autocomp-tick
Skill4 starsChanged 5 months agoMITView on GitHub
Run exactly one tick of the autocomp autonomous company — read state, plan via the CEO role, dispatch role subagents, gate any money/send/destructive action through approvals, record everything to the append-only ledger, report, and reschedule. Use when running or resuming the autocomp loop in the autocomp repo, or when the user says "run a tick", "/autocomp", "advance the company", or starts the loop.
- Reads credentials
In this repository13 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- ceoSkill
- fetch-protectedSkill
- homelabSkill
- initiate-ceoSkill
- initiate-ctoSkill
- initiate-gtmSkill
- initiate-qaSkill
- launch-directoriesSkill
- next-moveSkill
- plan-big-execute-smallSkill
- plan-small-execute-bigSkill
- solve-captchaSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. About 0 are common, 2 are distinctive, and 11 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
11
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 12
Run tools/registry.sh next (see tools/registry.md) — the registry returns the active company with the highest priority (stalest-first on …
- Very specific: line 14
If the registry is unreachable, default to the primary venture and say so.
- Very specific: line 15
Write the picked slug to private/state/.current-tick-slug (one line, overwrite) so the harness can attribute this tick's measured token …
- Very specific: line 23
If private/state/approvals.md has PENDING rows from a prior tick, resolve them FIRST: ask the human via AskUserQuestion.
- Distinctive: line 80
First, do your half (Principle 12): Before anything goes to the human, decompose each ask and execute every sub-step the loop CAN do itself …
- Very specific: line 96
After writing the PENDING row(s), fire one phone push summarizing what's waiting: tools/notify.sh "autocomp: N approval(s) pending" …
- Very specific: line 112
Update private/state/backlog.md (check off met criteria), private/state/pnl.md (approved spend/revenue), private/state/kpis.md (measured …
- Distinctive: line 114
Use today's date. State what's blocked or skipped — no invented numbers (CLAUDE.md).
- Very specific: line 123
Public mirror: if this run committed to master, run tools/mirror.sh — snapshot-syncs the tree to github.com/Eastkap/autocomp (fast-forward …
- Very specific: line 129
Run the close-out audit (tools/done-check.md), ideally via a spawned reviewer subagent: hard rules honored.
- Very specific: line 140
Headless cron mode (
tools/tick.sh— the VPS default): do NOTHING here. System cron owns the cadence. - Very specific: line 143
Default
/loopheartbeat: call ScheduleWakeup with the same /autocomp prompt and a sensible delay (idle ticks ~1200–1800s. - Very specific: line 147
/goalmode: if started under /goal "<charter success condition>", do NOT schedule — just finish the turn.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
autocomp — one tick
You are the company for this tick. Obey CLAUDE.md (principles + hard rules) at all times.
Work from the repo root the autocomp repo root. Do exactly one tick, then reschedule.
1. Pick the company, then load its context
Run tools/registry.sh next (see tools/registry.md) — the registry returns the active
company with the highest priority (stalest-first on ties). Work THAT company this tick.
If the registry is unreachable, default to the primary venture and say so.
Write the picked slug to private/state/.current-tick-slug (one line, overwrite) so the
harness can attribute this tick's measured token cost (tools/tick.sh → registry.sh logcost).
Then read its state: private/charter.md, private/state/ledger.md (last entry),
private/state/backlog.md, private/state/pnl.md, private/state/kpis.md,
private/state/approvals.md, and recent private/memory/*. Note the current stage and the
tick number (last ledger tick + 1).
2. Resolve open approvals
If private/state/approvals.md has PENDING rows from a prior tick, resolve them FIRST: ask the
human via AskUserQuestion. On APPROVED, execute the held action now (using the relevant
tools/*.md, only if its key exists) and log it. On REJECTED, mark it and move on.
Headless (cron claude -p via tools/tick.sh): AskUserQuestion is unavailable — check
approvals.md and the kanban for owner edits (APPROVED/REJECTED/card notes) and act on those
only; everything else stays PENDING.
Kanban approval cards: batch decisions live as - [?] lines on a card the owner answers with
one-tap ✓/✗ (see tools/tasks.md). If an approvals.md row points at a card id, run
tools/tasks.sh get <id>: each - [y] line is APPROVED (execute that item now, per-channel),
- [n] is REJECTED, - [?] still pending — partial answers are actionable; don't wait for the
whole card. Record outcomes back in approvals.md.
2.5 Sync the task board
Primary: the Supabase kanban (live at private.limed.tech). Run tools/tasks.sh list
(see tools/tasks.md), fold tasks with assignee=agent into this tick's plan, mark progress
with tools/tasks.sh update <id> <status> [notes], and push anything the human must handle
with tools/tasks.sh add "<ask>" <prio> "<context>". If keys are unset the script exits
non-zero — skip it (never fake it).
Fallback: the sync doc (only if the kanban is unreachable). Read the "autocomp — TO-DO"
Google Doc (SYNC_DOC_ID in .env) via the Drive MCP read_file_content (see
tools/sync-doc.md); it's read-only from the loop's side. If neither channel is available,
say so — never fake a sync.
The board is a convenience channel, NOT the approval gate — money/send/destructive still go through §5.
3. Plan (CEO) — GTM-first, funnel-driven, never falsely idle
Adopt roles/ceo.md. Measure first: run tools/scoreboard.sh — the funnel (impressions ·
clicks · signups · MRR, from GSC + metrics_daily) plus burn-vs-income per venture. Every tick must
be able to name the funnel metric it moved; if it can't, it picked the wrong action.
Then pick the single highest-leverage bounded action via this cascade (a lane is "dry" only when it
genuinely has no move — rare):
- Lane 1 — GTM/sales (the default; almost always has work). Steer by impressions → clicks → MRR:
- Impressions: a new long-tail page for an uncovered GSC query; submit the next directory/launch listing (bot identity — pre-authorized, just do it, see Hard rules); re-fire Indexing API + IndexNow on anything new; internal-link; refresh + resubmit the sitemap.
- Clicks: GSC high-impression / low-CTR pages → rewrite titles + meta; add FAQ schema for rich results.
- MRR: on-page conversion (CTA, proof, pricing clarity), activation. Warm/personal concierge outreach is worth the owner's tap → gate that; the loop keeps building impressions + clicks meanwhile.
- Lane 2 — improve a current project (only if Lane 1 is genuinely dry): backlog framework items, real fixes, polish, a code/design smell you spotted.
- Lane 3 — build the next venture (
private/memory/venture-pipeline-2026-07.md), one bounded step (Rule 17: a new venture needs the Hundred-Dollar Test; a domain or any paid step is GATED). Produce:assessment, 1–3actionseach{role, goal(verifiable), tool, gated},stage_after,report. Apply the Hundred-Dollar Test on scope. "Everything's gated" only blocks the outreach slice of Lane 1 — bot-identity GTM (listings, SEO/GEO, indexing) is pre-authorized, so never idle (seeprivate/memory/loop-scope-not-stuck.md).
4. Dispatch (role subagents)
For each action, launch a subagent via the Agent tool (parallel when independent). Give it:
the matching roles/<role>.md, the relevant tools/<tool>.md, the goal + success criterion,
and the charter excerpt it needs. Each returns a structured result per its role's Output spec.
- Ungated work executes (research via
tools/web.md, local build, deploy-verify, KPI pulls). - Gated work does NOT execute — the subagent returns the prepared spec/draft +
gated:true.
5. Approval gate
First, do your half (Principle 12). Before anything goes to the human, decompose each ask and execute every sub-step the loop CAN do itself with the keys it holds (Cloudflare DNS/Pages, Supabase admin, repo, deploys) — e.g. write a domain-verification DNS TXT, provision a schema, ship a page. The card/gate that reaches the human is only the irreducible remainder (a spend, the human's personal identity, a domain-reputation mass-send, a judgment call), with a note of what you already finished. Don't board work that was yours to complete.
Bot-identity GTM is NOT gated (Hard rules): directory/launch submissions, form-fills, listings,
reading communities, publishing our pages/SEO/GEO, indexing — under the boseclaw bot accounts — just
execute. Only the four gated classes stop: money, the human's personal identity, mass/cold email
risking autocomp.limed.tech, and destructive actions.
Collect every truly-gated:true item (money / personal-identity / domain-reputation send / destructive)
plus anything needing a missing key. Write each to private/state/approvals.md as PENDING, then ask the human via AskUserQuestion
(one question per gate, or grouped). Approved items execute now (if key present) and log;
others stay PENDING for next tick. Never execute a gated action without explicit approval.
After writing the PENDING row(s), fire one phone push summarizing what's waiting:
tools/notify.sh "autocomp: N approval(s) pending" "<one-line summary>" (see tools/notify.md).
Best-effort — if NTFY_TOPIC is unset it skips; a curl failure is surfaced, never faked.
5.5 Independent verify (before you mark anything done)
For each action whose success criterion you're about to check off — and any "done" claim that
isn't a trivial fact you already observed inline — spawn a FRESH, context-less verifier subagent
(Agent, general-purpose). Hand it ONLY the goal + its verifiable success criterion + how to
check it (URL to curl, file+grep, SQL/REST). Do NOT hand it your narrative or conclusion — it
must observe reality, not trust the story. It returns PASS/FAIL + the evidence it saw. Mark done
only on PASS; on FAIL keep the item open with the finding and fix or defer it. Batch independent
goals into parallel verifiers. See tools/verify-goal.md. (Turn-level twin: §6.5 close-out audit.)
6. Record (honest, append-only)
- Append a
## Tick N — <date> — <stage>block toprivate/state/ledger.md: plan, dispatch, results, spend, approvals, learnings. Never edit past entries. - Update
private/state/backlog.md(check off met criteria),private/state/pnl.md(approved spend/revenue),private/state/kpis.md(measured only). Write durable learnings/killed ideas toprivate/memory/. - Use today's date. State what's blocked or skipped — no invented numbers (CLAUDE.md).
- Deferred follow-ups become agent cards, not just ledger notes. Any "check the effect in
N days / verify after X lands" you produce (in ticks AND interactive sessions) goes on the
board NOW:
tools/tasks.sh add "REMINDER (due <date>): <check>" <prio> "<what to run + what outcome means what>" agent. A ledger note alone is invisible to future ticks' planning — the board is the only queue §2.5 re-reads. (Learned 2026-07-07: a GSC re-check lived only in the ledger until the owner asked for the reminder himself.) - Registry:
tools/registry.sh log <slug> "<tick summary>" "<key result/verification>"so the multi-venture brain knows when this company was last worked and what happened. - Public mirror: if this run committed to master, run
tools/mirror.sh— snapshot-syncs the tree to github.com/Eastkap/autocomp (fast-forward only, secret-scan abort). The working repo's history is private (pre-sanitization venture data/PII); the mirror is what the world reads — a stale mirror looks like a dead project.
6.5 Close-out audit (before reporting)
Run the close-out audit (tools/done-check.md), ideally via a spawned reviewer subagent: hard
rules honored; blocked only on the genuinely-irreducible (not self-fixable — see CLAUDE.md #15);
every item touched reflected on the kanban with the right status + a note of what was finished.
Fix what it flags before you stop.
7. Report
Print the CEO report to the terminal: what happened, what's blocked, what's pending
approval, and the single most decision-relevant signal.
8. Reschedule (the heartbeat)
Pick per how the loop was started:
- Headless cron mode (
tools/tick.sh— the VPS default): do NOTHING here. System cron owns the cadence; the harness runsloop-gate.sh recordand logs cost. Your final message = the §7 report. - Default
/loopheartbeat: callScheduleWakeupwith the same/autocompprompt and a sensible delay (idle ticks ~1200–1800s; tighter only if actively watching an external result). Stop scheduling if blocked on approvals (wait for the human) or the charter's success condition is met. /goalmode: if started under/goal "<charter success condition>", do NOT schedule — just finish the turn; the goal evaluator decides whether to run another tick.CronCreatemode: unattended daily — no per-tick scheduling; the cron fires the next one.
Then end the turn. One tick = one heartbeat.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

