You've already wired Claude Code into Feishu, WeChat, Slack or Telegram — bridges solved the plumbing. Then the scheduled agents start failing: lost context, double fires, zombie processes, the same mistake every morning. The Kit is the missing operations layer: battle-tested agent templates plus the playbook that keeps a real fleet reporting in every single day.
feedback.md — crypto demoted, AI infra boosted. Future runs adjust automatically. ✅Free bridges give you a scheduler in ten minutes. What they don't give you is an agent that's still useful in month two. The Kit ships the five operational systems a long-running agent actually needs — each as ready-to-paste prompt sections and file conventions:
Lessons, journal, and improvement files with strict append rules. The agent reads its playbook before work and logs after — every run learns from the last.
Your chat replies become structured feedback that provably changes future runs — likes, vetoes, and constraints all land in files the agent must consult.
Last-seen cursors, seen-lists, and idempotent run design. No double fires, no re-announcing yesterday's news, no losing your place after a crash.
Hard gates the agent can't talk itself out of: "no source, no claim", verification before completion, kill-criteria for low-quality output days.
Missed-fire recovery, health checks, and a month-two survival checklist for the failure modes every fleet hits: silent stalls, drift, context loss.
Agent templates extracted from production — each shipped with the memory/state scaffold most "templates" leave out. Two ready today (digest, researcher); three more shipping in the founder batch.
The Kit isn't written from imagination — each pattern is extracted from this fleet, which fires daily into a real chat workspace. A sample:
Gathers the last 24 hours of headlines and writes a clean morning brief into the group.
Picks one unread Anthropic / Claude Code post each day and pushes a deep, illustrated summary with source links.
Pushes a daily learning drill with a recap list, tracks progress that survives restarts, and nudges you to keep the streak.
Watches a creator / feed / repo for new uploads and pings only on genuinely new items — never re-alerts on what you've seen.
Validates one opportunity a day against ≥2 independent sources with hard kill-criteria — and reports an honest "nothing today" when nothing clears the bar.
Market intel, source-code study, a nightly self-improvement agent that opens PRs for review — all from one prompt each.
The Kit is plain markdown — prompts, file scaffolds, and checklists. It rides on whatever you already run; there's nothing to install and nothing that can break:
Digest, watcher, researcher, learning coach, or self-improving agent — each ships as a complete prompt plus its memory/state file scaffold.
Works with cc-connect, OpenClaw, any chat bridge, or bare cron + Claude Code. If your scheduler can fire a prompt on a timer, the Kit runs on it.
Wire the memory protocol, guardrails, and health checks. Your agent now learns from every run instead of repeating yesterday's mistake.
kit/
playbook/ the ops manual
01-memory-protocol.md lessons · journal · improvements
02-state-and-dedup.md cursors · seen-lists · idempotent runs
03-feedback-loop.md chat replies → behavior change
04-guardrails.md hard gates the agent can't skip
05-anti-zombie-ops.md missed fires · stalls · drift
templates/
daily-digest/ watcher/ researcher/
learning-coach/ self-improving/
checklists/
pre-flight.md month-two-survival.md
An unedited passage from the playbook — Chapter 05, Anti-Zombie Ops. This is the voice and the depth of all five chapters.
The most expensive zombie is the one that already succeeded. When the run's contract is fulfilled — output delivered, state persisted, journal appended — the agent must terminate. It does not "stay available in case." A supervisor timeout reaping a finished session is not the lifecycle working; it's a bug report that says nothing told the agent to stop.
The bonus-compute trap: when a finished agent has time left and you let it keep building, that continuation is exactly where hangs happen. Put a hard deadline on it — "build until T, then stop and exit" — or the finished agent becomes the hung one.
You could — after your agents fail the same way ours did. The Kit is the post-mortem: the memory, state, and guardrail conventions that survived months of real daily runs, so you skip the month-two graveyard. You're buying the failures we already paid for.
Plain-markdown prompts plus the file scaffolds — memory, state, dedup — for each template, and the five-chapter ops playbook. The Full Kit is delivered as a private GitHub repo invite: clone it, drop the pieces into your own agents.
If your scheduler can fire a prompt on a timer, yes. Bridge-agnostic — cc-connect, OpenClaw, any chat bridge, or bare cron + Claude Code. No SDK, no framework, nothing to install or lock into.
Templates if you just want working agents to copy. Full Kit if you want the playbook — the why behind each pattern and the ops discipline — plus repo updates as the fleet learns new lessons. Most people running agents daily want the playbook.
Full Kit buyers keep private repo access. When the production fleet learns a new lesson, the repo gets the fix and you git pull. You're buying the Kit's future, not a snapshot.
14-day refund, no questions asked. If it doesn't survive contact with your fleet, reply to your receipt and get your money back.