founder-os-pain-tagger

FounderOS pain tagger and pain-log quality check. Ranks raw posts from a watering hole with Laya so the founder reads the painful ones first (Sales Safari pre-filter), and checks the founder's pain lo

SKILL.md

GitHub

Pain tagger and pain-log check

Sales Safari is Amy Hoy and Alex Hillman's method (30x500). Use their terms and send the founder to their material. The founder reads their audience's own words, in the places that audience already talks, and logs the pain in quotes. Laya's job is the boring half: reading every post so the founder and you only read the ones worth reading.

When to run

  • Ranking: the founder has a batch of posts or comments from a watering hole, as a JSON list of {"id", "text"}.
  • Log check: the founder has logged pains with add_pain and wants to know if the log is good enough, or is heading to the stage 2 gate.

Laya pass

Ranking:

python3 "$FOUNDER_OS_DIR/scripts/audit.py" pain-tagger --input posts.json --lang <en|pt|...> [--jobs jobs.json]

Log check:

python3 "$FOUNDER_OS_DIR/scripts/audit.py" pain-log --founder-id <founder_id> --lang <en|pt|...> --record

Pass --lang every time. What the question set does with it comes from the measured run in new-jaxity #603:

  • is_pain runs on the english checkpoint for English text (AUC 0.78). On the multilingual checkpoint it was at chance (AUC 0.53), so for any other language its trust is chance, ranking falls back to money_or_workaround, and a quote is never flagged as not-pain on that basis. You read those for pain yourself.
  • money_or_workaround runs on multilingual for every language (AUC 0.84). Use it to fill the "what they buy, what they built" column.
  • intensity and pain_category are advisory. The category options are the founder's own job clusters, passed as --jobs for ranking and taken from the log for the check. Do the clustering yourself; Laya's pick is a hint.
  • Long posts are split into chunks that fit the checkpoint and the answer is taken from the strongest chunk, so a pain at the end of a long post is not cut off. excerpt shows which chunk carried it.

Claude pass

Ranking. Read the top of the list, not all of it: roughly the top 30 or the first quarter, whichever is smaller. Pull quotes in the author's own words. Say which posts Laya put high that are not pain (launches, advice, opinions about "people") so the founder learns the difference. For each quote you keep, help the founder add_pain with source_url, watering_hole and a job.

Log check. Go through the issues per quote and the log block:

  • not_a_pain: the quote reads as news, advice, a launch or an opinion. Check it; drop it or replace it with the author's actual complaint.
  • not_a_pain_unverified: Laya's is_pain was at chance for this language. Judge it yourself.
  • no_source_url, no_watering_hole, no_job, duplicate_of:, too_short.
  • thin_jobs (fewer than 3 quotes), too_few_watering_holes (fewer than 3), unclustered, and money_or_workaround_share: a log where nobody has paid or hacked around the pain is a warning sign about the pain's size.

End with a verdict: is this log strong enough to cluster into jobs, and what is the single most useful thing to go and collect next?

Without Laya

If mode is claude-only, ranking returns the posts unscored: read them all if there are few, or ask the founder to cut the batch. The log check still runs every structural check (sources, jobs, duplicates, counts); only the Laya checks are missing. Judge pain yourself and say in one line that Laya was unavailable. Write no labels.

Label capture

Ranking writes no labels, because raw posts are not ledger rows. Labels are written by pain-log --record, one per quote and question, on that quote's pains row. Ask the founder to confirm the tags you disputed and record them:

python3 "$FOUNDER_OS_DIR/scripts/labels.py" accept --founder-id <id> --by <founder_id> <label_id> ...
python3 "$FOUNDER_OS_DIR/scripts/labels.py" reject --founder-id <id> --by <founder_id> <label_id>
python3 "$FOUNDER_OS_DIR/scripts/labels.py" reject --founder-id <id> --by <founder_id> <label_id> --value 2

is_pain corrections on English text are the labels the fine-tune needs most. The job category is not recorded: its options are this founder's own, so it cannot be replayed as a training example. The cross-founder export is operator-only.

Ledger writes

  • pains rows, by the founder with your help, only from quotes that survive.
  • Labels from pain-log --record; corrections from labels.py.
  • One audits row per quote you judged a problem: record_audit.py --table pains --id <pain_id> --auditor claude --check pain_quality --verdict flag|fail --findings "<why>".