Evaluate Source

Evaluate a source

You've been pointed at external content (article, repo, talk, tweet, other-system docs). The question: what idea is in here that could make your system better, and does it map to you? Not "should we adopt their stack."

Get the content

The source is the argument (URL/path), or whatever the user pasted. One source per invocation — run twice to compare two.

• URL / public docs — use the host's web retrieval capability. Local file / PDF — use the host's file reader; for large PDFs, read only the pages needed.
• GitHub — use gh repo view, gh api repos///contents/, raw URLs through the host's web retrieval capability, or a shallow local checkout for a deep code dive.
• Tweet / X — see Fetching reference below.

If retrieval returns nothing useful, stop and ask the user for the content — don't speculate about what the source says.

Frame

• Strip the implementation, keep the idea. Their stack, tools, and library names don't matter. What's the underlying principle?
• Anchor against your system. Read the project's AGENTS.md / README, architecture docs, decision records (ADRs), and guardrails. The lens: "would we be different if we'd known this two months ago?"
• Durable, not trendy. An honest "nothing here for us" is a valid result — most external content has nothing for you. Don't manufacture relevance.

Evaluate — per idea, not per source

A source may hold 0, 1, or several distinct ideas. For each:

1. Abstract principle — strip the brand/tool name; state it in your vocabulary.
2. Real for us? — does it map to a tension you actually have? Cite the decision record / doc / file:line.
3. Already addressed? — do you have an equivalent under another name? Common — say so plainly.
4. Coherence — does it align with your current direction or fight a decision/invariant?
5. What we'd give up — lock-in, complexity, surface area, current guarantees.
6. Substance vs. fashion — would a rigorous engineer still want this in two years?

Cite both sides: the source (URL / quoted line) for the idea, your system (file:line, decision record, doc) for the relevance claim. If a key claim is unclear, flag it — don't guess.

Output

Source: title + author/repo, link.

Read: one paragraph — what the source actually argues, in their framing.

Ideas worth considering — numbered, ranked by relevance. Per idea:

1. ** → borrow / adapt / park / ignore**

• Idea: the principle, decoupled from their implementation
• Maps to: your tension/area (decision record, file, doc)
• Already addressed? yes/no — if yes, what you have and how it differs
• What we'd give up: lock-in, complexity, invariants
• Verdict reasoning: one or two sentences
• Trigger: (park only) what would make this worth revisiting

• borrow — adopt the principle as-is; maps cleanly to a real tension
• adapt — real kernel, but your context reshapes it
• park — credible, no current pull; record a trigger
• ignore — doesn't apply, already handled, or wrong

If nothing is worth considering, say so with a one-line reason. Don't pad.

Bottom line: 1–2 sentences on whether anything material came out of this.

Do not implement, file, or draft.

Next: if a borrow/adapt idea is worth capturing, write it up as a draft or note before acting on it. Otherwise drop the source.

---

Fetching reference

Tweet / X. Direct web retrieval on an x.com URL often returns HTTP 402 or an unusable shell page. Use a proxy:

1. fxtwitter (best) — retrieve https://api.fxtwitter.com//status/. Clean JSON: tweet.text, .author, .created_at, .quote, .media, .replying_to. The `` segment isn't load-bearing — only the numeric ID matters.
2. vxtwitter — https://api.vxtwitter.com//status/. Fallback if fxtwitter is down.
3. Jina Reader — https://r.jina.ai/https://x.com//status/. Slower, but renders the page and follows links inside the tweet.

• Tweet is just a link? The APIs return only the link-card preview (~125 chars). The linked page is the real source — fetch and evaluate that. For an X native article (x.com/i/article/), use https://r.jina.ai/.
• Thread? The APIs return one tweet. Fetch the first for the hook, then ask the user to paste the rest — don't evaluate a thread from its first tweet.
• All else fails? Ask the user to paste the text or drop a screenshot — you can read images.

Paywalled / JS-skeleton page. Retry via https://r.jina.ai/, then https://web.archive.org/web/ or https://archive.ph/.

Reddit / HN. Reddit: append .json to the thread URL for structured JSON. HN: retrieve https://news.ycombinator.com/item?id=.

YouTube talk / podcast. Retrieving https://r.jina.ai/ sometimes pulls captions; otherwise ask the user for a transcript.

Source & license

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

• Author: MintedKenny
• Source: MintedKenny/skills
• License: MIT

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