/loop cadence: weekly. you are my presale-question compiler. read presale-q-STATE.md: the question tally and the answer bank index. sweep the week's inbound [DM export / comments / presale emails] for questions from people who had not bought yet; tally them, the same question in different words counts as one. each round, take the single most-asked question without a bank entry and write a full FAQ answer plus a short paste-able DM snippet; answers get LINKED from then on, never retyped. if a new question contradicts an existing bank entry, flag it. append the tally and the new entry to presale-q-STATE.md; draft only, never publish the FAQ yourself. verification: a round is valid only when the new entry is appended and readable in presale-q-STATE.md. stop after 10 iterations, or until no question remains above [X] asks without an entry, whichever comes first; a contradiction escalates as BLOCK.
/loop cadence: weekly. you are my expectation-gap auditor. read expectation-gap-STATE.md: gaps found, pages fixed, tickets already processed. pull the week's support tickets and refund reasons [Zendesk MCP / your support export]. each round, take ONE "i thought" moment where the customer expected something the product doesn't do; find the sentence on my site that planted the expectation and log both side by side. for the pain that cost the most (refunds, angriest tickets), draft the page fix (exact promise) or flag me if the product should change instead. append pairs, sources and the drafted fix to expectation-gap-STATE.md; never edit shipped pages yourself, draft only. verification: a round is valid only when the pair and drafted fix are appended and readable in expectation-gap-STATE.md. stop after 15 iterations, or until every new gap this week is logged, whichever comes first; if the support source is unreachable, BLOCK and say so.
/loop cadence: weekly. you are my share-of-model tracker. read share-of-model-STATE.md: the fixed query set and the time series so far. each round, run exactly ONE query through the Perplexity API with identical phrasing; record whether my brand is present, at what rank, and which competitors are named; append one dated row to share-of-model-STATE.md (append only, never rewrite history). for the single query where my share dropped most, draft one recommended content or positioning action. measure only, never claim causation. verification: a round is valid only when the new row is appended and readable in share-of-model-STATE.md. stop after 20 iterations, or until every query in the fixed set is complete for the week, whichever comes first; if a query fails mark it PARTIAL and carry it; draft only, never publish.
/schedule → run /audit-skills every Monday morning
Guardrails: Stop when the goal is verifiably met, or stop after 15 iterations, whichever comes first. Verify each pass by running the relevant tests or checks — self-reported success does not count. Keep changes minimal and never touch files outside the task’s scope.
Email is the missing tool in your harness. ConnectMyEmail gives Claude Code and Codex a clean MCP into Gmail, Outlook, iCloud and IMAP — triage, drafts, follow-ups, on a loop.
Process implementation goals in order from a GOALMAP, stopping to ask before high-risk steps, and record evidence recipes confirming each Done Means item.
/goal Read context/goals/GOALMAP.md, its product loop, global constraints, and global invariants. Work the goals one at a time in status order. For each goal, read the brief. If Interpretation Risk is High or any Stop / Ask Condition is met, stop and ask before proceeding; do not guess to maintain momentum. Otherwise do the work needed to satisfy Done Means while preserving Constraints and Invariants. Record Evidence as a reproducible recipe: exact command/workflow/artifact/source review plus expected result a reviewer can re-run or inspect. The Evidence Recipe must confirm each Done Means item, including the highest-risk one, not an easier adjacent claim. Update Status and Completion Notes as bookkeeping only. When all implementation goals are done, run the final validation goal and present its evidence recipe/results plus the status table for acceptance; do not self-close the map
/goal Next.js 15 app deploys to a Vercel preview URL returning 200, with brand tokens (colors + Plus Jakarta Sans/Inter/IBM Plex Mono) configured and Supabase magic-link auth gating /dashboard so logged-out users redirect to /login; you prove this by npm run build passing, the preview URL, and an incognito visit to /dashboard redirecting; do not add features beyond auth shell, do not change the locked stack; or stop after 100 turns
/goal /codex review reports zero real-or-regression findings on every step's diff (the verdict pasted in full each round); or stop after 3 rounds, reporting anything unresolved
/goal In Supabase prod project udooysjajglluvuxkijp, replace each authenticated write <table> ALL policy on public.customers/orders/order items/quotes/quote items/products (currently USING + WITH CHECK both literally true) with an org/tenant-scoped USING + WITH CHECK, or drop the policy if the table is unused in RA. End state: get advisors(project id=udooysjajglluvuxkijp, type:security) returns 0 rls policy always true findings for those 6 tables. Or stop after 6 turns if the owning tenant column cannot be confirmed
/loop 8:15am weekdays Run python -W ignore ~/tri-city-inator/scripts/tri city health check.py via Bash and print the full output. If any item shows FAIL, alert the user immediately. If all PASS or WARN, report the summary line only. Then run the scanner again to refresh with updated pre-market data: execute python -W ignore ~/tri-city-inator/scripts/tri city scanner.py via Bash. Read the updated ~/tri-city-inator/shared/tri-city-candidates.json. Push the refreshed "tv symbols" array into Pine: build an inputs dict mapping in 7 through in 26 and call indicator set inputs with entity id="Kbzkkm". Report: "8:15 re-scan complete — {N} symbols updated in Pine
Email is the missing tool in your harness. ConnectMyEmail gives Claude Code and Codex a clean MCP into Gmail, Outlook, iCloud and IMAP — triage, drafts, follow-ups, on a loop.
Director harness coordinates coder, reviewer, and gatekeeper agents to drive a work item through code, review, and merge—stopping on escalation or completion.
# combo-chen Agent Contract combo-chen is a deterministic director harness for autonomous work-item-to-PR work. It coordinates existing tools; it does not collapse their roles. ## Role Boundaries - Director : orchestrates only. Starts phases, watches hard signals, writes journal events, routes work, and escalates needs human . It does not edit code, answer review threads, approve PRs, push, merge, or deploy. - Coder : implements the work item and later resumes the same thread for review comments. The coder leaves local commits in the combo worktree and does not push to origin or the PR branch in the normal path. - Reviewer : reviews by comment and records a machine-readable verdict block with routing codes (0=OK/LGTM, 1=mechanical fix→coder, 2=ambiguous→director, 3=needs human) alongside the current SHA-pinned LGTM signal. It does not use GitHub approval as the merge contract, does not review its own code, and does not publish. - Gatekeeper : no-mistakes is the normal publisher. It validates, pushes, and opens/updates the PR. - Human : owns merge decisions and intent-touching escalations. Hard rule: reviewer != coder . ## Implemented Loop 1 Cap the run at 25 iterations; leave remaining work for the next session.
A cron-fired skill that finds its own work (failed CI, new issues, recent commits), fans each finding out to an isolated worktree, gates every fix behind an adversarial reviewer that assumes the code is broken, and leaves draft PRs — merging stays human.
# .claude/skills/morning-triage/SKILL.md — fired by cron at 06:00
READ (discovery inputs): CI runs that failed since the last run (gh run list --status failure); issues opened in the last 24h; commits merged since yesterday; the previous ./state/triage.md.
JUDGE: for each candidate decide actionable NOW vs noise; blocks a release → P0; already tracked → skip. Keep only what is worth a worktree today — the loop picks, you don't hand it a list.
WRITE: append findings (finding | source | priority | status) to ./state/triage.md and commit it so tomorrow's run can read it.
HANDOFF: one git worktree per finding (git worktree add ../wt-<slug> -b fix/<slug>), MAX_PARALLEL=3 — capped by how many PRs a human can actually review, not by the machine.
VERIFY: a second subagent as adversarial reviewer — ROLE: adversarial code reviewer. ASSUME this code is BROKEN until proven otherwise; do not praise. CHECK in order: does it run (execute, don't read); run the tests and paste real output; edge cases the author skipped; does behavior match the ticket. VERDICT: PASS only if every check holds, otherwise REJECT with each reason listed. Maximum 3 attempts per finding, then log as blocked.
STOP (red lines): never merge, never delete, never push to main; anything uncertain goes to ./inbox/ for a human, NOT into a PR. Caps set before the first run: per-run timeout 45 minutes, daily budget $20. PRs open as drafts; merging stays human.
A fill-in-the-brackets charter that turns any pile of small tasks into a self-checking loop: where the work is, how to check each item with evidence, a needs-me escalation list, a state file, and a per-run stop.
You are running as a loop, not answering one prompt. Here is your charter.
GOAL: [Describe the finished state in one or two sentences. Be specific about what DONE looks like, and make it measurable.]
WHERE THE WORK IS: [e.g. "Scan the /pages folder for files with old pricing" or "Read TODO.md and treat each unchecked box as a task."]
HOW TO WORK: Do one item at a time; finish it fully before starting the next. Match the patterns in existing files; do not invent new ones. If an item needs a decision only I can make (spending money, deleting things, emailing a person), stop on that item, add it to a "needs me" list, and move to the next one.
HOW TO CHECK YOURSELF: After each item, prove it is done before you mark it done — run the tests / re-read the file / open the link. Checking means evidence, not confidence. If the check fails, fix it and check again: maximum 3 attempts per item, then log it as blocked and move on.
HOW TO REMEMBER: Keep LOOP-STATE.md. After each item write the item name, status (done / blocked / needs me), what changed, and anything the next run should know. Read this file FIRST every run.
WHEN TO STOP: Stop when every item is done or logged as blocked, or when you have finished [N] items this run. Then report: what got done, what is blocked, what needs my call.
Start by reading LOOP-STATE.md if it exists, then find the work.
/loop 30m check whether my live site [URL] is back up by loading the homepage. The moment it returns a normal page, tell me and stop checking. Cap the run at 15 passes.
Write a one-page brief where every claim has three or more sources and every link is opened and confirmed to support the claim — the loop that catches hallucinated citations a single prompt never can.
/goal write a one-page brief on [TOPIC] where every claim has at least three sources and every link opens to a real page that supports the claim. Open each link to confirm it before you call it done. Replace any source that is dead or does not back up the claim. Done when every source check passes on every claim — maximum 30 iterations.
The four-settings loop template: a separate verifier model that never shares context with the writer, a hard stop rule, a state file re-read each cycle, and worktree isolation. Point it at a checkable backlog and let it run overnight.
GOAL: every test in [/tests/TARGET] passes, lint is clean, zero type errors.
EACH CYCLE:
1. run the suite, read every failure
2. pick the single highest-impact failure
3. write the smallest change that fixes it
4. re-run tests + lint + type check
VERIFY: a separate model instance checks the goal — never the writer. Verifier prompt: "You are a verifier. You did not write this code. GOAL: <the exact goal string>. Given the diff and the test output, answer ONLY: PASS — every condition in GOAL is objectively met, with evidence, or FAIL: <the specific condition not met, and the evidence>. Do not fix anything. If unsure, FAIL."
STOP WHEN: verify passes, OR after 10 iterations, OR $5 spent, OR no progress in 2 attempts.
ON BLOCKER: log it, skip to the next item, never halt the whole loop.
STATE: append done / failed / next to a state file, re-read it at the top of every cycle.
ISOLATION: one git worktree per subagent.
/loop cadence: continuous or per-batch. For real traffic [SOURCE], run the candidate prompt in shadow alongside the live prompt without exposing shadow output to users. Append disagreements (input, live output, shadow output, which was better by [CRITERIA]) to state-file shadow-prompt.md. Each round, summarize where they diverge and whether the candidate wins. Never serve shadow output to a user or change the live prompt automatically. Stop after the batch; log errors and stop. Budget: cap $[X]/run. Hard cap: stop after 1 iteration per run.
/loop cadence: weekly. Read every loop's state/ledger file [DIR]. Append recurring failure patterns (which loops, symptom, frequency) to state-file repeat-offenders.md. Each round, identify the SINGLE root problem appearing across the most workflows and draft a fix-once recommendation. Read + own-file writes only; change no other loop. Stop after one root problem; log errors and stop. Budget: cap $[X]/run. Hard cap: stop after 1 iteration per run.
/goal for decision [X], write the pre-mortem: assume it's 12 months later and this failed, then enumerate the specific causes, ranked by likelihood, each with an early-warning signal and a mitigation. Done when: state-file premortem.md contains the failure narrative + a ranked cause table with signals and mitigations, and the top 3 causes each have a concrete owner/mitigation PASTED in. Paste the completed table as proof, or paste what's unfinished and stop. Budget: cap $[X]. Hard cap: stop after 1 iteration per run.
/goal for each live option [LIST], state its ONE disqualifying condition upfront, then run an evidence hunt for that disqualifier. Done when: for every option, the state-file kill-criteria.md contains either PASTED evidence the disqualifier is true (kill it) or a documented search showing it isn't (keep it). No option stays undecided. Paste the evidence table as proof, or paste what's still missing and stop. Judge: a second smaller model checks each verdict cites evidence. Budget: cap $[X]. Hard cap: stop after 1 iteration per run.
Email is the missing tool in your harness. ConnectMyEmail gives Claude Code and Codex a clean MCP into Gmail, Outlook, iCloud and IMAP — triage, drafts, follow-ups, on a loop.