitches may share an opening sentence. Vary structure, not just the name.
- Merge duplicate people appearing under two handles into one row, keeping both URLs.
- British English throughout. No em dash characters.

## Return

Print, in this order:
1. The four filenames written, with absolute paths.
2. Counts: candidates scored, candidates marked research, shortlisted, pitches written.
3. The full `shortlist.md` table.
4. Any candidate you dropped and the one-line reason.
5. A line reading `UNKNOWNS:` followed by every field you could not fill.