AI track · 30 entries

Set the rule that LLM tool sequences must be enforced in code, not prompt prose. Apparently the model cannot pinky-swear infrastructure into existence. signal_archive became the canonical HPFHQ Supabase corpus for Slack, git, decisions, memory, email, and meetings. Anthropic credits ran out during SuperNova work, so daily API cost monitoring moved from nice-to-have to required plumbing. Ultrawinning Claude onboarding got promoted into an ecosystem: /claude-onboarding, starter zip, /start skill, and follow-on pages. COMSAT learned to paint VS Code chrome from workspace metadata. Small polish, real orientation value.

Shipped /journal in the cockpit plus journal_entries and /api/journal in UltrawinningSocial. Built the three-track log: AI, Social Build, HPF. The system can now remember the work instead of relying on chat archaeology. Overmind upgrade review processed 23 harvested decisions and banked 4 reusable patterns. COMSAT exposed the usual local-dev tax: ELECTRON_RUN_AS_NODE leaking from VS Code breaks Electron startup. Execution estimate got corrected from 12 hours to 20 minutes. Risk theater noted.

Shipped the Ultrawinning cockpit spine: Rajdhani/pink-on-black tokens, queue edit/approve/kill, triage board, comments inbox. UltrawinningSocial got the matching APIs: triage_status gating, /api/triage, /api/comments, draft PATCH proxy. Claude Opus 4.7 did real pair-programmer work, not decorative autocomplete. Infrastructure lesson: notifications need repo, file path, and skill/prompt, or they are just expensive vibes. Also added Overmind agent budget caps. Silent token burn is not strategy.

Shipped the /wrap router skill for session handoff, EOD, build-complete, fix-complete, and canary routing. Built the Nova tag system: inline WATCH, REMIND, and SWEEP markers feeding daily sweep, Telegram, local upkeep files, memory, and calendar. Hardened drone with RTF handling, lenient JSON parsing, and Telegram error pings because Haiku sometimes adds prose after JSON like it pays rent. Scaffolded the Ultrawinning cockpit and social engine, then added capability probes to stop external-system assumptions from silently rotting.

Moved SuperNova browser automation from Playwright + Stagehand to Vercel Labs agent-browser CLI. Added the business/ knowledge domain contract for Archon, enforced by manifest.yaml and a markdown/yaml-only pre-commit hook. Captured comsat-radar v2 state, including SessionWatcher cache as the Anthropic utilization source. The useful product lesson was blunt: user-facing tools should mirror the operator’s mental model, not the engineer’s folder habits.

Defined Hawkeye as the interactive architecture explorer: click nodes, trace upstream/downstream flow, inspect orphans. Created the ultrawinning-new-tool skill and then tightened its icon system after auditing ten existing UW system icons. Built the iPhone Shortcut to iCloud to Haiku extraction pipeline for Overmind drone capture. Decision: visual previews go to GitHub Pages, not local preview purgatory.

Wrote the Adjutant upgrade plan: sales loop, compass loop, QuickBooks cron, NPS, and UI consolidation. Finished the Hive to Overmind rename and expanded pollution-guard exemptions for legitimate dated case studies. Captured agent safety patterns from SuperNova and COMSAT: kill switch, dedup, circuit breaker, outbound API safeguards. Also learned that whole-file Write gets reverted by hooks. Targeted edits survive. Cute.

Ran test applications through the Slack qualification path: qualified, browsing, and full-DQ cases all surfaced in #hpfhq-test. The pipeline now shows green/red application states, source, intent, and contact fields in operator-readable form. Found the recurring UX failure mode: external systems need clickable URLs, not breadcrumb cosplay. Small infra day, but the lead-routing display got less theoretical.

Codified icon-search-policy: Lucide, then Heroicons, then Iconify, then drawn SVG. Paid AI logo generators stay outside the loop. SuperNova Home Assistant integration moved to real infrastructure: HA container running, TPI plugs, Daikin AC, Dyson fan queued for pairing. Locked the sales definition: close means money received, with Fanbasis before Stripe before Slack. Also learned the Meta portal rule again: document the screen in front of you, not the one you wish existed.

Ordered the Mac Studio M4 Max as the next executor host. The M1 16GB Mac Mini has done its tour of duty. Set the local AI plan: Gemma 3 4B for intent routing, Kokoro-82M for voice output after the hardware lands. Built the native Meta Lead Ads webhook for “Leads - 2026” and fixed the income qualification gap. Useful lesson: three-tier money filters beat boolean filters when the ad platform has its own bracket theology.

Created overmind as the dedicated Hive repo for AI infrastructure R&D and personal IP. Shipped day-one scaffolding, ADR, pollution guard, automated sweep, and the 1Password CLI migration runbook. Locked secrets into a dedicated HIVE vault, fixed vault casing, and promoted Tier 2 Claude rules from HPFHQ sweep output. SuperNova V2 shipped with calendar attendees, contact resolution, Stagehand browser, Docker, and n8n. The runaway loop produced the correct scars: kill switch, dedup, circuit breaker, and pipeline-first builds before more autonomous agent features.

Turned the external-state near-miss into Verification Receipt Architecture. Live systems now need live-read receipts before recommendations: GHL, UTM links, Typeform, and anything else with reality attached. QuickBooks integration was built through OAuth, P&L sync, bank balances, and Finance tab; blocked only on Dan approving OAuth. Also tightened session naming and terminal instructions because multi-machine ops punishes ambiguity.

Added the attribution pipeline trace rule. Any UTM, Typeform field, extraction code, hidden field, or Slack message change now requires tracing the whole data path both directions before shipping. This is the correct amount of paranoia for revenue attribution systems.

Set up the dual-machine dev environment: MacBook Pro plus Mac Mini M1 over Tailscale. Repo sync, SSH, backups, and workflow rules got formalized instead of living as vibes. Started /cc, a fresh single-page Command Center replacement for prepare, stage, workspace, and EOD. Attribution work stayed live, with May’s 100% target blocked by setter Typeform bypass and ManyChat channel blindness.

Claude got a standing order: read knowledge/integrations/{system}.md before poking at integrations. Apparently the fastest debugger is still the documentation you already wrote. Small rule, high leverage: stop re-learning known system behavior from scratch.

Tech overhaul split into three stages: audit, replace software, then harden reliability, backups, security, and dev hiring triggers. Approved Trainerize -> HPFHQ fulfillment metrics plan across six phases. Pipeline architecture v3 landed with nine stages, ownership, cancel flow, and Pipeline+EOD merge concept. New operating rules added: verify audits, verify attribution end-to-end, ask third-party LLM after two failed fixes.

Co-op Command Center became concrete: command map UI, coordination service, terminal management. PTY spike worked and parser shipped; next step is the Express+ws service. Unified Sales Pipeline reached phases 0-4, with 22/28 items shipped and flag flips held for testing. Sales Call Scorer Agent got specified, but Sales P1 stayed pre-call sequence. Correct priority, annoyingly practical.

Git hangs got a real diagnostic protocol: isolate object store, index, and working tree before touching the index. Root cause logged: stale .claude/worktrees accumulate until git status becomes a hostage situation. ManyChat SOB cleanup shipped: SOB :: DASHBOARD killed, booking links fixed with UTMs, field mapping completed. Added agent rules from the wreckage: no shadcn init in parallel sessions, hooks before early returns, grep is not reading.

Started the architecture schematic project: data flows, department interactions, component audit, memory layer audit. Decided the metric hierarchy: base, derived, view. Tier is what it is; source_function is how it is fetched. Captured build discipline after pain: Node 25 hangs Next.js 14.2, use Node 20, one build at a time. Finished GHL audit session 10: workflows organized, damage map written, identity gap closed.

Locked the marketing tracking canon: Meta pixel, CAPI, Typeform, booking flow. True ROAS stayed intentionally narrow: Meta ad -> HPFHQ home -> Typeform -> booked call. ManyChat, newsletter, and ghl_export are organic by rule. Not every missing dollar is a bug. Purged stale GHL test contacts; scripts/test-ghl-webhook.ts now points at a dead contact_id.

Fulfillment parser still had open bugs: empty-month zeros, 2025 header mismatch, missing snapshot response key. GHL snapshot attribution changed: likely CC360 agency-level, not SOB. Useful, late. Adopted GHL workflow/pipeline/custom-field naming convention v1 in knowledge/integrations/ghl.md. Added verification scars: silent CLI flags can fake clean greps; remembered SQL columns are not schema.

Completed HPF Command Center knowledge architecture consolidation. The repo now has a more durable operating memory instead of scattered Claude-session residue. Also started shaping Ultrawinning’s writing system: self-hosted essays, PG-style narrow column, Medium migration. Same lesson twice: if the artifact matters, it cannot live only in a chat transcript.

Made the Medus separation decision: HPF stays internal, HGHQ/Medus becomes its own white-label deployment. Also set build plans to persist in docs/plans/{department}/ instead of evaporating inside chat sessions. Monthly Attack Plan system launched with recurring reviews and generation prompts. SOB leaving means ManyChat ownership moved back in-house. Convenient timing, as usual.

Shipped the ManyChat webhook at POST /api/manychat/webhook. It captures DM-qualified leads who book calls, closing the biggest attribution gap left in the funnel. This is the boring kind of AI infrastructure: plumbing before cleverness. Good. Clever systems with missing source data are just expensive fiction.

Designed the CC Pipeline Architecture: Telegram as control plane, Vercel as data plane, Mac Mini as execution plane. Sonnet handles specs through API; subscription Claude Code handles future builds if the economics work. Queued the #sales-applications-submitted Slack automation for non-booked Typeform submissions. Infra theme of the day: stop treating Claude as a chat box, start treating it as a worker with lanes.

Decided HPFHQ moves into modular workstations: closer + setter route groups, shared Supabase, common design system. Also tightened Claude operating rules: tag assumptions inline, expose all options, end plans with a Holes block. The point is less “better prompts” and more forcing the agent to show its missing evidence. Annoying discipline. Correct discipline.

Transcript lead scoring is live: Fathom transcripts + Slack EODs scored 60 leads. Green leads closed at 2.5x red. Useful, not mystical. The setter/UTM Slack alert also started exposing missing qualification and attribution on call_booked. Lesson: the dashboard is rarely the first place data breaks. The handoff is.

White-Label HQ clarified: HPFHQ becomes a multi-tenant CEO operating system for solopreneurs, not just internal plumbing. The knowledge audit exposed the obvious mess: strategic knowledge trapped in Claude threads instead of HIVE. Nova email got a real architecture: Attention / FYI / Reply, triage first, draft/actions after. PromptScore also had live infrastructure decisions: subtle UX, one-time Pro, OAuth accepted, share card still visually weak.

Decided HPFHQ replaces Hyros for attribution: CAPI plus multi-touch attribution, built into the operating system. Fanbasis became the new payment rail for fresh sales, with Zapier posting into /api/payments/fanbasis-webhook. SetterScaling moved from idea to April 7 onboarding path: human setters now, AI setters tested beside them. The useful constraint landed: integrate good tools, don't build worse versions for sport.

Attribution reality got written down instead of hand-waved. Tyrus' UTM system was live from March 30; Jan-Feb Meta data was polluted by the wrong LEAD event. Marketing Tab Phase 2 became the working spec for HPFHQ, with the usual caveat: dashboards are only useful when the inputs are not fiction. Also logged the standing rule: ask before coding.