video-brief

name	video-brief
description	Browse the research corpus for video ideas or generate a single-file HTML planning brief for a chosen idea. Browse mode prints a table in chat; brief mode writes ./briefs/<slug>.html to the local working directory. Pulls everything from the Connapse container via MCP.

Video idea browser + brief generator (local tool)

Two modes, selected from the user's prompt:

• Browse mode — answers "what should I film?" in chat. Reads the research corpus, prints a table, writes nothing to disk.
• Brief mode — resolves to a single candidate and produces a self-contained HTML page at ./briefs/<slug>.html for the human to open in a browser and plan a shoot from.

This skill runs locally in Claude Code, invoked on-demand by the human. It is not called by the cloud routines. The daily/weekly routines produce the research corpus; this skill consumes that corpus.

Briefs are personal planning artifacts, not part of the research corpus — they are never uploaded to the Connapse container. Do not call mcp__connapse__upload_file from this skill.

Invocation dispatch

Match the user's prompt to one bucket. If ambiguous between modes, or if a brief-mode prompt resolves to multiple candidates, ask before proceeding. Write no files until confirmed.

Browse mode triggers

Print a table (and a short follow-up) in chat. Write no files.

Prompt pattern	Reads	Output
"show me today's ideas" / "what should I film today" / "today's menu" / "give me ideas from today's research"	/daily/<latest>/candidates.md	Today's menu table
"show me this week's ideas" / "this week's menu" / "what's in the last 7 days" / "ideas this week"	last 7 /daily/*/candidates.md	Weekly menu table
"what trends this week" / "what's hot" / "weekly themes" / "what should I lean into"	latest /wiki/strategy/positioning-theses.md + /wiki/strategy/pillar-balance.md	Trends summary

Brief mode triggers

Resolve to exactly one candidate, then run §0–§6.

Prompt pattern	Resolver
"brief me on " / "brief "	Exact slug match in latest daily
"brief " (legacy IDs, pre-slug dailies)	Exact C-number match in latest daily
"brief today's top pick" / "brief #1"	Rank 1 in latest daily
"brief #" / "brief the -ranked"	Rank n in latest daily
"brief this week's top"	Highest weighted across last 7 dailies
"brief the top P" / "brief the P one"	Rank 1 filtered to pillar P in latest daily
"brief the undersupplied pillar"	Rank 1 in the pillar flagged UNDERSUPPLIED in latest pillar-balance
"brief the one" (e.g., "brief the token-diet one")	Fuzzy keyword match across hooks in latest daily
Free-text topic ("brief a video about X")	mcp__connapse__search_knowledge(path="/daily/", query=<text>, topK=3, mode="Hybrid")

If the request is ambiguous or multiple candidates match, return the top 3 and ask. Do not fall through to a different day silently.

---

Browse mode

Today's menu

Resolve containerId and read /daily/<latest>/candidates.md via mcp__connapse__get_document. Print a Markdown table:

#	Pillar	ID	Hook	Why it's live

• # — rank from the bottom-of-file ranking table (1 = highest weighted).
• Pillar — P1, P2, or P3.
• ID — slug for new dailies, C-number for legacy.
• Hook — the candidate's hook, verbatim from YAML.
• Why it's live — one sentence distilled from the candidate's rationale field.

Sort rows by rank (1 first). Include every candidate in that day's file.

After the table, add exactly ONE pro-tip line choosing the highest-leverage pick in this priority order:

1. News-cycle decay: if a news-reaction candidate (HN-cited in rationale) sits at rank ≤3, flag it: "News-cycle window closes ~<date+2d>; <id> decays fastest."
2. Undersupplied pillar: if pillar-balance flags a pillar as UNDERSUPPLIED and a candidate in that pillar exists in the file, flag it: "<id> is the P pick — filming it corrects the pillar imbalance."
3. Fallback: flag the rank-1 candidate: "<id> is the strongest weighted pick this cycle."

End the response with: Say "brief me on <id>" to generate a full HTML planning brief.

This week's menu

List /daily/ via mcp__connapse__list_files, take the 7 most recent date folders. For each, read candidates.md. Aggregate all candidates into one ranked list by weighted score.

If two candidates share a slug (recurring topic across days), merge into one row and annotate the Hook column with [recurring N×].

Columns: same as today's menu, with an added Date column (or "recurring" for merged rows).

If fewer than 2 dailies have run in the 7-day window, say so explicitly and fall through to the today's-menu view.

Weekly trends

Read /_manifest.md first to resolve the latest versions of strategy/positioning-theses and strategy/pillar-balance. Read both files.

Print each hypothesis as:

T —

• Actionable: <what this means for filming this week — be specific about which candidates or pillars benefit>

Below the hypotheses, append one line citing the undersupplied pillar from pillar-balance: This week's undersupplied pillar is P<n>: filming a P<n> video is the highest-leverage move.

End the response with: Want me to brief one of these? Say something like "brief the P3 one" or "brief today's top pick."

Write no files in any browse-mode path. The response ends at the in-chat output.

---

Brief mode

(Only enter §0 onward if brief mode was dispatched AND exactly one candidate was resolved.)

0. Setup

• Resolve and cache containerId via mcp__connapse__container_list (name: connapse-youtube-research).
• Verify Connapse connectivity. If the MCP is unreachable, abort and tell the user — do not write a stub HTML.

1. Resolve the candidate

Record for the resolved candidate:

• id — the candidate's ID (slug for new dailies, C-number for legacy)
• source_daily_date — the date folder the candidate came from
• The full YAML block (pillar, hook, angle, five score axes, rationale, cloud_compatible, promotion_surface)
• Weighted rank and score from the ranking table at the bottom of that day's candidates.md

Output path: ./briefs/<id>.html. Re-running for the same ID overwrites (intended — briefs reflect the latest research state).

Fuzzy keyword resolution: when the user asks for "the token-diet one" or similar, search hook texts (and slugs if present) for the keyword(s). If zero matches, list the available IDs and ask. If multiple matches, list them and ask.

By free text ("brief a video about X"): mcp__connapse__search_knowledge(path="/daily/", query="<user text>", topK=3, mode="Hybrid"). Pick the candidate in the highest-ranked chunk. Confirm the interpretation in the return message.

2. Pull evidence

The candidate's rationale cites concrete sources. Resolve them:

• YouTube references: mcp__connapse__get_document('/raw/youtube/<source_daily_date>/pull.json'). For each video ID or channel+title phrase in the rationale, capture video_id, title, channel_title, view_count, published_at. Keep the 5 most relevant. Video URL: https://www.youtube.com/watch?v=<video_id>.
• HN references: if the rationale cites HN #<id> or "Hacker News", fetch /raw/hn/<source_daily_date>/pull.json and match thread IDs or titles. Capture id, title, score, descendants, url. Thread URL: https://news.ycombinator.com/item?id=<id>.
• Trends: fetch /raw/trends/<source_daily_date>/pull.json entirely. Note the keyword with the highest positive slope relevant to the pillar; flag explicitly if all direction=flat.

3. Pull wiki context

Use /_manifest.md to find the current version for each topic, then mcp__connapse__get_document:

• /wiki/strategy/positioning-theses.v<latest>.md — find the hypothesis (T1–T5) matching this candidate's pillar + angle. Quote verbatim; cite by T-number.
• /wiki/strategy/pillar-balance.v<latest>.md — note the current count and whether the pillar is flagged UNDERSUPPLIED.
• /wiki/audience/<persona>.v<latest>.md — pick by pillar + surface:
  • P1, P2, any surface → claude-code-power-user
  • P3 or promotion_surface=oss|both → agentic-dev-indie
  • Ambiguous → include both.
• /wiki/voice/connapse-brand-voice.v<latest>.md — tagline, pain phrasing, hosted-vs-OSS framings, complete don't-say list.

4. Render HTML

Produce one self-contained HTML file — all CSS inline, no scripts, no CDN, no external images. Opens correctly offline. Print-friendly (max-width ~860px, system font stack, decent line height).

Required sections, in this order:

1. Header — <h1> = <id>: <hook>; subtitle with pillar badge (P1 blue, P2 red, P3 green), promotion_surface, weighted score, source daily date.

2. Angle — one paragraph verbatim from the candidate YAML.

3. Score breakdown — five rows (outlier_precedent, trend_slope, pain_density, audience_fit, signup_pull) + weighted composite. Each row: numeric score and horizontal bar (<span> with inline width: <score*200>px, solid color background).

4. Reference videos — table: Title (linked to youtube.com/watch?v=…), Channel, Views (formatted with thousands separators), Published (date only). 5 rows.

5. HN discussions — table if the rationale cited HN: Title (linked), Points, Comments, Age. Omit the section entirely if no HN citations.

6. Trend tailwind — short paragraph: which keyword is rising (if any) with slope value, or "No trend tailwind this cycle — all direction=flat."

7. Lead with these pain words — bullet list from the matching persona's pain signal words. Two labeled sub-lists if both personas were included.

8. Brand voice — canonical tagline in a <blockquote>; pain phrasing below.

9. DON'T say — red-accented list, verbatim from the brand-voice don't-say.

10. Positioning frame — "This candidate maps to T: ".

11. Pillar balance context — one sentence: current count + flag.

    • If this candidate's pillar is UNDERSUPPLIED → "→ filming this corrects the imbalance."
    • If this candidate's pillar is NOT the undersupplied one → "→ P is the undersupplied pillar this week; consider whether a P candidate should jump the queue."
    • If no pillar is flagged → "→ pillar mix looks balanced; proceed on weighted-score merits."

12. CTA copy — pick by promotion_surface:

    • hosted → "60-second setup", "free beta", connapse.com
    • oss → "Docker 60-second deploy", "MIT, self-hosted", GitHub link
    • both → both stacked, hosted first Suggest ≤20 seconds of end-screen copy.

13. Script outline scaffold — four slots, rendered as dashed-border <div> with grey italic placeholder text:

    • Hook (0–15s) — "Use the tagline verbatim or the pain phrasing — don't warm up, cut straight in."
    • Problem (15s–1:30) — "Show the pain on screen. One concrete metric visible."
    • Solution demo (1:30–7:00) — "Before/after with the same metric. Show Connapse doing the work."
    • CTA (last 30s) — "Per the CTA section above. Under 20 seconds. No outro music filler."

14. Footer — generation date, candidate source path (/daily/<source_daily_date>/candidates.md). Small grey text. No session URL needed — this is local.

CSS target: under ~120 lines, no vendor prefixes, system font stack, light theme, readable at A4 print.

5. Write locally

Create ./briefs/ relative to cwd if it doesn't exist. Write the HTML to ./briefs/<id>.html. Use standard file writing tools (Write); do not call MCP upload.

/briefs/ is already in the repo .gitignore. If the cwd is a different git repo that doesn't already ignore it, suggest (don't auto-apply) adding briefs/ to .gitignore.

6. Return

Report to the user:

• The local path: ./briefs/<id>.html (absolute path also fine)
• One-line summary: "Briefed <id> '<hook>' — <pillar> <surface>, ranked #<rank> score <score>. Open in your browser."
• If the pillar was UNDERSUPPLIED, mention it as a pro-filming signal.
• If you short-circuited on ambiguity in §1, no file was written — list the candidate options and ask which to brief.

Abort conditions

• Connapse MCP unreachable → abort, no stub file. Clear error to user.
• Requested candidate not found in the most recent /daily/<date>/candidates.md → list the available IDs and ask for a retry. Do not fall through to a different day silently.
• No /daily/<date>/candidates.md exists at all → tell the user the research routine hasn't run yet; point them at the cloud routine to produce a daily first.