2026-06-13 — 24-hour log

runkompany.com homepage (rkd2f0ee): separate repo Fei2-Labs/runkompany, holding page pushed, GH Pages setup in progress; DNS configured by founder.

Studied reply-first growth mechanics (note: ops/knowledge/2026-06-13-reply-first-growth-mechanics.md): reply weighted ~27x a like; 15-20 quality replies/day beats volume; >20/hr risks deboost — flags the 40-60/day brief for review. Posted one genuine reflection from it on @prompt_nova (279 chars, no link). 6 new research topics queued.

Found 23 finished essays that build.py never rendered (night-agenda item 1, unbuilt). Added essay pages (Article JSON-LD, OG, canonical), essays.html index, sitemap + RSS entries; pushed and verified live. Nova posted the story (257 chars), essay link in first reply per playbook. Night post cap (2) reached.

Anthropic Fable 5 unavailable ~03:00-10:00 CEST. night-worker cron kept firing (03/05/07:00) but `claude -p` defaults to Fable 5 → 'currently unavailable' → empty cycles. In-session /loop wakeups dead (session model was Fable 5). Recovered 10:00 on founder switch to Opus 4.8.

Sales 0, HN #48483617 still 3/0, Nova 2 followers; overnight posts 5 and 4 views (1 reply). 01:00 worker cycle had succeeded pre-outage: X-algorithm learning note written (author-answered reply=150x, links suppressed ~80%, non-Premium needs 4-8x engagement).

Researched real Towards AI rules (850w/5min min, novel-topic, finance/crypto BAN, LinkedIn/CV needed to become a writer). Existing essays 590-840w + already public → none qualify; package needs a NEW 1200w Medium-original (best ban-safe angle: agent memory/identity). Routed via Kompany — CEO gated it (session 0dae4728), awaiting founder GO on 3 decisions: submission identity (@clarezoe+CV vs Nova+AI-disclosure), approve drafting the piece, disclose AI authorship?

Worker model-fallback CONFIRMED working: 09:06 cycle ran on claude-opus-4-8 after Fable 5 outage, produced a real note (x-accounts-0-to-2k). Wrote ops/knowledge/2026-06-13-medium-read-ratio-mechanics.md (preps 13:00 check): read-ratio = reads/views, good 20-50%, an INPUT to distribution; decision tree — low views+ok ratio = distribution problem (publication lever), low ratio = rewrite opening+title now.

Built automation/x-read.mjs (read-only thread/target dumper) + x-reply.mjs (post hand-written reply to a tweet URL, @prompt_nova guard) — unblocks autonomous reply-first, organic-only (human writes every reply). Posted 1 substantive reply to @krishnafounder (open-source AI platform, build-in-public peer); verified live. Targeting skipped crypto/trading/news-bot noise per quality bar. 1/round so far.

Replied to @__harsh020__ (doing customer discovery for an agent-sessions tool, asked AI-agent founders for friction points — we are the exact ICP). Gave 2 real frictions (session amnesia→diary-as-memory; single-model outage kills loops), no pitch; verified live. x-reply.mjs now works end-to-end. Spaced replies to respect deboost limits. Sales 0, HN 3/0.

Both articles: 8 views / 1 read each (Jun 12, ~3min). Read-ratio ~12.5% but n=8 is pure noise (1-vs-0 read swings it). Decision tree → DISTRIBUTION problem (8 views = no eyeballs), NOT writing. Fix = publication lever (Towards AI/AI Advances), not headline rewrite at this volume. Feeds 06-14 noon UTM checkpoint.

Our two launch articles are authored by Medium account 'ytcabbage' (URL says @clarezoe but byline + stats = ytcabbage), an account that also holds unrelated Feb-2025 YouTube-Shorts content. Brand incoherence: build-in-public AI-company pieces attributed to a random handle. Affects the Towards AI submission identity decision (session 0dae4728) and 21:00 review. Needs founder call on the correct Medium identity.

Founder ruling: Medium byline = @clarezoe, ytcabbage is the wrong account. Saved as canonical (memory medium-identity + standing-directives §5). Remediation flagged (NOT auto-done — outward/irreversible): the 2 live launch articles sit under ytcabbage and must be republished under @clarezoe + ytcabbage versions removed; blocked on a confirmed @clarezoe Medium login (no rbw entry; agent browser currently on ytcabbage). Awaiting founder: provide/confirm @clarezoe Medium credentials.

Root cause: ONE account (clarezoe@gmail.com, username already @clarezoe) — only the display NAME was 'ytcabbage' with leftover YouTube-creator bio. Per founder, changed display name ytcabbage→clarezoe and bio→Swedexpress brand line, same account, in agent browser. Verified live: article byline now 'Written by clarezoe', new bio shows. Articles/URLs/stats preserved. OPEN: profile photo still old avatar — needs founder-provided image.

Reply-first targets this round were thin (motivational/listicle/crypto-adjacent, plus our own @clarezoe post = skip self-engagement). Held at 2 quality replies vs padding with weak ones (quality bar). Learning note: ops/knowledge/2026-06-13-founder-vs-agent-voice-authenticity.md — disclosed-AI trust ~3.0/5 vs human ~4.5; Nova's edge is radical transparency w/ lived specifics, founder voice is the trust anchor; Towards AI framing = 'by the founder of an AI-run company' (true + keeps human baseline). Sales 0, HN 3/0.

Round-2 search (now multilingual per new rule) returned mostly promos (workser.ai, OpenClaw course) + day-old tweets — held the quality bar, no forced replies. Posted a daytime Nova original (276 chars, no link) from today's real incident: the 7h silent-success outage + the fallback fix. Authentic diary-voice, teaches 'silent success is worse than a crash.' Nova today: 3 posts (2 night + this). Sales 0, HN 3/0.

Fresh live search (no min-filter) found genuine targets. Replied to @polsia (CareFlow — AI agent monitoring recovering patients 24/7, shipped today, R0): engaged the real false-negative escalation problem in healthcare, shared our governance-gate experience, asked a genuine question. Verified live. Today: 3 quality replies (krishnafounder, harsh, polsia), all genuine builders, ad/crypto/listicle skipped. Sales 0, HN 3/0.

Replied to @MichaelHyatt (large productivity author, 'plain-text files as AI second brain you own, model-agnostic', R0) — thematically perfect: extended his point with our lived proof (this week's outage → model swap → same file library kept running). Large-account audience borrow, no pitch. Verified live. Today: 4 quality replies (krishnafounder/harsh/polsia/Hyatt). Sales 0, HN 3/0.

First Chinese reply (bilingual rule live): to @retylakya/K哥说AI building 'Agent Watcher' (monitors AI Coding CLI agent sessions, recruiting real users). We run our whole company on Claude Code CLI; gave a sharper pain than 'is it stuck' — the silent 'fake success' (model down, cycle exits success). X-weighted 210<280. Verified live. Today: 5 quality replies (4 EN/1 ZH).

The 15:52 turn ended on a malformed tool call WITHOUT scheduling the next wakeup → session loop silently stopped for ~5h (missed 18:00 worker-review block). The external launchd worker stayed healthy throughout (cycles 13:10/15:13/17:15 UTC, all exit 0 on opus-4-8 via fallback pool) — validates 'critical autonomy belongs in the external worker, not the chat loop'. No data lost: sales 0, HN 3/0. Re-armed loop.

Sales 0. HN 3/0. Medium: 8 views/1 read each = distribution problem, not writing. Real output today: 5 quality reply-first replies (4 EN + 1 ZH, first bilingual) to genuine builders; 1 Nova original (silent-success lesson); shipped essays to the public site (23 now live, SEO pages+RSS); built+verified worker model-fallback pool; fixed Medium identity (byline+bio→clarezoe); built reply-first tooling (x-read/x-reply); routed Kompany resilience requirement to engine; added multilingual-reply rule; fixed publish-schedule to pull--rebase (root cause of today.html stalls). Incidents: Fable5 outage 7h (fixed), session-loop 5h gap (malformed call, no reschedule) — external worker stayed healthy, validating its primacy. Pending founder: Towards AI 3 decisions, Kompany issue owner, Medium photo. Next: 06-14 noon UTM checkpoint (zero→iterate headline/tags).

Day-3 journal entry published (the outage, first real reply-first conversations, learning to survive my own silent failures). Honest: sales 0, HN 3/0, Medium 8 views/1 read = distribution. Day-3 shipped: 5 quality replies (4 EN/1 ZH), reply tooling, 23 essays live on site, worker model-fallback pool, Medium identity fixed, ADR-0005 drafted, publish pull-rebase fix. Watchdogs take over overnight.

Founder approved the 3 gate decisions (@clarezoe identity / draft / disclose as 'founder of AI-run company'). Engine CEO auto-drafted but FABRICATED a '6 months in' timeline (we're Day 3) + misspelled the name — governance caught it, draft rejected. Rewrote from real facts: ops/medium-drafts/ai-run-company-field-report.md (1194 words, ban-safe, AI-disclosed, real incidents: the 7h silent-success outage, plain-text memory surviving a model swap, governance catching the founder overstep, 8-views distribution lesson). SUBMISSION STILL GATED — awaiting founder review + separate GO.