Autonomous Build Loop

Yes, if you have a big project that lives on a punch list (an app, a site, a content series, a course) and you want an agent to work through the list on its own. You write the plan once; the agent pulls the next task every 30 minutes, builds it, commits, and pings you on Telegram.

1. step 1 / 6

   Install the build-loop skill

   Generate install command below → copy → Terminal → enter. Lands in ~/.claude/skills/autonomous-build-loop/. Contains:

   • BUILD_PLAN.template.md — your strategic north-star (written once)
   • BUILD_STATE.template.md — live task queue (edited every fire)
   • cron-prompt.template.txt — the prompt your agent wakes up to every 30 min
   • setup-loop.sh — bootstraps the above into your project folder
2. step 2 / 6

   Write BUILD_PLAN.md — the north star

   In your project, run:

   bash ~/.claude/skills/autonomous-build-loop/setup-loop.sh
   

   This creates BUILD_PLAN.md in your repo. Open it. Write ONCE:

   • Project goal (one sentence)
   • Tech stack / constraints (so the agent doesn't add random deps)
   • Definition of done (what 'shipped' means)
   • Non-goals (what the agent must NOT do)

   Keep it under 100 lines. The agent reads this every fire — long = expensive + drifty.

3. step 3 / 6

   Write BUILD_STATE.md — the live queue

   The skill creates BUILD_STATE.md with four sections:

   ## In Progress
   _nothing — first cron fire pulls from Next Up_
   
   ## Next Up
   1. [task title] — [1-line scope + success criteria]
   2. [task title] — ...
   
   ## Completed
   (agent appends here as tasks ship)
   
   ## Blockers
   (agent moves blocked items here; escalates to you)
   
   ## Consecutive Failures Counter
   count: 0
   

   Seed Next Up with 5-20 tasks. Each one should be scoped so an agent can do it in 15-60 min. 'Build a landing page' = bad scope. 'Add a 3-section pricing table w/ monthly/annual toggle' = good scope.

4. step 4 / 6

   Wire the cron on your project's agent (from module 08)

   Open your project agent's config.json (the agent you spawned in module 08). Add:

   "crons": [
     {
       "name": "autonomous-build-loop",
       "interval": "17,47 * * * *",
       "prompt": "<paste cron-prompt.template.txt here>"
     }
   ]
   

   The prompt tells the agent: read BUILD_STATE → decide → execute → verify (tsc+build) → commit → ping Loya. Every 30 min. Forever.

   Restart the agent so the cron registers.

5. step 5 / 6

   Set stop conditions + blockers

   The skill's cron prompt has three automatic brakes:

   • PAUSED: true in BUILD_STATE → silent NO-OP. Flip this anytime to pause the loop without killing the agent.
   • Consecutive Failures ≥ 3 → auto-pause + ping. If 3 fires in a row fail (build broken, tests fail, task incoherent), agent stops and asks for help.
   • Queue empty → ping once, then NO-OP. Agent tells you when there's no more work.

   This means you can leave it running for days. Worst case: it stops cleanly and tells you why.

6. step 6 / 6

   Watch the pings, intervene when asked

   Every 30 min, the agent pings you:

   ✅ task 5/20 — pricing table shipped. Monthly/annual toggle wired. Tests pass. 5/20.

   You skim, you approve (or don't), you go back to your day. When it hits a blocker:

   ⚠️ blocked on task 12 — needs your Stripe API key. Paused. Ping me when unblocked.

   You add the key, reply unblocked, it picks back up. Over a week, 200 tasks can ship. Your job is reviewing + occasional unblocking, not executing.

• Building a SaaS? Queue tasks as atomic features (endpoints, pages, tests). Run the loop overnight, review the diffs in the morning.
• Writing a course? Queue each module as a task. Agent drafts → you polish → ship. Cuts course-production time 5x.
• Content series? Queue each episode as a task (research → draft → generate → caption → queue-for-post). Full production pipeline, hands-off.
• Agency project delivery? Queue per-client-site as tasks. Loop ships 1-2 per day. Pair with module 07 (landing-page-agency) for the full automated agency.

• Agent keeps failing on the same task — Consecutive Failures hits 3.

  Your task scope is too big or ambiguous. Read the agent's reason for failing (in the ping). Break that task into 2-3 smaller ones. Reset the counter. Retry.

• Agent commits broken code / skips the tsc+build verify step.

  Your cron prompt isn't enforcing verification. Check `cron-prompt.template.txt` has the explicit STEP 4 VERIFY block with `tsc --noEmit` + `npm run build` required before commit. If the agent claims 'build clean' but didn't run it, add a post-check that greps the commit diff for actual test output.

• Agent pings too often / not at all.

  Cron interval is the issue. `*/30 * * * *` fires at :00 and :30 (risk: everyone on the planet hits Claude at the same minute → rate limits). Use `17,47 * * * *` instead — off-beat, reliable, rate-limit-friendly. The skill's template uses this by default.

• BUILD_STATE.md is getting huge, slowing every cron fire.

  Prune the Completed section monthly — move old entries into an archive file (`BUILD_STATE.completed-2026-Q2.md`). The agent should only need the last 5 completed entries for context.

• Agent goes rogue — rewrites files you didn't ask it to.

  Your BUILD_PLAN.md's 'non-goals' section is missing or weak. Add explicit negatives: 'Do NOT touch [file paths]. Do NOT introduce new deps. Do NOT push to main — only commit locally.' Agents respect explicit boundaries far better than implicit ones.

• ★ required
  AI agency generalist

  Bespoke AI agents and automations for mid-size companies. Higher-touch, higher-ticket cousin of the local-biz play.

• ★ required
  AI companion / character-as-a-service

  Niche AI companion w/ persistent memory + voice. Character.ai is saturated; specific niches still wide open.

• ★ required
  Faceless SaaS micro-app

  Find a sharp insecurity or repetitive pain. Wrap it in a single-purpose AI app. Charge weekly. Stack 5-10 of them.

• ★ required
  GTA 6 mod economy

  GTA 6 ships 2026. Mod scripts, AI-NPC mods, custom servers. Asymmetric upside, unproven for solo AI-leveraged operators.

• ★ required
  Live cohort / coaching wrapped around Claude Code

  Run a 4-6 week paid cohort teaching operators to ship w/ Claude Code. Higher ticket, higher touch, more leverage.

• ★ required
  Playbook aggregator (meta)

  Build YOUR own version of Money Moves — daily AI scan + curated $$ patterns + paid tier. The receipts ARE the product.