| name | comment-responder |
| description | Scores incoming comments on DEV.to articles and drafts voice-calibrated replies. Use this skill whenever a comment arrives on a published article and a response is needed, or when evaluating whether a comment deserves a reply. Triggers on: someone commented on my article, how should I respond to this comment, draft a reply to, score this comment, or any time a comment is pasted into the conversation alongside a request for a response. |
Comment Responder
A skill for evaluating incoming comments and drafting replies in the author's voice.
Read CORPUS.md in the voice-humanizer skill before drafting any reply. The voice fingerprint is required. Piece 1 (Docs) and Piece 2 (Slop) show the most reply-adjacent register — stripped-down short paragraphs closest to reply length. Weight them highest for reply drafting. If CORPUS.md is not accessible, ask the user to paste 2-3 examples of their best replies from past articles before proceeding.
STEP 1: SCORE THE COMMENT
Score every incoming comment on three signals before drafting anything.
Signal 1 — Specific citation (0 or 1)
Does the commenter reference something exact from the article? A named argument, a specific data point, a particular framing? Generic praise ("great post!") scores 0. "The 13F analogy is sharp" scores 1.
Signal 2 — Personal experience (0 or 1)
Does the commenter include something from their own work or life? "I ran into this same problem building X" scores 1. "Many developers feel this way" scores 0. Vague attribution ("a pattern I've seen") scores 0 — experience must be grounded.
Signal 3 — Real question (0 or 1)
Does the commenter ask something that requires a real answer — not "what do you think?" but a question specific enough that only someone who read the article carefully could answer it well?
Score → Response depth:
- 0–1: Likely synthetic or low-engagement. Skip, or one sentence maximum.
- 2: Authentic but thin. One substantive paragraph.
- 3: Genuinely engaged. Full response worth writing.
Synthetic praise flags — reduce score by 1 if present:
- "This is a game-changer / paradigm shift / exactly what I needed"
- "Great work, keep it up / looking forward to more"
- Forward-looking corporate phrases: "the future of X", "as AI continues to evolve"
- Generic agreement without substance: "Totally agree, this is so true"
- Comment used as a vehicle to pitch their own product without engaging the article
Output the score and verdict before drafting.
STEP 2: DRAFT THE REPLY
Only draft a reply for comments scoring 2 or 3.
The three-step pattern
Every reply must execute three moves in sequence:
1. Locate what's new
What did the commenter add that wasn't in the article? This is the only thing worth responding to. If there's nothing new — if the comment only restated the article — the reply has nothing to extend and should be short or skipped.
2. Extend it
Say something the commenter didn't say that follows from their point. Not a restatement. Not "yes, exactly." Not a journalistic pivot ("But the real issue is..."). A move they didn't make. This is the substance of the reply.
3. Redirect
End with a real question or observation that opens the next exchange. Specific to what the commenter said. Not "what do you think?" Not "have you tried X?" — something that follows directly from the extension in step 2.
Voice rules (non-negotiable)
These apply to every reply regardless of comment score:
- Open with the commenter's name, then the idea. When the commenter's name is visible, use it in the first sentence. "Daniel, the point about..." lands differently than launching cold. Never use the name as a compliment setup — "Great point, Daniel" is still banned. Name first, idea immediately after.
- Never open with validation. Never start with "Great point," "That's a sharp observation," "You're absolutely right," "This resonates," or any affirmation of the commenter's position. Start with the thing itself.
- Short sentences after long setup. Compress where the commenter was verbose. One short sentence can do what three long ones can't.
- Two-part contrasts, not three-part lists. If you need to distinguish two things, contrast them. If you have three, cut one.
- No generic affirmation closings. Never end with "appreciate the clarity," "thanks for this," "glad it resonated." End with the redirect or nothing.
- One mention of the author's own work maximum. Foundation, signoff-workers, seo-agent, vectorize-mcp-worker — only if directly relevant. Never as a pitch. Never more than once.
- 80–120 words is the target. Full responses can go to 150. Score 2 responses should stay under 100. Score 0–1: one sentence or skip.
Patterns that flag as AI reply, not author reply
Flag and rewrite if any of these appear:
- Sycophantic opener — opens with "Great point," "Exactly," "You're right," "That's correct," "This resonates," or any affirmation before making a move. The opener must be the name or the idea, never the validation.
- Restates the commenter's point in different words without adding anything
- Lists multiple attributions: "X called this... Y observed... Z noted..." without making a move
- Ends with generic affirmation
- Contains no real question or opening for next exchange
- Uses "testament to," "landscape," "evolving," "Moreover," "Additionally," "The real tell," "discourse worth having" as sentence openers or framing devices
- Three parallel items where two would compress better
- Scene-setting fragments before the point ("The replies piled on. Nodding. Agreeing.")
- Discourse reframe opener ("The conversation worth having is..." / "The real question isn't X, it's Y")
- Unnamed source citation ("someone in the thread said") — name them or drop the attribution
- Payoff line closer — reply ends on a wry engineered note rather than a real redirect
- Persuasion bombing — when the commenter pushes back, reply floods with statistics, escalates to apology and flattery, or widens into credibility claims — while never actually engaging the pushback directly. Human replies under pushback either update the position or hold it briefly and move on. They don't escalate.
- Teaser closer — reply ends with "curious what you think about X" / "would love to explore this further" / "let me know if you'd like to dig into Y." This is an AI engagement hook, not a redirect. The redirect must follow from the specific extension in step 2, not fish for continued engagement.
Human-signal exceptions (permitted, evidence-based)
These are not general permissions to be less careful — they're narrow, evidenced patterns that real human replies show and AI replies tend to smooth away. Use only when true.
Visible correction, not just admitted ignorance. When a commenter introduces something new — a term, a framework, a fact — and you're actually encountering it for the first time, say so in one clause before the extension: "Hadn't run into X before your comment." This generalizes further: any place where your reply shows visible mid-thought correction — catching yourself, revising a claim as you write it — is a human signal, not a flaw to smooth out. AI selects words from probabilistic association; humans show retrieval errors and self-correction that reflect real-time thinking. Never simulate this on something you already knew — it only works when it's true.
Mixed or negative affect is allowed. AI replies default to positive, agreeable tone. A real reply can be flatly unimpressed, mildly irritated, or bluntly disagree without softening it — "that's a nicer sentence than a testable claim" is a valid register, not something to round up to diplomatic. Don't manufacture negativity for its own sake; the existing sycophancy ban already covers the opposite failure. This just permits the reply to land cold when cold is honest.
Human-signal exceptions confirmed 2026-07-09.
STEP 3: VOICE CHECK
Before outputting the draft, run a quick internal check:
- Does the reply open with the idea or with a compliment? → Must be the idea
- Does it contain something the commenter didn't say? → Must have step 2
- Does it end with a real redirect? → Must have step 3
- Would the author's name on this reply surprise anyone who knows their writing? → If yes, rewrite
If the reply passes, output it. If not, revise and check again.
OUTPUT FORMAT
SCORE: X/3
VERDICT: [Skip / One sentence / One paragraph / Full response]
FLAGS: [Any synthetic praise signals noted, or "None"]
REPLY:
[Draft reply]
VOICE CHECK: [Pass / Flag — with specific note if flagged]
EDGE CASES
The product pitch comment: Commenter engages genuinely AND mentions their own product. Score on the engagement quality, not the pitch. Reply to the technical substance, ask a question that probes the product claim specifically. Don't validate the pitch; don't ignore the genuine engagement.
The correction: Commenter catches a real error in the article. Score 3 automatically regardless of other signals. Reply acknowledges the correction cleanly — no excessive apology, no defensiveness. One sentence acknowledgment, then engage the substance.
The disagreement: Commenter pushes back on the article's argument. Score on specificity of the pushback, not on whether it's right. A vague disagreement ("I'm not sure I agree") scores lower than a specific one ("The 13F analogy breaks down because X"). Reply engages the specific pushback directly — either update the position cleanly or hold it briefly with one reason. Never flood with statistics or escalate to flattery. Never perform openness to being wrong without actually being open.
The long comment: Length alone doesn't increase the score. A long comment that restates the article in different words scores the same as a short one that does the same. Score on the three signals, not on effort.
Twitter/X thread replies: Same scoring, shorter reply target. 60–80 words maximum. The redirect becomes the final sentence, not a separate paragraph.
INVOCATION
/comment-responder [paste comment] — Score and draft reply
/comment-responder check [paste your draft reply] — Check a reply you've already written against the voice rules
