Idea capture and evaluate
Skill clarkhager/cowork-skills/skills/idea-capture-and-evaluate
Score URLs against Clark's 12-column weighted rubric and append them to the Ideas Inbox database in Notion. Use this skill whenever Clark pastes one or more URLs in chat - especially with verbs like "evaluate", "score", "capture", "add to inbox", "research", "analyze these links", "run the rubric", "rank these", "deep dive on these", or even a raw URL dump with no explicit instruction (the implicit ask is always evaluate + capture, never ignore). Handles Instagram reels, Instagram carousel posts, GitHub repos, YouTube videos, Google Drive videos and PDFs, X/Twitter posts, Reddit posts, and generic web URLs. Auto-reads MEMORY.md when present to ground Wayfinder Fit scoring. Always runs mandatory verification of external factual claims before final scoring. Writes to the canonical Ideas Inbox database in Notion under the Jadyly Dev Shop page (creates the DB on first run, appends thereafter, skips duplicates with a note).From its SKILL.md
npx -y skills add clarkhager/cowork-skills --skill idea-capture-and-evaluateAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
12.8 KB, as published. Nobody here has run it
Idea Capture and Evaluate
Clark constantly finds ideas (tools, repos, reels, articles, videos) that he wants to evaluate and remember. Without a system, they get hopelessly bookmarked or opened in tabs and nothing comes of them. This skill is the system: paste URLs, get them classified, fetched, scored, verified, and persisted into a single canonical Ideas Inbox he can return to.
When this skill fires
Any time Clark pastes 1+ URLs in chat, regardless of explicit instruction. The default assumption is he wants evaluation. Specific high-confidence triggers:
- "evaluate these URLs" / "score these for me" / "rank these"
- "add to my ideas inbox" / "capture and evaluate"
- "run the rubric on these" / "research and score these" / "deep dive on these links"
- "what should I do with these" / "are any of these worth adopting"
- A raw URL dump with no verb (implicit ask: evaluate + capture)
If you see 1+ URLs and the user hasn't given an unambiguous non-evaluation instruction (e.g., "just open this and read it to me"), fire this skill.
What this skill does (high level)
- Read context - load MEMORY.md if it exists in the current working directory (grounds Wayfinder Fit scoring)
- Identify and classify - parse URLs from the user's message, classify each by content type
- Estimate cost - sum expected Apify spend across all URLs; warn if > $1, hard-stop if > $5 without confirmation
- Fetch content - per type via the right Apify actor or GitHub MCP; see
references/content-types.md - First-pass scoring - score each item against the 12-column weighted rubric; see
references/rubric.md - Mandatory verification pass - identify external factual claims, verify via WebSearch, update scores if claims were wrong; see
references/verification-pass.md - Duplicate check + Notion write - look up existing rows in Ideas Inbox by Source URL; if found, skip with a note; otherwise append with full rubric + rationale
- Synthesize and report - chat summary of each item with verdict + Notion row link
Step 1: Read context
Before processing any URLs, check for MEMORY.md in the current working directory.
if [ -f ./MEMORY.md ] || [ -f ./bizzabo-academy/MEMORY.md ]; then
# read relevant slices: active sprint, locked decisions, design discipline, methodology rows
fi
Use Grep to pull just the rows likely to inform Wayfinder Fit scoring (search for: "Wayfinder", "TDVR", "v2-elevated", "RFD", "multi-source", "Phase 3", "V1 seed"). Don't read the entire file - it's typically 25k+ tokens.
If MEMORY.md doesn't exist (skill running outside the bizzabo workspace), skip this step. Score Wayfinder Fit as 0 with a note: "Not in Wayfinder context; T2 Wayfinder Fit = 0 by default."
Step 2: Identify and classify URLs
Extract every URL from the user's message. For each URL, match against these patterns to determine type:
| Pattern | Type |
|---|---|
instagram.com/reels/ or instagram.com/reel/ | IG reel |
instagram.com/p/ | IG post (may be carousel) |
github.com/{owner}/{repo} | GitHub repo |
youtu.be/ or youtube.com/watch | YouTube video |
drive.google.com/file/ | Google Drive file (could be video or PDF - detect by trying video-intelligence first) |
twitter.com/ or x.com/ | X/Twitter post |
reddit.com/r/ | Reddit post |
| anything else | Generic web URL |
If a URL doesn't match any pattern, treat it as Generic web. Don't fail - fall through gracefully.
Step 3: Estimate cost
Quick reference (full table in references/content-types.md):
- IG reel: ~$0.005
- IG carousel: ~$0.10 + $0.01/slide
- GitHub repo: free (uses GitHub MCP)
- YouTube: ~$0.40-$0.75 (depends on length tier)
- Google Drive video: ~$0.40-$0.75
- Google Drive PDF: free (bash curl + pdf skill)
- Generic web: ~$0.005
Sum across all URLs. Apply guardrails:
- Estimate < $1: proceed silently
- Estimate $1-$5: report estimate to user in chat before fetching, then proceed
- Estimate > $5: report estimate and require explicit user confirmation ("ok", "yes", "proceed") before fetching
Step 4: Fetch content
See references/content-types.md for the full mapping including exact actor names, input schemas, and fallback chains. Brief summary:
| Type | Primary | Fallback |
|---|---|---|
| IG reel | apple_yang/instagram-transcripts-scraper | bulletproof/instagram-transcript-extractor |
| IG carousel | hikayatlabs/instagram-post-to-article | flag for manual review on INTERPRETATION_FAILED |
| GitHub repo | GitHub MCP get_file_contents (README + root dir) | chunked Read if README > 25k tokens |
| YouTube | marielise.dev/video-intelligence | akash9078/youtube-transcript-extractor |
| Google Drive video | marielise.dev/video-intelligence (handles Drive URLs) | bash + ffmpeg + local whisper |
| Google Drive PDF | bash curl https://drive.google.com/uc?export=download&id=$ID + pdf skill | - |
| Generic web | apify/rag-web-browser | WebFetch (provenance permitting) |
| X/Twitter | apify/rag-web-browser (works for public posts) | search Apify for dedicated actor |
apify/rag-web-browser with .json suffix | dedicated actor |
Critical patterns:
- Always set
previewOutput: falseoncall-actorto prevent overflow onget-actor-run - Use
get-actor-outputwithfieldsparameter to extract only the data you need (massive token savings) - Poll
get-actor-runto check status before fetching output - If output overflows even with
fields, save to file (paths returned in tool output) and Read selectively
Step 5: First-pass scoring
Apply the 12-column weighted rubric to each item. See references/rubric.md for full definitions, scoring guidance, and verdict thresholds. Brief summary:
Tier 1 (Clark's priorities, 2x weight, 0-6 each): Design Ceiling, Design Floor, Autonomy / 24-7, Cost Optimization
Tier 2 (standard, 0-3 each): Replaces/Upgrades, Adds Capability, Maturity Signal, Integration Cost, Jadyly Studios Fit, Wayfinder Fit
Tier 3 (risk guardrails, 0-3 each, 3 = good): Drift Risk, Reversibility
Total range: 0-48 Verdict thresholds: 35+ Adopt / 25-34 Pilot / 15-24 Hold / <15 Pass
Be honest. Don't pad scores to make items look better. Cost Optimization is intentionally hard to score high on (most items don't move compute spend). That's by design.
Step 6: Mandatory verification pass
Critical lesson from the May 15 session: first-pass scoring missed the Anthropic billing change being real because it was dismissed as "fear-based growth hack." See references/verification-pass.md for the full discipline.
Brief version: after first-pass scoring, identify external factual claims in the source content (especially: pricing claims, billing changes, performance benchmarks, vendor announcements, future-dated events, comparative numerical claims). For each material claim, use WebSearch to verify. Update scores if claims were wrong - either direction. Document verification attempts in the row's Sources Cited field with URLs.
If you find no verifiable claims (e.g., a personal story reel), note "no external claims to verify" in the row.
Step 7: Duplicate check + Notion write
The canonical Ideas Inbox lives at: parent page 3299123d-1fc8-80e0-8c9f-dff5174c208a (Jadyly Dev Shop).
First-time setup (only if Ideas Inbox doesn't exist yet):
- Use
notion-searchfor "Ideas Inbox" under that parent - If not found, create it via
notion-create-databasewith the schema inreferences/notion-schema.sql - Cache the data_source_id for the session
Duplicate check:
For each URL about to be written, query the Ideas Inbox for existing rows where Source URL matches. If a match exists:
- Skip the write
- Tell user in chat: "Already in inbox - scored X/48 on [date], verdict [verdict]. Want me to re-evaluate? (paste the URL again with 're-eval' to force)"
Write the row:
Use notion-create-pages with all 12 rubric columns filled, verdict + rationale, per-Tier-1-column notes, guardrail flags (use JSON array string format: "[\"DriftRisk\",\"LockIn\"]" for multiple), source URL, sources cited (with verification URLs), date scored, and the Source-Channel + Research-Session-Tag columns.
Step 8: Synthesize and report
Output a tight chat summary. Format:
Captured N URLs to Ideas Inbox. DB: <link>
- Item 1 title - score/48 - Verdict - one-line rationale - Notion row link
- Item 2 title - score/48 - Verdict - one-line rationale - Notion row link
- ...
Notable:
- [highest scorer] - why
- [anything verified-wrong from claims] - what the source got wrong
- [duplicates skipped] - count and list
Open items:
- [URLs that couldn't be fetched]
- [verification gaps that need manual eyeballs]
Don't over-format. Don't list scores in a giant table unless there are 10+ items.
Edge cases
- Failed fetch (actor returned 0 items, error, or empty transcript): flag the item in the chat output as "fetch failed - manual review needed"; still create a Notion row with Status = "Researching" so it's not lost
- Massive README (>25k tokens): use chunked Read with offset, or fetch a subset of files (just the README + directory listing); document in Sources Cited that full README wasn't read
- INTERPRETATION_FAILED on carousels: the multimodal LLM couldn't extract content; ask Clark to describe the carousel OR offer to use a different Instagram actor (e.g.,
futurizerush/instagram-profile-posts-scraperfor raw image URLs) - Cost overrun mid-fetch: if an actor unexpectedly costs more than estimated, stop, report, ask for confirmation before continuing
- Notion API multi-select validation failures: the API accepts
"single_value"for single multi-select option but requires JSON array string"[\"a\",\"b\"]"for multiple. Empty string is invalid - omit the field instead. - Verification surfaces a fact that changes verdict significantly: rewrite the verdict rationale to explain the verification finding; don't bury it
What NOT to do
- Don't score items without reading the source content. If fetch failed, mark as such; don't guess.
- Don't pad scores to push borderline items into a higher verdict tier. Honest 24/Hold > inflated 25/Pilot.
- Don't skip the verification pass even on items that look "obviously" promotional or "obviously" legitimate. The May 15 lesson: I dismissed a "fear-based growth hack" reel that was actually surfacing a real billing change.
- Don't write to Notion before verifying. If verification surfaces a material change, the row should reflect post-verification scores, not first-pass.
- Don't ask permission before fetching unless cost > $1. Clark wants this autonomous.
- Don't write a separate DB per session. Append to the canonical Ideas Inbox.
- Don't read all of MEMORY.md. Grep for the specific terms that inform Wayfinder Fit; reading the full file burns context.
Reference files
When you need detail beyond what's in this SKILL.md, read these:
references/rubric.md- full 12-column rubric definitions, scoring guidance per column, verdict tier definitions, examples from the May 15 sessionreferences/content-types.md- full URL pattern -> fetch method mapping with exact actor names, input schemas, cost estimates, fallback chains, known failure modesreferences/verification-pass.md- which claims warrant verification, how to verify each type, how to update scores after verificationreferences/notion-schema.sql- exact SQL DDL for creating the Ideas Inbox database
Don't read all of them on every invocation. Read the one you need for the step you're on.
Voice and writing rules for the chat output
Clark has explicit voice preferences:
- No em dashes - use space-hyphen-space
- No ALL CAPS, no "MUST", no "Absolutely", no "Definitely", no "Great question"
- No "straightforward", "It's worth noting", "It's important to note", "Let's dive in", "Let's break this down"
- No "In terms of"
- Direct and honest tone
- Match length to the task - tight when tight is right
- No filler sentences
Apply to all chat output the skill produces.