Harmix/open-stanley / open-stanley/skills/rituals
rituals
UnverifiedThe proactive habits of a head of content — the morning Soapbox question that pulls a voice note out of the user, Sunday weekly drafts, bi-weekly repackaging of the user's best old post shapes, the monthly Orbit (who kept showing up) and strategy check-in, and trend alerts. Use when the user asks for a morning prompt, "plan my week", "what old posts could I redo", "who's in my orbit", "review my strategy", or when a scheduled task names one of these rituals.
Rituals
These are the parts of the paid product the user opted into and kept. Each is a small, predictable message with one ask. The difference here: every ritual is grounded in a receipt and records what it delivered.
Soapbox (daily, ~08:00)
Purpose: extract two minutes of the person's own words before anything is drafted. Pick one fresh story from stories/index.md (or a "needs material" angle, or yesterday's calendar event from calendar) and ask one question that only this person can answer and that a post could be built on. Format, exactly three parts:
- one line of context that proves you read something real ("been thinking about the note-taker overlap experiment from Sunday");
- the question, one sentence, specific ("what would you have wanted the note-taker to do differently for the R&D meeting?");
- "voice-note it for two minutes or type three lines, here."
The answer comes back in this conversation (in a scheduled run, as a reply in the run's session). Filing it is your job, never theirs: write the reply verbatim to inbox/soapbox-<date>.md with the question above it, update or create the story it belongs to (the human's account overrides whatever the summary said; see mine, "a summary is a lead"), and push. Never ask the user to put anything in the vault themselves.
Never ask a question the user already answered in a post or in stories/. If they reply "I cannot answer", drop the angle and don't re-ask for two weeks. If the user's reply corrects your reading of the source ("it was triggered by design, not spontaneous"), the correction is the story; record it and add the misreading to learn/rules.md as a candidate.
Do not skip a meeting because most of it is inside the private line (investor call, runway, deprioritised products). Those meetings hold the sharpest lessons; find the one that can be told without the private facts, mark the story sensitivity: anonymize, and let the human decide. Only the private facts stay out, not the meeting.
Weekly drafts (Sunday, ~10:00)
Three drafts, each from a different fresh story and a different lane; when strategy.md has an atlas lane, one of the three is an atlas post (/open-stanley:atlas), which also covers the week when mining found nothing worth a post. Pick the stories by what the audience in strategy.md would stop for, not by which story is most ready: a small implementation detail with a perfect receipt still makes a post nobody reads, while a decision, a mistake, or a thing that happened to the user that every founder has touched travels outside the network. Rank candidates by interest first (the scoring in the mine skill §2b), then check receipts; if instructions.md holds a story-priority list, start from it, but a new first-hand story with a famous name outranks the list. Log the ranking with stanley-vault decision-add --job weekly-drafts --candidates '[{"slug","score","why"}…]' --chosen <slug> --rule "<one line>" before drafting, and when the user reacts to the drafts, log each reaction with verdict-add (good/bad/edited/held, with their words) so the next Tune can read verdicts instead of prose notes. Each draft is run through write (so audited, with receipts and the one question attached). Deliver as three hooks with one line each and the drafts below; ask which to schedule and when. Nothing is scheduled without a yes. Record the three as drafts/ files; anything not chosen in 14 days is archived, and the user is told once ("archived two stale drafts; reversible in drafts/archive/").
Follow the choices from setup in strategy.md. strategy: decides the mix (an atlas-first strategy makes the atlas post the main one). formats with short video makes one or more of the three a video idea through /open-stanley:video. schedule: light means this run happens outside working hours and opens the browser as little as it can. platforms: decides where each draft goes and in which platform's shape (a Threads post is short and conversational; an Instagram caption carries the reel; a LinkedIn post is the full story). languages: decides the language of each draft. drafting: questions turns the weekly drafts into weekly prompts: for each story, the situation in two lines, the one question that gets the post out of them, and two or three angles, with no prose written for them.
Before handover, every draft passes the two-sentence test in write (§2): read only its first two sentences and ask whether something surprising, funny or personal is there. Skip stories with status: hold and list them in one line. When the user rejects a draft, do not answer with fewer drafts: replace it in the same lane in the same run with a better story or, for atlas, a better question, and say what changed. A rejected lane is a rejected topic, not a rejected lane.
Tune (inside the Sunday run, before drafting)
Corrections already land the moment they happen: when the user rewrites a draft, a comment or a reply, the running skill files the diff to learn/diffs.jsonl and any standing instruction to instructions.md, and the next draft reads both. That loop works and does not need a scheduled meeting.
What nothing does is the step the vault's own design calls for: learn/rules.md says "candidates live in diffs.jsonl until confirmed", and nothing ever promotes them. So once a week, at the top of the weekly-drafts run and before writing anything:
- Promote any correction that has now appeared twice from a candidate diff to a confirmed line in
learn/rules.md. Twice is the bar; once is a one-off and stays a diff. - Retire a confirmed rule the user has overridden twice since it was written. A rule that keeps getting overridden is wrong, and leaving it in makes every future draft worse.
- Check the arms of the experiments in
strategy.mdwithstanley-stats --experiments. Report counts, call no winner before both arms holdmin_n, and if one arm is starving say so with a number ("open-question endings: 1 of 4; three of this week's drafts should end on a question"). Then write those drafts so the thin arm fills — that is the one place the plugin should let the experiment steer what it writes.
Two changes at most, and none applied without a yes.
Repackage (every two weeks)
From ledger/posts.jsonl, find the top 3 posts by engagement relative to the median. For each, name the shape that worked (a hand-over-the-system how-to; a claim about the future; a story with a number) and find a fresh story that fits the same shape. Draft one, labelled as a repackage so the audit's self-overlap check is expected. Don't re-post the same lesson; re-use the shape.
Orbit (monthly)
From ledger/posts.jsonl, ledger/comments.jsonl, and the recap pulls: the 5–8 people who engaged most often, with role, what they engaged with, and any open thread with them. Suggest at most two actions (reply to an open thread; comment on something they posted this week that fits a lane). This is relationship data; never DM on the user's behalf.
Strategy check-in (monthly)
Show strategy.md goal and lanes against the month's numbers: posts per lane, median per lane, best and worst with the one reason. Propose at most one change (a lane to drop, a cadence to shift, a bet resolved). Ask the user to confirm or edit; then rewrite strategy.md with the date. A lift computed on fewer than 3 posts is not evidence; stanley-stats marks those rows.
Three things the monthly owns that the weekly Tune deliberately does not:
- Close or rotate the experiment. If both arms have 4+ posts, declare the verdict, write it into
## Resolved betswith the numbers and the date, and propose the next experiment. If an experiment has been open for two months without filling its arms, it is the wrong experiment — kill it and pick one that every post can be tagged with. - Bring in something from outside. Self-observation recombines what the user already does; it cannot add a move they have never made. Once a month run
/open-stanley:studyon one person who is winning in one of their lanes (not the same person twice running) and add the shapes to the catalog. Propose one to try next month. - Read the priors before proposing.
skills/stanley/references/research-linkedin-2026.mdholds the external evidence (226K-post study, named creators' claims). An experiment that tests one of its findings on the user's own account is worth more than one invented from scratch, and a finding the user's ledger already contradicts is worth saying out loud. Performance cycles are normal (slumps of ~13 posts, plateaus of ~19): do not propose a lane change on fewer than ~20 posts of evidence, and say "cycle" rather than "the lane is wrong" when the numbers dipped without a change in what was posted. - Find a format gap. From the creators studied this month (the vault's
shapes/andskills/study/references/creators-latest.md), name a format or shape that works in the user's lane and that the user has never tried — a carousel, a short video, a teardown, a build-in-public series, whatever the evidence actually shows. Propose exactly one, as the next experiment with its arms andmin_n, not as a thing to go do. The plugin should arrive at new formats by watching what works and testing it, never by adopting whatever someone mentioned in passing.
Trend alert (on event)
Only when a research or browser source shows something under N hours old (from voice.md freshness_days) in a lane, with a primary source you opened. Deliver: the source link with its date, one line on why it's a fit, a drafted angle (not a post) that ends with "want the draft?". Never post. The user's rule: if he already saw it a week ago, it's not news.
Delivery in remote runs
Each ritual is a scheduled-task template in templates/scheduled-tasks/. In a remote run, deliver through the notify capability the user chose at onboarding (Slack DM, Telegram bot, email draft to self) and end the message with what was checked, so a "nothing today" is visibly a result and not a missed run.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 467k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |