How do I install the Derive Thesis skill?

Run npx skillmds@latest add jperrello/derive-thesis in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Derive Thesis safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Derive Thesis?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Derive Thesis free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Derive Thesis?

jperrello (@jperrello) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Derive Thesis by jperrello — Agent Skill

Name: Derive Thesis
Author: jperrello

Derive Thesis

Name the CENTRAL SUBJECT (spine) of a long-form source so selection can tell on-theme moments from clip-shaped tangents. Claude reads the topic chapter list (topics.json) + source duration and emits thesis.json = {subject, thesis_sentence, key_threads[]} — the one durable source-level "what is this video about" artifact the pipeline lacked. Runs AFTER segment-topics and BEFORE pick-segments; pick-segments feeds it as a theme prior + scores each pick's theme_fit against it, and the confidence-floor backfill is gated on it so a high-standalone but off-spine tangent can no longer be auto-injected. Deterministic fallback (subject from the longest topic) on any failure. Non-fatal, idempotent, DERIVE_THESIS=0 skips.

jperrello 1d1cbe1 4 files · 7.5 KB Updated Sep 17, 2026

File contents

derive-thesis

A 2-hour talk wanders through one core throughline plus dozens of entertaining tangents. Nothing in the pipeline named the throughline, so pick-segments maximized standalone virality per span and shipped clip-shaped asides (a beetle anecdote, a DoorDash story) that misrepresent the source. derive-thesis computes that missing source-level subject once, as a durable JSON artifact both pick-segments (theme prior + theme_fit score + backfill gate) and selection-report read.

Invoke

.claude/skills/derive-thesis/derive-thesis.sh <transcript.json> [topics.json] [out.json]

transcript: full-source transcript (for duration; the digest is the topic list)
topics: segment-topics output (auto-discovered next to the transcript)
out: thesis.json (auto-discovered next to the transcript)

Output

thesis.json:

{
  "subject": "Donald Hoffman on consciousness and reality",
  "thesis_sentence": "Perception is a species-specific interface, not a window onto objective reality.",
  "key_threads": ["conscious agents", "fitness beats truth", "spacetime is not fundamental", "the hard problem"]
}

On any Claude/parse failure it writes a deterministic fallback (subject from the longest topic's title, key_threads from the leading topic titles, "fallback": true) so downstream always has a subject.

How

build_prompt.py renders the whole chapter list and asks Claude to name the SPINE (not summarize every chapter) — the guest + core theme, the throughline sentence, and the on-spine sub-themes. parse_reply.py validates and clamps, falling back to a topic-derived subject when the reply is empty/unparseable. Idempotent via mtime (topics.json → thesis.json); DERIVE_THESIS=0 skips.

---name: derive-thesisdescription: Name the CENTRAL SUBJECT (spine) of a long-form source so selection can tell on-theme moments from clip-shaped tangents. Claude reads the topic chapter list (topics.json) + source duration and emits thesis.json = {subject, thesis_sentence, key_threads[]} — the one durable source-level "what is this video about" artifact the pipeline lacked. Runs AFTER segment-topics and BEFORE pick-segments; pick-segments feeds it as a theme prior + scores each pick's theme_fit against it, and the confidence-floor backfill is gated on it so a high-standalone but off-spine tangent can no longer be auto-injected. Deterministic fallback (subject from the longest topic) on any failure. Non-fatal, idempotent, DERIVE_THESIS=0 skips.---# derive-thesisA 2-hour talk wanders through one core throughline plus dozens of entertainingtangents. Nothing in the pipeline named the throughline, so `pick-segments`maximized standalone virality per span and shipped clip-shaped asides (a beetleanecdote, a DoorDash story) that misrepresent the source. `derive-thesis`computes that missing source-level subject once, as a durable JSON artifact both`pick-segments` (theme prior + `theme_fit` score + backfill gate) and`selection-report` read.## Invoke```.claude/skills/derive-thesis/derive-thesis.sh <transcript.json> [topics.json] [out.json]```- `transcript`: full-source transcript (for duration; the digest is the topic list)- `topics`: `segment-topics` output (auto-discovered next to the transcript)- `out`: `thesis.json` (auto-discovered next to the transcript)## Output`thesis.json`:```json{ "subject": "Donald Hoffman on consciousness and reality", "thesis_sentence": "Perception is a species-specific interface, not a window onto objective reality.", "key_threads": ["conscious agents", "fitness beats truth", "spacetime is not fundamental", "the hard problem"]}```On any Claude/parse failure it writes a deterministic fallback (`subject` fromthe longest topic's title, `key_threads` from the leading topic titles,`"fallback": true`) so downstream always has a subject.## How`build_prompt.py` renders the whole chapter list and asks Claude to name theSPINE (not summarize every chapter) — the guest + core theme, the throughlinesentence, and the on-spine sub-themes. `parse_reply.py` validates and clamps,falling back to a topic-derived subject when the reply is empty/unparseable.Idempotent via mtime (`topics.json` → `thesis.json`); `DERIVE_THESIS=0` skips.

Frequently asked questions