The Reply Desk · A field report for operatorsFiled: May 23, 2026

Reply to the right post.
At the right time.
In your voice.

ViralReply finds the conversations worth entering before they peak, drafts replies that sound like you, and keeps the final post where it belongs — in your hands.

No auto-postingNo cookie uploadWorkspace scopedHuman in control
The Numbers · I

A reply cockpit, not a one-click bot. Built around the figures operators actually watch.

Post → alert latency5minFrom official monitoring to your inbox.Monitoring SLO · 28-day
Approval rate38%Drafts accepted with only minor tone edits.Approval log · v3 voice
Voice match88%Drafts held inside your voice cluster.Hold-rate · per-platform
Readiness gates35Production checks before anything ships.Release manifest · main

† Methodology in the production runbook · figures refreshed weekly

The Voice · II

It sounds like you. Because it learned from you.

A questionnaire, a sample reader, and a rule engine. Re-run any time the brand shifts.

A voice you can version, test, and ban from bad habits.

Connect X, LinkedIn, blogs, podcasts, newsletters — or paste raw samples directly. ViralReply builds a Voice Profile that captures cadence, vocabulary, structural habits, and the AI-tell phrases you never want in your replies. Every profile is versioned. Re-run the questionnaire after a brand shift, A/B two voices side by side, or freeze one for a specific identity.

  1. Questionnaire pulls cadence, vocabulary, and stance from real samples
  2. AI-tell filters ban em dashes, hedging openers, list-of-three crutches
  3. Punctuation rules respected per-platform, per-identity
  4. Playground drafts against any post before you change anything
Founder voice · v3Active
Avoid em dashesEnforced
Avoid "in essence"Enforced
Avoid list-of-three crutchEnforced
Conversational X registerEnforced
Allow first-person mentionsOff

Sample size1,284
AI-tell filters11
Hold rate88%
The List · III

Curated targets.
Not endless feeds.

Enter a handle. ViralReply reads your bio, posts, product and audience, and recommends the accounts you should reply to — with the specific reason each one belongs on your list.

ICP cluster · founders & operators · 12 of 184 candidates
01
@karpathyResearcher · OSS maintainer

Adjacent technical audience. Threads compound across the AI-infra community.

92ICP score
02
@swyxDevTools founder

Curates AI-infra conversations early. Reply windows hold long.

88ICP score
03
@jeremyphowardResearcher

Drives high-comment threads on training topics. Strong technical credibility.

84ICP score
04
@vvoovvInvestor

Posts that the right operators reply to. Engaged with founders' tools.

79ICP score
05
@dharmeshFounder · platform CTO

Long-form posts on growth that operators read end-to-end.

76ICP score
06
@sahilbloomWriter

Frame-establishing threads. Strong follower-of-follower overlap.

71ICP score
The Window · IV

Five minutes between post and alert.
Three drafts between alert and reply.

Every monitored post is scored by author fit, timing, relevance, safety, and visibility. You get an alert while the reply window is still valuable — and three reply variants written in your voice.

Opportunity · 027F4Window · early
@karpathy · 4 min ago

distribution is the only loop that compounds. capability is table stakes now

Variant A · witty · 174 chars

capability gets you in the room. distribution decides whether anyone hears you when you talk. nobody admits this until they're shipping into the void

Variant B · helpful · 218 chars

the gap between “works on benchmarks” and “twitter knows it exists” is where most of the EV is rn. only ones doing it well imo are anthropic and cursor. everyone else is in stealth on linkedin


Author fit92
Est. visibility21k
SafetyClear

A scored window. Not a notification dump.

ViralReply tags every reply window early, useful, or late. You stop chasing posts that already crested, and you stop missing the ones that haven't. Three drafts per opportunity, platform-tuned for X cadence. A safety pre-check on every draft — policy, product, persona.

  1. Three drafts per opportunity, platform-tuned
  2. Safety pre-check on every draft — policy, product, persona
  3. Approval-likely cluster surfaces what your audience accepts
  4. Email, Slack, side-panel — dedup and quiet hours built in
The Composer · V

Wherever you're already drafting.

A Manifest V3 extension that adds a VR Draft button next to the X composer. It reads only the visible thread, drafts a reply in your voice, and pastes it in. You always hit Post.

Native to the composer. Invisible to the platform.

The extension is held to the same safety posture as the platform. No cookies, no passwords, no auto-clicking the final button. It captures only what you can see, and only when you ask. Tokens are scoped to your workspace, with no shared keys.

  1. VR Draft button injected into the reply composer
  2. Side panel for live alerts, captures, and draft history
  3. Token-scoped to your workspace — no shared keys
  4. Works in x.com and pro.x.com · LinkedIn ingestion in v0.2
Replying to @karpathy

capability gets you in the room. distribution decides whether anyone hears you when you talk

VR Draft204 / 280Reply

The chrome stays exactly where you expect it. Only the draft changes — written in your voice, never sent without you.

Six Rules · VI

Built for trust.
Not built around it.

Account safety is a product feature, not a clause. It shows up in the architecture, the extension, the API, and every default. Six rules you can read in under a minute.

01

No auto-posting

Every draft requires your manual confirmation on the platform itself. The extension never clicks the final button. The API will not post on your behalf without a flag your operator has explicitly enabled.

02

No cookie upload

ViralReply does not read, copy, or upload platform cookies, sessions, or local auth state. The extension cannot — the manifest does not request cookie permissions.

03

No platform passwords

Your X password stays with X. Connections use official OAuth scopes. We do not impersonate, scrape behind login, or store credentials we should not.

04

Workspace scoped

Every session is bound to a workspace and a role. Data never crosses workspaces. The API enforces tenancy at the route, the service, and the database row.

05

Audited by default

Every draft, edit, approval, post, and admin action lands in an immutable audit log keyed to the operator who performed it. High-risk admin actions require a typed audit reason.

06

Emergency disable

One form in the admin console halts publishing, connectors, LLMs, and alerts in a single workspace-scoped action. The kill switch is auditable, reversible, and rehearsed in the runbook.

The Receipts · VII

Measure what actually moved.

Edit distance, approval rate, reply-window timing, target quality, voice match, and per-reply outcome. Every event flows back into target ranking, voice tuning, and prompt shaping.

Reply window · last 28 daysCohort · founder voice v3
Immediate51%
Early47%
Useful32%
Late11%

Edit distance0.18
Voice match88%
Variant keptwitty

A learning loop. Not a vanity dashboard.

Drafts that earn edits become signal, not noise. Targets that stop earning replies leave the list. The reporting layer keeps receipts on every decision the system makes, and every decision you override — so the model gets better at being you, not better at being itself.

  1. Edit distance from draft to posted reply, per variant
  2. Approval rate per voice version, per platform
  3. Target quality trend per cluster, with provenance
  4. Best-variant lineage so you can see why a style wins
Backstage · VIII

Run it like a real platform.

An internal operations console for the people keeping the system honest — queue health, connector and provider state, feature flags, cost monitoring, incident management, and a one-form emergency disable.

Production overviewInternal · audited
Queues reporting6
Failed or dead-letter jobs0
Providers degraded0
Open incidents0

P95 reply latency1.8 s
LLM cost · 24h$4.82
Tokens · 24h10.4M

Operational by default. Not bolted on later.

Thirty-five readiness gates run before a build is allowed to call itself production. Kubernetes manifests, migration jobs, pod disruption budgets, network policies, observability rules, and an incident response runbook all ship in the same repo as the code.

  1. Job queue: waiting, active, failed, dead-letter — per logical queue
  2. Model providers: per-task routing, P50 / P95 latency, token usage, cost-24h
  3. Feature flags: kill switches that require a typed audit reason
  4. Incidents: severity tracking, commander, channel, runbook references
Inside This Issue · IX

Nine things worth keeping.

ViralReply is not a single product. It is a small set of products wired together carefully — and a long list of decisions about what not to build.

01

Voice questionnaire

Capture cadence, banned phrases, and AI-tell signals in fifteen minutes. Re-run any time the brand shifts.

p. 04
02

ICP discovery

Generated from your real bio, posts, product, and audience — explained in plain language, not embeddings.

p. 08
03

Live monitoring

Official X API plus extension capture plus internal connectors. Reply windows tagged early, useful, late.

p. 12
04

Reply scoring

Author fit × timing × relevance × safety × estimated visibility. Three drafts per opportunity, voice-matched.

p. 14
05

Multi-provider LLM

Route across OpenAI, Anthropic, Azure AI Foundry, Xiaomi MiMo. Per-task routing, fallback recovery, cost tracking.

p. 18
06

Workspace isolation

Multi-workspace from day one. Data never crosses tenants. Tenancy enforced at route, service, and row.

p. 20
07

Audit log

Immutable record of every draft, edit, approval, post, and admin action. High-risk operations require a typed reason.

p. 22
08

Alert delivery

Email, Slack, and side-panel. Quiet hours and idempotent dedup so you never get the same alert twice.

p. 24
09

Production-grade

Kubernetes, pgvector, BullMQ, pod disruption budgets, observability rules, runbooks, 35 readiness gates.

p. 28
Apply for access · X

Start replying with intent.

ViralReply is invite-only for operators, founders, and teams who treat replies as distribution. Tell us about your handle and we'll be in touch — usually within one business day.