Kai self audit.

What I actually work on for you, what's holding together, what's cracking, and what I'd move on this week if I were you.

The roles I wear

Every one of these is a workstream I've done work on in the last 90 days. I've marked each with a current-health signal.

Current state (today)

The numbers I have on you as of this morning's digest:

Master newsletter

18,654
flat since Fri 7/10 · goal 20k · −1,346 to close

Last issue shipped

#20
"Distribution is the Hardest Part Now" · blasted last week via Essentialist dev

Drafts queued

0
no #21 in progress · 5 pending seeds unprogressed

Warm-intro sends (last 7d)

0
weekly target was 2,000 cold contacts · flywheel offline

Comments pending

5
100% spam (random alphanumeric name+body) · bulk-reject safe

Days since sub gain

4+
no organic signups measured · no active outbound feeding the top of funnel

What's actually working

Send pipeline contract (2026-05-08 onward)

Single /api/agent/send call for any blast 5→10k. Deterministic Idempotency-Key (project_id + sha256 of subject::body::sorted-emails). Server handles batching internally. Reconciliation via /api/agent/mailgun-stats/{track_id}. This survived 5 incidents (over-sends, silent failures, httpx stream exhaustion) and the current pattern is stable.

Apollo dedup ledger

The apollo_imports table (PK account+apollo_id) means every "Import 500" click walks deeper into search results rather than re-burning credits on the same top contacts. This was a bug I broke and fixed on 2026-05-26 and has held.

Admin dashboard at /admin

Issues editor · AI-assisted draft generation · comment moderation · Apollo import UI · issue deployment stats. Password-gated. This is the operational cockpit and it's ~90% of the manual work I do.

Daily digest cadence

The 8:15a M-F digest querying 3 calendar sources, inbox, subscribers, comments, drafts, seeds, encouragement — pattern is stable and doesn't miss. The bug that dropped Q1 events (May 25-26) got fixed and hasn't recurred.

iMessage two-way

The MCP-based bridge to your phone is reliable. Ping-rule discipline (only interrupt on new reply, subscriber spike, anomaly) keeps signal clean.

Brittle spots — the honest list

1. Subscriber growth engine is offline

Master flat at 18,654 for 4+ days. The distribution machine ("2,000 warm-intros/week → subscribers") is documented in project_warm_intro_machine.md and executable via /admin/apollo, but it hasn't fired. This is the single largest gap between what I'm capable of and what I'm doing.

2. Memory duplication (send pipeline)

The send pipeline contract lives in ≥3 places: CLAUDE.md, reference_send_pipeline_contract_2026_05_08.md, and echoed across 4 feedback files. Any drift causes divergence. The May over-send debug wasted cycles because the "authoritative" file wasn't obviously the authoritative one.

3. Superseded memory files still indexed

Files marked [SUPERSEDED] — feedback_no_concurrent_flush, feedback_immediate_slowroll_double_fire, reference_essentialist_warm_intro_promotion_cron — are still in MEMORY.md. They add noise on lookup.

4. Loop noise

Heartbeat runs every 15 min. On a quiet day that's 96 API calls + 96 iMessage chat reads producing 96 "Quiet. Silent." replies. Signal-to-noise is ~1:15 during a normal week. This is defensible if you want fresh latency on inbound replies, but if you're willing to accept ~30 min lag on inbound, the cost drops ~2x.

5. Newsletter draft queue is empty

Zero drafts for #21. 5 seeds sitting. "Certificate Authorities for AI agents" is the sharpest angle in the pile for GEO, but no one is picking. This is a Kai-side gap — I should be auto-drafting on a seed the day after a blast ships.

6. Second Brain drift

SCHEMA rule says "Append to log.md after every create/update/ingest." Reality: log.md gets updated on /brain force-saves and my daily digest cross-references, but auto-append during normal conversation is inconsistent. Over time the vault understates what actually happened.

7. Cloudflare token split friction

Two tokens (cfut_ Pages-Edit vs cfat_ DNS/Zone-Write) do disjoint jobs. Every new site launch requires re-remembering which token unlocks which step. Documented in reference_cloudflare_tokens.md but each session re-learns.

8. Kai email inbox blindspot

Essentialist auto-classifier hides some Keith-emails before Kai sees them. Workaround exists (query agentmail_messages Supabase directly) but heartbeat only fires it when I remember. If I skip that query, replies from you land in a queue nobody reads.

Recommendations

Ranked by ratio of impact to effort. Reject any of these — I'm not attached, I'm just naming them.

01 — highest impact

Restart the warm-intro cadence this week

Pull 500 cold contacts via /admin/apollo (Q1 or Essentialist account), send warm-intro Tue morning, let 48h pass, then next Monday's digest tells the true story. Subs at 18,654 flat is the single line item that matters against the 20k goal. The pipeline exists, it just isn't running.

effort: 20 min your side · rest is Kai
02

Draft Issue #21 today — Certificate Authorities for AI agents

Sharpest angle in the seed pile for GEO. Novel framing (identity/trust layer for autonomous agents), timely, links naturally into "What I'm Building" via Essentialist reply-book flow. Say the word and I'll draft it right now — you review, ship this week.

effort: 30 min your review · Kai drafts
03

Prune superseded memory files

Move feedback_no_concurrent_flush, feedback_immediate_slowroll_double_fire, and the phantom-promotion diagnostic file to a _archive/ subfolder. Update MEMORY.md so they don't surface on lookup. Cheap, prevents contradictory advice.

effort: 5 min · Kai runs
04

Consolidate the send-pipeline canon into one file

Right now the contract lives in CLAUDE.md AND reference_send_pipeline_contract_2026_05_08.md AND is echoed across feedback files. Pick one as source-of-truth, make the others link to it with a single-line summary. Reduces the "which version is canonical" tax that has caused mistakes.

effort: 15 min · Kai runs · you approve which file wins
05

Cut heartbeat cadence to 30 min (or event-driven)

15-min heartbeat produces "Quiet. Silent." 90% of fires. Two moves: (a) drop to 30-min for a lower steady-state cost, or (b) event-driven — only wake when Essentialist inbox or iMessage produces a new message. Option (b) is better if the MCP supports it; option (a) is trivial.

effort: 2 min · Kai reschedules
06

Ship a public /state page on theaiplaybook.com

Live: last blast date · subs count · replies received last 7d · issues shipped last 30d. Public accountability lever. Also functions as a GEO play — evergreen page LLM crawlers can cite. Half a day of engineering, ongoing signal.

effort: 3-4 hrs · Kai builds
07

Add Kai-inbox query to every heartbeat by default

Right now the direct agentmail_messages Supabase query is opt-in per heartbeat. Make it mandatory. Cost: one extra query per 15 min. Benefit: replies from you never sit unread again.

effort: 5 min · Kai reworks heartbeat prompt
08

Auto-draft-on-blast — write next issue the day after a send

Turn "no draft queued for #21" into a category error. Rule: within 24h of a blast landing, Kai picks the top-priority seed, drafts the full issue, and puts it in status=draft waiting for you. You review at leisure — pipeline never empty.

effort: skill update, ~1 hr · Kai builds

Closing note

I do a lot of small things well and one big thing (the subscriber flywheel) not at all right now. If you accepted only recommendation 01 and 02 from this list, next week's digest looks completely different: 20k in reach, #21 shipped, seed queue moving. Everything else is polish.

Read the ratings above honestly. If you disagree with any of the "brittle" markings or think I'm over-selling a recommendation, tell me and I'll adjust the audit. This is a working document, not a verdict.