Clippos

Clippos Skill

Use this skill to run the local clipping engine end-to-end from a video path, attached video file, or video link. The engine does deterministic media work locally; the current agent supplies the semantic scoring step by reading scoring-request.json and writing scoring-response.json.

Inputs

Accept:

• A local video path, including a path resolved from an attached video file.
• A YouTube URL pasted directly. Auto-capped at 1080p height before

WhisperX transcription (4K @ 60 fps would OOM on most laptops).

• A direct http / https video URL (signed S3, CloudFront, plain mp4

hosts) — downloaded with urllib, validated with ffprobe before mining.

• A messaging-platform attachment URL (Discord CDN, Telegram api.telegram.org

bot-file URL, WhatsApp/Signal-provided HTTPS URL). On Discord and Telegram, when the user drops a video into the channel, the attachment appears in the message as {filename, url, size} — pass that url straight into advance --source. The helper detects CDN hosts and downloads them directly via urllib (yt-dlp is skipped for those, since they are signed direct mp4s).

• Any other URL yt-dlp supports (Twitch VODs/clips, Vimeo, X/Twitter,

Reddit hosted video, TikTok, Instagram, Facebook, 1000+ more sites) — not platform-specifically tested in Clippos beyond YouTube; should work since the download step is generic yt-dlp with a 1080p cap.

Optional user intent:

• Ratios: default all three (9:16, 1:1, 16:9). Respect explicit requests

such as "vertical only", "square", "wide", or --ratios 9:16,1:1.

• Clip count: default to config CLIPPOS_APPROVE_TOP or 5. Treat 5 as the

normal minimum when the video has enough valid candidate windows.

• Quality threshold: default to config CLIPPOS_MIN_SCORE or 0.70.
• Output directory: default to config CLIPPOS_OUTPUT_DIR or

~/Documents/Clippos.

If no video source is present, ask one short question for the video link or file path.

Slash-command shape

This skill is harness-agnostic. The surface differs per harness but the workflow below is the same.

Hermes exposes a single /clippos command and treats extra text as a subcommand or source argument:

• /clippos — run the main clipping workflow.
• /clippos config [options] — check or write local defaults.
• /clippos package [workspace] — generate publish packs after rendering.
• /clippos status — run the preflight config check.

Claude Code / Codex expose three slash commands via commands/*.md shims:

• /clippos — same main workflow.
• /clippos-config [options] — same as /clippos config.
• /clippos-package [workspace] — same as /clippos package.

When this SKILL.md says "use /clippos config" or "use /clippos package", Claude Code / Codex users substitute /clippos-config or /clippos-package. The underlying helper scripts are identical across harnesses.

Creator Profile (harness memory)

The skill is memory-aware. Whenever the harness has persistent memory (Hermes memory, Claude Code CLAUDE.md + ~/.claude memories, Codex equivalents), check it for creator-profile facts and apply them at the scoring and packaging handoffs. Creator-profile facts are the kind a content creator would tell the agent once and expect respected every run:

• Target platform and format (e.g. "TikTok-first, 9:16, 15–45s clips").
• Clip style (e.g. "story-beat > one-liner", "never pick intro music").
• Caption style (e.g. "clean, punchy, no fake hype, no all-caps").
• Brand tone (e.g. "indie-founder, self-deprecating, technical specifics").
• Banned phrases, emoji, or hashtags.
• Title/hook formatting preferences.

Treat these as contextual lens, not rubric overrides: the embedded rubric_prompt and response_schema remain authoritative. When no creator profile is in memory, score and package from the rubric alone and, when the run finishes, offer to save the user's stated preferences to the harness's memory tool for next time — do not write directly to ~/.hermes/memories/ or other memory stores.

Feedback Loop (self-improving profile)

The skill learns from the user's own keep/skip choices. After every render, advance emits a feedback_prompt with the clip IDs. Ask the user which clips they actually posted (or plan to post) and record the answer:

"$CLIPPOS_PYTHON" "$CLIPPOS_ROOT/scripts/hermes_clippos.py" feedback \
  "$WORKSPACE" --kept c1,c4 --skipped c2,c3 --note c2='too long'

Or, for structured payloads from the harness, use --json and pipe {"entries": [{"clip_id": "c1", "posted": true, "notes": "..."}, ...]} on stdin.

Each feedback call writes feedback-log.json in the workspace and appends rows to the global ~/.config/clippos/history.jsonl. On the next clippos run, advance attaches a creator_patterns section to the score and package handoff payloads. That section contains:

• summary — total clips, keep rate, per-ratio and per-spike rates.
• patterns — detected regularities, each with a rule (human-readable),

confidence (low / medium / high), sample_size, and a suggested_memory string.

Apply the patterns alongside rubric_prompt when scoring and packaging. High-confidence patterns (e.g. "user skips clips over 60s, 92% skip rate over 18 samples") should strongly bias rubric weights; low-confidence ones are informational. When a pattern is high confidence and not already in the harness memory, offer to save its suggested_memory via the harness's memory tool so the rule becomes a stable preference.

Never write directly to ~/.hermes/memories/. The skill captures outcomes; the harness owns the memory store.

Preflight

Run this before the first clippos job in a session. The prologue resolves the skill directory across harnesses:

• Hermes substitutes ${HERMES_SKILL_DIR}.
• Claude Code / Codex plugins substitute ${CLAUDE_PLUGIN_ROOT}.
• Any other harness must either pin CLIPPOS_ROOT directly or invoke the

prologue from inside the repo checkout so $PWD resolves correctly.

# Resolve CLIPPOS_ROOT — env var > harness substitution > known install
# locations > persisted config > $PWD. The candidate must contain
# scripts/hermes_clippos.py to be accepted.
# Why so many fallbacks: Hermes substitutes HERMES_SKILL_DIR reliably, but
# Claude Code's CLAUDE_PLUGIN_ROOT does not always expand inside command
# bash blocks (Anthropic issue #9354), so we also probe known install
# locations (Claude Code plugin cache, Codex plugin cache, Hermes skill
# dir) and the persisted CLIPPOS_ROOT in the config file.
_CLIPPOS_HERMES_HOME="${HERMES_HOME:-$HOME/.hermes}"
for candidate in "${CLIPPOS_ROOT:-}" "${HERMES_SKILL_DIR:-}" "${CLAUDE_PLUGIN_ROOT:-}" \
                 "$_CLIPPOS_HERMES_HOME/skills/clippos" \
                 "$HOME/.claude/skills/clippos" "$HOME/.codex/skills/clippos"; do
  if [ -n "$candidate" ] && [ -f "$candidate/scripts/hermes_clippos.py" ]; then
    CLIPPOS_ROOT="$candidate"; break
  fi
done
# Claude Code + Codex marketplace installs land under versioned caches whose
# exact shape can vary by harness. Search for the helper script and pick the
# newest matching checkout if no env var hit.
if [ -z "${CLIPPOS_ROOT:-}" ]; then
  for cache_root in "$HOME/.claude/plugins/cache" "$HOME/.codex/plugins/cache"; do
    [ -d "$cache_root" ] || continue
    candidate="$(
      find "$cache_root" -mindepth 2 -maxdepth 5 -type f \
        -path "*/scripts/hermes_clippos.py" -print 2>/dev/null \
      | while IFS= read -r script_path; do
          root="${script_path%/scripts/hermes_clippos.py}"
          [ -f "$root/SKILL.md" ] || continue
          mtime="$(stat -f "%m" "$root" 2>/dev/null || stat -c "%Y" "$root" 2>/dev/null || printf 0)"
          printf '%s\t%s\n' "$mtime" "$root"
        done \
      | sort -nr \
      | head -1 \
      | cut -f2-
    )"
    [ -n "$candidate" ] && [ -f "$candidate/scripts/hermes_clippos.py" ] && \
      { CLIPPOS_ROOT="$candidate"; break; }
  done
fi
if [ -z "${CLIPPOS_ROOT:-}" ] && [ -f "$HOME/.config/clippos/.env" ]; then
  CLIPPOS_ROOT="$(awk -F= '/^CLIPPOS_ROOT=/{gsub(/^["'"'"']|["'"'"']$/,"",$2); print $2; exit}' "$HOME/.config/clippos/.env")"
fi
[ -n "${CLIPPOS_ROOT:-}" ] && [ -f "$CLIPPOS_ROOT/scripts/hermes_clippos.py" ] || \
  { [ -f "$PWD/scripts/hermes_clippos.py" ] && CLIPPOS_ROOT="$PWD"; }
# Hard guard — fail fast and explicit. Without this, an empty CLIPPOS_ROOT
# falls into bash "/scripts/bootstrap-venv.sh" with a confusing error far
# from the real cause. Surface the install instructions instead.
if [ -z "${CLIPPOS_ROOT:-}" ] || [ ! -f "$CLIPPOS_ROOT/scripts/hermes_clippos.py" ]; then
  printf '[clippos] Could not resolve CLIPPOS_ROOT.\n' >&2
  printf '[clippos] Install Clippos with one of:\n' >&2
  printf '[clippos]   Hermes:      git clone https://github.com/dylan-buck/Clippos %s/skills/clippos\n' "$_CLIPPOS_HERMES_HOME" >&2
  printf '[clippos]   Claude Code: /plugin marketplace add dylan-buck/Clippos\n' >&2
  printf '[clippos]   Codex:       codex marketplace add dylan-buck/Clippos\n' >&2
  printf '[clippos] Or set CLIPPOS_ROOT=/abs/path/to/Clippos and re-run.\n' >&2
  exit 1
fi
# v1.x bootstrap: native plugin managers (Claude Code's /plugin, Codex's
# `codex marketplace add`) clone the repo but do not run pip — there is
# no PostInstall hook for engine extras. The bootstrap script creates a
# .venv at the install root and pip-installs the engine extras (~5 min,
# ~700 MB of wheels). Idempotent — no-op once a completed install marker
# exists, but resumes if a prior pip install failed after creating .venv.
bash "$CLIPPOS_ROOT/scripts/bootstrap-venv.sh"
CLIPPOS_PYTHON="${CLIPPOS_PYTHON:-$CLIPPOS_ROOT/.venv/bin/python}"
[ -x "$CLIPPOS_PYTHON" ] || CLIPPOS_PYTHON="$(command -v python3)"
"$CLIPPOS_PYTHON" "$CLIPPOS_ROOT/scripts/clippos_skill.py" config-check

If discovery fails (no env var, no install paths, no persisted config, not in the repo), persist the path once and every future invocation resolves cleanly:

"$CLIPPOS_ROOT/.venv/bin/python" "$CLIPPOS_ROOT/scripts/clippos_skill.py" \
  config-write --root "$CLIPPOS_ROOT"

bootstrap-venv.sh runs that step automatically as its last action.

Diarization (zero-config by default)

The skill ships with a zero-config open-source diarizer (silero-VAD + SpeechBrain ECAPA-TDNN + spectral clustering). No HuggingFace token, no license click-through. Models are public and auto-cache on first use.

CLIPPOS_DIARIZER (env var or --diarizer flag) chooses the path:

• speechbrain (default) — open-source, no setup. Recommended.
• pyannote — opt-in upgrade. Requires HF_TOKEN and one-time license

acceptance at . Slightly higher quality but the gate is a real onboarding cost.

• off — skip diarization entirely. Every segment gets SPEAKER_00. Use

for single-speaker content where speaker labels add no value.

Only mention the pyannote path if the user explicitly asks for higher diarization quality.

The real pipeline needs FFmpeg, ffprobe, an FFmpeg build with the ass subtitle filter (libass), and engine extras. HF_TOKEN is no longer required for the default open-source diarizer; the preflight reports it as missing but the run will succeed without it. Only ask the user for an HF token if they explicitly want the pyannote upgrade (see "Diarization" above).

Every subsequent bash block assumes CLIPPOS_ROOT and CLIPPOS_PYTHON are resolved with the same four-line prologue.

Main Workflow (agent loop)

Prefer the scripts/hermes_clippos.py driver (named for its Hermes-first design, but harness-agnostic — it works anywhere a Python script can shell out + read JSON). It advances the pipeline state machine and always prints a single JSON payload describing the next action, so each /clippos turn is exactly one tool call plus (when needed) one model handoff.

1. Start or resume a job:

"$CLIPPOS_PYTHON" "$CLIPPOS_ROOT/scripts/hermes_clippos.py" advance --source "$SOURCE"

Use --ratios 9:16,1:1 or --clips 2 only when the user asks. The payload contains workspace, next_action, and—when a model handoff is required— handoff_request_path and handoff_response_path.

1. When next_action == "brief" (v1.1):

• Read the request at handoff_request_path (brief-request.json).
• Follow its embedded brief_prompt and response_schema.
• Read the transcript_excerpt. The full transcript is provided when

short; for long videos, transcript_truncated: true and the excerpt contains the head + tail with a marker where the middle was dropped. Infer the global shape from what you can see plus speakers + duration_seconds.

• Author a tight, opinionated VideoBrief: theme, video_format,

3–5 expected_viral_patterns, 0–3 anti_patterns, optional audience / tone / notes. The brief must be scoring-relevant — what to up-weight and down-weight in this specific video — not a summary.

• Write handoff_response_path (brief-response.json) with a valid

VideoBriefResponse.

• Re-run advance on the workspace. The brief is cached for the rest

of the workspace's lifetime; the next per-clip scoring call sees the brief as video_brief on the scoring request.

This step is one model call per video and is the highest-leverage moment in the loop — it is where the agent's context-synthesis ability outperforms the per-clip rubric. Skip it only when next_action != "brief" (i.e. output_profile.video_brief: false in the job).

1. When next_action == "score":

• Read the request at handoff_request_path.
• Follow its embedded rubric_prompt and response_schema.
• Check loaded harness memory for any creator profile preferences before

scoring — target platform (TikTok/Reels/Shorts), preferred clip length, caption style, brand tone, hook style, and banned topics or phrases. Let those bias scores, spike categories, and penalties. The rubric stays authoritative; creator preferences act as tiebreakers and contextual lens.

• If the advance payload includes a creator_patterns field (populated from

past feedback), treat high-confidence patterns as strong bias signals and medium/low as informational. See the "Feedback Loop" section above.

• Score every ClipBrief, preserving clip_id and clip_hash verbatim.
• Write handoff_response_path with a valid ScoringResponse.
• Re-run advance on the workspace:

"$CLIPPOS_PYTHON" "$CLIPPOS_ROOT/scripts/hermes_clippos.py" advance --workspace "$WORKSPACE"

Advance builds the review manifest, auto-approves the top-scoring clips above the configured threshold, fills from the next-best clips when needed to reach the requested count, and renders the approved clips.

1. When next_action == "done-renders", the payload includes clips_dir,

clips[] with renders keyed by ratio, and a feedback_prompt with the clip IDs. Return the MP4 paths plus the clips directory and workspace path to the user, then ask which clips they kept or plan to post. Pipe the answer into hermes_clippos.py feedback (see the "Feedback Loop" section) so the creator profile keeps learning. Mention if any requested ratio was skipped.

1. When next_action == "error", surface the error string and the stage

it happened in. Do not retry silently—diagnose or ask the user.

Deterministic fallback (raw primitives)

If the harness cannot use hermes_clippos.py, the older step-by-step flow still works. Run prepare → clippos.cli run --stage mine → score → --stage review → clippos_skill.py approve → --stage render → clippos_skill.py outputs. See git history for the long form; hermes_clippos.py encodes the same sequence.

Packaging Workflow

Use /clippos package after a render finishes to produce per-clip publish packs (5+ title candidates, thumbnail overlay lines, a social caption, hashtags, and opening-line hooks). The hermes_clippos.py driver handles work

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dylan-buck
• Source: dylan-buck/Clippos
• License: MIT

Install and usage instructions live in the source repository linked above.