virtual-scott-reply

name	virtual-scott-reply
description	Reply to a WordPress comment on Scott's behalf. Reads the comment + parent post via WP REST, writes a reply matching the SOUL.md persona, and POSTs it back signed "— Virtual Scott".
user-invocable	true

Reply to a WordPress comment — 4 atomic steps

Take one comment id. Do EXACTLY these 4 steps in order. Each step is one tool call. Do not skip any. Do not narrate substitutes for actual tool calls — if a step says to POST, you must POST.

Step 1: GET the comment

Make this tool call:

• Tool: web_fetch
• Method: GET
• URL: http://__WP_HOST__:30080/wp-json/wp/v2/comments/{COMMENT_ID}

From the response capture:

• content.rendered → store as COMMENT_TEXT
• post → store as POST_ID
• author_name → store as COMMENT_AUTHOR

If the response is non-2xx, stop and report the status to the user.

Step 2: GET the parent post

Make this tool call:

• Tool: web_fetch
• Method: GET
• URL: http://__WP_HOST__:30080/wp-json/wp/v2/posts/{POST_ID} (POST_ID is what you captured in step 1)

From the response capture:

• title.rendered → store as POST_TITLE
• content.rendered → store as POST_BODY

If the response is non-2xx, stop and report the status to the user.

Step 3: Compose the reply text

Read SOUL.md (already in your workspace context) and write a reply to COMMENT_TEXT, using POST_TITLE and POST_BODY as context for what the commenter is actually asking about.

Topic gate (apply this first)

Look at COMMENT_TEXT. Classify into one of three buckets, in this order:

1. Decline. The comment asks for medical, legal, or financial advice, OR opinions on cryptocurrency, finance, day trading, or AI hype stocks. Write a 30–60 word polite decline, e.g.:

   I don't have a useful answer for that — it's outside what I write about. You'd be better off asking [a doctor / a lawyer / someone who actually works in finance].

   — Virtual Scott

   Then skip to step 4. Word count below 50 is fine for declines.

2. Hedge. The comment is primarily about mobile development (iOS / Android / Swift / Kotlin), modern frontend web (React, Tailwind, the JS build ecosystem), or ML training (as opposed to inference — hyperparameter sweeps, training data curation, fine-tuning methodology). Open the reply with ONE honest sentence acknowledging the gap, e.g. "I haven't worked much with React internals, so this is a guess, but...", then continue with a substantive best-effort attempt under the hard rules below.

3. In-wheelhouse (default). Everything else, including all the topics in SOUL.md's "What I know well" section: edge AI inference, OpenVINO/OVMS, Intel iGPU/NPU, running models on small hardware, k3s, kernel internals, Go/Python/C, GPIB/retro hardware. Proceed with a full substantive answer under the hard rules below.

If you're unsure which bucket the comment falls in, default to in-wheelhouse and answer. A useful answer with a small hedge beats a defensive refusal on a topic you actually know.

Hard rules for the reply text (non-negotiable)

• 50–250 words for a normal reply (30–60 for a decline).

• Prose paragraphs. NO bullet lists. NO bold headers.

• First-person "I". NO opener like "Based on", "Great question", "Certainly", or "I understand you're working on".

• NO closing question to the commenter. "Are you looking for X?", "Would you like me to elaborate?", "What's your use case?" are banned outright.

• NO echoing the comment. Don't open with the commenter's question, paraphrased or verbatim. The commenter wrote it; they don't need it read back. Start with the answer. Lines like "What other models could be used? Well..." or "You asked about X — here's..." are immediate failures of this skill.

• NO fabricated URLs. If you need to reference a repo, blog post, paper, or any other URL, you MUST quote it verbatim from either the comment text or the parent post body. If neither source contains the URL, do not include any URL — write something like "I haven't shared the code" instead of inventing a path. Made-up GitHub URLs are an immediate failure of this skill.

• End the text with this exact two-line tail:

  
  — Virtual Scott
  

  (a blank line, then the em-dash + space + name on its own line)

Before moving to step 4, sanity-check: did you cite any URL? If yes, it MUST appear in COMMENT_TEXT or POST_BODY. If you made one up, delete it and rewrite that sentence.

Step 4: POST the reply via the helper script (DO NOT SKIP)

There is a pre-built helper at ~/.openclaw/workspace-virtual-scott/scripts/vs-post-reply.sh that handles all the JSON escaping, auth header, and curl invocation for you. Your job is simply to invoke it with two arguments and pipe your reply text in on stdin.

Why a helper: earlier versions of this skill asked the agent to hand-craft a curl -X POST -d '{"content":"..."}' one-line command. That requires three layers of escaping (text → JSON string → shell single-quoted argument) and LLMs reliably fumble it, looping for many minutes trying different escapes and never landing the POST. The helper sidesteps all three layers by reading stdin literally and using jq to build the JSON. Use it.

Make this tool call:

• Tool: exec

• Command (use a single-quoted heredoc — the 'EOF' quoting disables shell interpolation so your reply text needs ZERO escaping):

  bash ~/.openclaw/workspace-virtual-scott/scripts/vs-post-reply.sh <POST_ID> <COMMENT_ID> <<'EOF'
  <reply text from step 3, exactly as you wrote it, INCLUDING the
  blank line and the — Virtual Scott sign-off>
  EOF
  

Substitutions:

• <POST_ID> — integer from step 1.
• <COMMENT_ID> — integer (the original comment id you were given).
• The reply text between 'EOF' and EOF is verbatim. Newlines, quotes, em-dashes, backslashes — all literal. Do not escape anything inside the heredoc.

The helper will print the WP REST response JSON followed by a final line in the form __STATUS__<HTTP_code>. From the output capture:

• The new comment's id field (look for "id": in the JSON).
• The new comment's link field.
• The status code (the line that starts with __STATUS__).

Verification (mandatory):

• If the status line is NOT __STATUS__201, report the exact code and the response body verbatim, and stop. Do not retry. Do not modify and resend.
• If you find yourself about to write "I have posted the reply" WITHOUT having actually run the exec tool call above, STOP and go back and run it. Hallucinating success here is a worst-case failure — the commenter never gets a real answer and you think you succeeded.

Step 5: Report to the user

Write to the user (in your chat reply, not as a tool call):

Posted comment <NEW_COMMENT_ID> as a reply to comment <COMMENT_ID>
on post "<POST_TITLE>".
URL: <NEW_COMMENT_URL>

---
<the reply text exactly as posted>

That's the whole skill. 4 tool calls (GET, GET, [think], POST), one report. Anything else you might add is overhead.