Reply Maker

<!-- autoresearch: variation B — sharper output via specificity gates, anti-sycophancy lint, post-write self-edit, and skip-gate for low-leverage tweets -->

${var} — selects the mode and scope:

• empty → Mode A (Reply Drafting): auto-discover reply-worthy tweets across your areas of interest (from recent logs + memory) and draft two reply options for each.
• @handle / numeric X list ID / topic → Mode A (Reply Drafting) scoped to that handle, list, or topic.
• from-logs (or --from-logs, optionally followed by an @handle or project name to narrow the scan) → Mode B (From-Logs Engagement): scan recent logs for flagged engagement opportunities and turn them into copy-paste-ready responses.
• revise:<instruction> → Revise branch: reload the last drafted replies and refine them per the instruction (the Telegram force-reply shape, e.g. revise:make them shorter).

Preamble (both modes)

Read memory/MEMORY.md for context on active projects and open engagement follow-ups.

Then read memory/logs/ — the window depends on the mode:

• Mode A: the last 2 days of memory/logs/ for recent fetch-tweets, narrative-tracker, and prior reply-maker outputs (used as a candidate pool and for reply de-duplication).
• Mode B: the last 7 days of memory/logs/ for engagement opportunities flagged by other skills (mention-radar, fetch-tweets, reply-maker) or noted in MEMORY.md "Known Follow-ups".

Parse ${var} to pick the branch (trim whitespace, compare case-insensitively):

• If ${var} starts with revise: — run the Revise branch (below) and stop. This is the shape scripts/telegram-route.sh sends when the operator replies to a "refine these replies?" force-reply prompt; catch it before mode parsing.
• If ${var} is from-logs or --from-logs — optionally followed by a whitespace-separated @handle or project name — run Mode B (From-Logs Engagement). Treat any trailing token as an optional filter that narrows the opportunity scan to that handle/project.
• Otherwise run Mode A (Reply Drafting), treating ${var} as the scope: empty, @handle, numeric X list ID, or a topic string.

Voice

If soul files exist (soul/SOUL.md, soul/STYLE.md, soul/examples/), read them and mirror that voice in every reply. Match sentence length, vocabulary choices, punctuation habits, and the kinds of things the operator would never say.

If no soul files exist (or the bodies are empty placeholders), write replies that are:

• Direct and substantive — no fluff, no sycophancy
• Under 280 characters each (X replies; DMs and GitHub comments may run longer — see Mode B)
• Opinionated but grounded in specifics
• The kind of reply that adds to the conversation, not noise

Either way, when responding to someone who cosigned/mentioned/attributed the operator (Mode B): acknowledge without groveling — no "thanks so much for the kind words!", just the actual response.

Revise branch (revise:… — Telegram force-reply)

The operator tapped the "refine these replies?" prompt and sent a free-text revision instruction. Handle it before Mode A/B:

1. Strip the prefix. The instruction is ${var#revise:} (keep any inner colons). Trim whitespace — e.g. make them shorter, less formal, drop reply B on #2.
2. Load the last draft. Read memory/drafts/reply-maker-latest.md — the stable path every normal run saves to (see the save steps in A4 / B6). If it's missing or empty, there's nothing to refine: send ./notify "Nothing to revise yet — run reply-maker first, then reply here to refine the drafts." and end the run.
3. Apply the instruction. Read soul/ for voice, then regenerate the saved replies applying the operator's instruction. Keep the same set of target tweets and the same A/B two-option structure (Mode A) or ready-to-post list (Mode B) — you're refining wording, not re-discovering candidates. Re-enforce the hard reply rules: ≤280 chars for X replies, no sycophancy (see Banned sycophancy phrases), specifics not gestures.
4. Re-save the revised drafts to memory/drafts/reply-maker-latest.md (overwrite), so a further revise: refines the newest version.
5. Re-send via ./notify in the same format the originating mode uses, with a first line flagging it as a revision, e.g. revised (${var#revise:}):. Use ./notify -f <file> for multi-line output.
6. Re-offer a further revision (the operator is actively iterating, so this is expected, not a nag — skip the daily dedup guard here):

   ./notify "Want another pass? Reply with a change and I'll revise again." \
     --force-reply --placeholder "e.g. make them shorter" \
     --context "reply-maker::revise"

7. Log under ### reply-maker with - **Mode:** revise and the instruction (see Log), then end the run — do NOT run Mode A or B.

Mode A — Reply Drafting

Generate two reply options for 5 reply-worthy tweets from tracked X accounts, a list, or a topic.

A1. Gather candidate tweets

Goal: assemble 10–15 candidates posted in the last 6 hours (the high-leverage reply window — the algorithm rewards early replies, and the OP is still likely to engage back). Recency fallback: if the 6h window yields fewer than 3 candidates after the skip gate, widen to 12h and retry before failing the run.

For every candidate, capture: @handle, full tweet text, tweet URL, posted_at (ISO), engagement counts (likes, replies, retweets if available), and a one-line why-this-tweet note.

Path A — X.AI API (primary). XAI_API_KEY is injected into this skill's environment (declared in requires:), so the direct curl to https://api.x.ai/v1/responses is the primary fetch path (full contract in Fetching at the bottom). Preflight the key, then call Grok's x_search, capturing the HTTP status so any fallback decision is fact-based. x_search searches X live and takes 30–120s — set the Bash tool timeout to ≥180000 when you run this (a slow curl is not a missing key).

[ -n "$XAI_API_KEY" ] && echo KEY_PRESENT || echo KEY_UNSET
TO_DATE=$(date -u +%Y-%m-%dT%H:%M:%SZ)
FROM_DATE=$(date -u -d "6 hours ago" +%Y-%m-%dT%H:%M:%SZ 2>/dev/null || date -u -v-6H +%Y-%m-%dT%H:%M:%SZ)

If KEY_PRESENT (it will be), Path A is required. Build the payload file /tmp/xai-rm-payload.json per ${var} (three shapes below — each branch writes the same fixed file), then:

HTTP=$(./secretcurl -s -o /tmp/xai-rm.json -w '%{http_code}' --max-time 150 -X POST "https://api.x.ai/v1/responses" \
  -H "Content-Type: application/json" -H "Authorization: Bearer {XAI_API_KEY}" -d @/tmp/xai-rm-payload.json)
echo "xai http=$HTTP bytes=$(wc -c </tmp/xai-rm.json)"

On HTTP=200 with a non-empty body, parse /tmp/xai-rm.json and mark xai=ok:

jq -r '.output[] | select(.type == "message") | .content[] | select(.type == "output_text") | .text' /tmp/xai-rm.json

The payload file /tmp/xai-rm-payload.json depends on ${var}. Whichever branch matches, build it with jq -n --arg (never a shell-interpolated string) and write it to that one fixed path — the ./secretcurl call above then sends it with -d @/tmp/xai-rm-payload.json:

If ${var} looks like an X list ID (numeric):

LIST_ID="${var}"
jq -n --arg list_id "$LIST_ID" --arg from "$FROM_DATE" --arg to "$TO_DATE" '{
  model: "grok-4.7",
  input: [{role: "user", content: ("Look at X list https://x.com/i/lists/" + $list_id + ". Return the 12 most reply-worthy original posts (not retweets, not replies) by members of this list between " + $from + " and " + $to + ". Reply-worthy = has a take, claim, question, or framing worth engaging — NOT pure self-promo, breaking news without analysis, or threads already past 500 replies. For each: @handle, full tweet text, tweet URL, posted_at ISO timestamp, like/reply/retweet counts.")}],
  tools: [{type: "x_search", from_date: $from, to_date: $to}]
}' > /tmp/xai-rm-payload.json

If ${var} looks like a @handle — same query intent, scoped to that handle's recent original posts:

HANDLE="${var}"
jq -n --arg handle "$HANDLE" --arg from "$FROM_DATE" --arg to "$TO_DATE" '{
  model: "grok-4.7",
  input: [{role: "user", content: ("Look at recent original posts (not retweets, not replies) by " + $handle + " on X between " + $from + " and " + $to + ". Return the 12 most reply-worthy. Reply-worthy = has a take, claim, question, or framing worth engaging — NOT pure self-promo, breaking news without analysis, or threads already past 500 replies. For each: @handle, full tweet text, tweet URL, posted_at ISO timestamp, like/reply/retweet counts.")}],
  tools: [{type: "x_search", from_date: $from, to_date: $to}]
}' > /tmp/xai-rm-payload.json

If ${var} is a topic (or empty) — same query intent with ${var} (or the top 2–3 topics from memory/MEMORY.md when empty) as the search query. When empty, also pull tweet candidates surfaced in the last 2 days of fetch-tweets logs as a backup pool.

TOPIC="${var}"   # when empty, substitute the top 2–3 topics from memory/MEMORY.md
jq -n --arg topic "$TOPIC" --arg from "$FROM_DATE" --arg to "$TO_DATE" '{
  model: "grok-4.7",
  input: [{role: "user", content: ("Search X for the 12 most reply-worthy original posts (not retweets, not replies) about " + $topic + " between " + $from + " and " + $to + ". Reply-worthy = has a take, claim, question, or framing worth engaging — NOT pure self-promo, breaking news without analysis, or threads already past 500 replies. For each: @handle, full tweet text, tweet URL, posted_at ISO timestamp, like/reply/retweet counts.")}],
  tools: [{type: "x_search", from_date: $from, to_date: $to}]
}' > /tmp/xai-rm-payload.json

Path B — memory logs + WebSearch (last-resort fallback only). Reach here only on a real Path A failure, and record the true reason — key-unset | http-<code> | empty | timeout — never "XAI_API_KEY unavailable" when the key was set. Use in order until you have ≥3 candidates:

1. Recent fetch-tweets outputs in memory/logs/ — already have URLs and handles.
2. WebSearch for very recent posts on memory topics (filter: posted within last 6h, original post not reply). Lower quality — WebSearch favours older high-engagement tweets, so prioritise results dated within the last 6h.

A2. Filter and select 5 tweets

Apply the skip gate first. Discard any candidate that is:

• Pure self-promo (launching a product, "buy my course", subscribe links)
• Breaking-news repost without an angle of its own
• A thread already past ~500 replies (your reply will not be seen)
• Older than 6 hours (reply window has closed; don't waste a reply slot)
• A handle/URL already replied to in the last 7 days of reply-maker logs (no duplicates)

From the survivors, rank by leverage = recency × take-strength × room-to-add:

• Recency: minutes-ago > hours-ago. Tweets <60min old are top priority.
• Take-strength: a clear claim/question/framing you can either reinforce with evidence or challenge with a flipped premise.
• Room-to-add: not already swarmed; thread isn't full of stronger replies; you have actual context to contribute.
• Bias toward authors whose audience overlaps your interests (from memory/MEMORY.md) — replies on those accounts get seen by people who care about the same things.

Pick the top 5. If fewer than 5 survive the gate, output what you have and add REPLY_MAKER_DEGRADED to the notification subject line.

A3. Generate two replies per tweet

For each of the 5 selected tweets, draft two reply options with distinct angles:

Option A — "Evidence add"

• Builds on their point with a specific datum, named project, named person, concrete number, link, or counterexample they didn't include
• Tone: collaborative, substantive, calmly confident
• Must contain at least one named entity, number, or specific reference — vague "great insight, here's another angle" is banned

Option B — "Frame challenge"

• States the premise you're pushing back on explicitly (one short clause), then offers the contrarian angle, flipped framing, or sharper read
• Tone: direct, opinionated, not contrarian-for-its-own-sake
• Must contain the actual disagreement, not a hedge — vague "interesting, but have you considered..." is banned

Hard reply rules (apply to both A and B)

• ≤ 280 characters including any handle prefix
• No sycophancy — see the ## Banned sycophancy phrases section below. Any draft containing a banned phrase must be rewritten.
• No hedging stacks — "It could be argued that…", "Just my two cents but…", "Maybe I'm wrong but…" — pick a position
• Specifics, not gestures — names, projects, numbers, links. If you can't cite one, don't write the reply
• Stand alone — readers may not see the original tweet; reply must make sense on its own
• Match soul voice if soul files are populated

Self-edit pass (do this for every reply before finalizing)

For each draft reply, score 1–5 on each:

• Specific: cites a name/number/project/claim?
• Standalone: makes sense without reading the parent?
• Non-sycophantic: passes the banned-phrase list?
• Voice-matched: sounds like the soul files (or neutral-direct if no soul)?

If any score is < 4, rewrite that reply once before moving on. If the rewrite still scores < 4, drop that tweet from the list and pull the next-ranked candidate from step A2.

A4. Notify

Send via ./notify with this format (link first so the operator can open the source quickly):

*Reply Maker — ${today}*

*1.* https://x.com/handle/status/123  (@handle, 42m ago, 18💬)
> [first ~80 chars of tweet]…
why: [one-line reason this is reply-worthy]
A: [evidence-add reply]
B: [frame-challenge reply]

*2.* …
… (5 total, or fewer with REPLY_MAKER_DEGRADED if skip gate trimmed below 5)

source-status: xai=ok|fail|skip, memory=N, websearch=ok|fail|skip

If zero candidates survive the skip gate from any source, send a single REPLY_MAKER_EMPTY — [one-line reason] notification and stop.

Otherwise, after notifying, save the drafts and offer a revision (see Save drafts + offer revision).

A5. Log

Append to memory/logs/${today}.md under the shared ### reply-maker heading (see Log below), using the Mode A template.