| name | distill-lessons |
| description | Crawl every episode transcript (chain → IPFS) and distill recurring themes + standout lessons into LESSONS.md and ALL-LESSONS.md. Use when the user asks to update the lessons files, mine the transcripts, or ask "what have we learned across episodes". |
Distill lessons from all episode transcripts
Four stages: fetch → extract per episode → synthesize → compose the flat
list. Stages 1 and 2 are cached on disk, so re-running after new episodes
only does the new work.
1. Fetch transcripts
node scripts/fetch-transcripts.mjs
Reads the episode list from the SlopComputer contract (mainnet), pulls each
manifest + transcript from IPFS into transcripts/ (gitignored). Idempotent —
cached by CID. Outputs per episode: raw <slug>.json (JSONL events),
<slug>.txt (rendered [h:mm:ss] speaker: text, speech only), <slug>.chat.json,
plus transcripts/index.json (titles, one-liners, participants, dates).
2. Extract per episode (fan out, cached)
For each <slug>.txt that has no transcripts/lessons/<slug>.md yet (or
whose transcript CID changed in index.json), launch one subagent per episode
— in parallel, one message. Each agent prompt:
Read transcripts/<slug>.txt — full transcript of a slop.computer livestream
episode: "" (<oneLiner from index.json>). Host: austingriffith.eth
(Austin Griffith); clawdbotatg.eth is Clawd, an AI cohost. Transcription is
imperfect — read through the noise.
Write a markdown extraction to transcripts/lessons/<slug>.md, ≤600 words,
starting with the line ## <slug>:
### Themes — 3-6 main themes, one line each.### Lessons — every concrete lesson/insight/strong opinion/advice worth
keeping (8-15 bullets), one sentence each, attributed (who, ~[h:mm:ss]).
Prioritize: working with AI/coding agents, crypto x AI, building &
shipping, career/philosophy. Skip pleasantries, logistics, show mechanics.### Quotes — up to 3 memorable near-verbatim quotes with speaker + timestamp.
After writing the file, reply with one line: the theme phrases, comma-separated.
3. Synthesize LESSONS.md
Read all transcripts/lessons/*.md (they're small) and rewrite LESSONS.md
at the repo root (committed — this is the deliverable).
Audience and framing (this is the key requirement): write for someone
learning how to be a sloperator — how to build real things with AI agents.
Not an analytical theme digest, not a dump: a teaching document. Distill the
BIGGEST, most important lessons — roughly 12-18 numbered lessons, ordered as
a learning path (start → workflow → craft → security → shipping → career).
Each lesson: a punchy bold headline stating the lesson, then a few sentences
of distilled guidance synthesized across episodes, tagged (slug) for the
episodes it came from. Weight lessons by how many episodes echo them, but a
single razor-sharp insight earns its place. Preserve real disagreements
between guests as "both are right at different stages" guidance rather than
false consensus. End with a short quote wall (best lines, speaker + slug) and
a footer: episode count + date + pointer to ALL-LESSONS.md + "regenerate with
the distill-lessons skill". Timestamps stay in the per-episode files only.
4. Compose ALL-LESSONS.md + pin to IPFS
python3 scripts/compose-all-lessons.py
bgipfs upload ALL-LESSONS.md --config ~/.bgipfs/credentials.json
The script concatenates every episode's Lessons bullets into ALL-LESSONS.md
(committed), one flat list, oldest episode first. The bgipfs upload (CLI +
credentials already set up on this machine; see https://www.bgipfs.com/SKILL.md)
pins it and prints a CID — the file is then live at
https://<CID>.ipfs.community.bgipfs.com/. Report the fresh CID to the user;
it changes on every regeneration.
Notes
transcripts/ is gitignored (refetchable public IPFS data); LESSONS.md,
this skill, and scripts/fetch-transcripts.mjs are committed.- To force a full re-extract (e.g. after changing the extraction prompt):
rm -rf transcripts/lessons and re-run stage 2.
- Contract address / gateway live at the top of
scripts/fetch-transcripts.mjs
(ETH_RPC_URL / IPFS_GATEWAY env overrides).
