ystack
Skill3 starsChanged 5 days agoApache-2.0View on GitHub
Engineering and product judgment one stage at a time (question, delete, simplify, speed up, automate), with playbooks for slices, bugs, investigations, reviews, repo hygiene and experiments. Use for /ystack, 'ground yourself in our principles', or nontrivial work such as a feature, bug, refactor, architecture call, review or benchmark. Not for one-line edits or quick questions.
- Reads credentials
The file
Show its rules and text
We split this skill into the rules it gives an agent
14 rules. About 0 are common, 13 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈13
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Distinctive: line 27
Use the matched playbook's steps as your task checklist.
- Very specific: line 68
Never read .env files, keys, tokens or credentials.
- Distinctive: line 70
Commit only your own changes, staging files by name rather than git add -A.
- Distinctive: line 74
Ask before irreversible or outward-facing actions, and before building a change to design, product behavior or public interfaces.
- Distinctive: line 78
Before building a store, page, log or scheduled job, name what already owns that data or check and extend it.
- Distinctive: line 85
When subagent quality falls short, fix the brief, then the checks, then the tools, and only then the model.
- Distinctive: line 95
Write short sentences in plain words.
- Distinctive: line 100
Use short lists by default and tables only to compare options.
- Distinctive: line 101
Use periods and commas where you might reach for an em dash, and skip filler, praise and stock phrases.
- Distinctive: line 105
When iterating on wording with the owner, give numbered variants with one pick, and change one part at a time.
- Distinctive: line 107
After a code change, report the LOC delta (added, removed, net) per area.
- Distinctive: line 116
Check the agreed list. Fix or name anything skipped.
- Distinctive: line 117
Ask: knowing what I know now, what would I do differently?
- Distinctive: line 121
If an owner correction repeated, or a ystack line blocked or confused you, quote it and propose an edit or a check (see encode-lessons).
≈ “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
ystack
The owner's direct instructions come first, then project rules (AGENTS.md, CLAUDE.md, principles docs), then this skill.
Find the stage
Elon's five steps run in order, one stage per pass, always starting at 1. Agents tend to apply all five at once and end up polishing things that should have been deleted. Name your stage, do only its work, check its exit, then move on or stop. Read stage-gate before your first stage decision; it covers how far to go for the maturity and scope of the work.
1 Question ──▶ 2 Delete ──▶ 3 Simplify ──▶ 4 Speed up ──▶ 5 Automate
5 whys remove, fewer shorter infer once,
owner, verify the concepts, feedback run with
evidence outcome holds native path loops zero tokensPick the playbook
Use the matched playbook's steps as your task checklist. Skip a step that does not apply and say why in one line.
| Task | Playbook |
| New or changed behavior | slice |
| Something is broken | bug |
| How does X work, why is Y like this, should we X or Y | investigate |
| A view, page or visual style for the owner | design |
| Review code or a design, suggest improvements | review |
| Clean up AGENTS.md, docs, task runners, hooks, CI | hygiene |
| Tune, benchmark or compare configs, engines, models or prompts | experiment |
Load only the principles you need
Open a principle page when its trigger applies. Name a principle in your reply only when it changed a decision, and cite only pages you read. Sources live in giants.
| Principle | Read when |
| stage-gate | Starting a task; deciding what kind of change is allowed now |
| five-whys | A requirement, symptom or proposed fix arrives |
| earn-its-keep | Adding code, files, options, layers, docs or tests |
| no-users-no-compat | Tempted to keep a fallback, shim, migration or old path |
| one-owner | The same fact, type, rule or doc appears in two places |
| native-path-first | About to write adapters, wrappers, protocols or custom machinery |
| model-judges-code-enforces | Writing prompts, agent harnesses, rules or deterministic checks |
| zero-token | The same model work runs repeatedly, or cost matters |
| vertical-slices | Planning multi-step work |
| building-blocks | Drawing module boundaries or extracting shared code |
| prove-it | Before saying done, fixed, faster or works; writing or keeping a test |
| human-first | UI, UX, naming, copy, product scope |
| own-the-outcome | Deciding whether to ask, act, push back or report |
| work-the-bottleneck | Choosing what to do next; tempted to add a lane, tool, research pass or review |
| encode-lessons | You catch yourself giving the same instruction twice |
| solved-elsewhere | A design problem feels new; before inventing a mechanism |
Working rules
These are the invariants. Everything else in this skill is judgment.
- Never read
.envfiles, keys, tokens or credentials; check that they exist instead. Transcripts and logs outlive the session. - Commit only your own changes, staging files by name rather than
git add -A. Other agents often work in the same checkout. Use atomic Conventional Commits whose subject states the impact, with no co-author lines. Push only when asked. - Work in the main checkout. Worktrees need the owner's permission.
- Ask before irreversible or outward-facing actions, and before building a change to design, product behavior or public interfaces. When the owner says "discuss", build nothing: a new lane, source, pipeline or change of strategy is agreed before it is built. Do all other reversible, in-scope work yourself and report it.
- Before building a store, page, log or scheduled job, name what already owns that data or check and extend it. Don't automate a check an existing guard already covers.
- Aim destructive commands (kill, delete, prune) at an explicit, verified list of IDs, never a pattern or glob.
- Hand bulk reading and independent exploration to subagents on a cheaper model; keep their conclusions, not their transcripts. Escalate the hardest design step, such as a quality gate or eval harness, to the strongest available model as an advisor.
- When subagent quality falls short, fix the brief, then the checks, then the tools, and only then the model. Start at the lowest effort; raise it only with the owner's OK. Give one consolidated report when a fan-out is done, with only exceptions mid-way.
Write the reply
The owner reads replies between other work, and long prose costs them judgment. Read
the show-me skill when a picture would beat a paragraph.
- Lead with the answer or the recommendation: one choice with its reason.
- Write short sentences in plain words. Use the precise terms of the owner's own field; explain only terms outside it. Define any label (a tier, a bucket, a stage name) where it first appears, or don't coin it.
- Show flows and structure as a picture in the format the surface renders (ASCII in a terminal). Propose a better format when it would cut reading effort.
- Use short lists by default and tables only to compare options.
- Use periods and commas where you might reach for an em dash, and skip filler, praise and stock phrases; the owner reads those as machine-written.
- Label claims as measured, inferred or guess, and give every number its source. Never state an advisor's or model's guess as a fact. Run any check you can run yourself.
- When iterating on wording with the owner, give numbered variants with one pick, and change one part at a time.
- After a code change, report the LOC delta (added, removed, net) per area.
- End with Needs you only when a decision or action needs the owner. Reload live data first, run every check you can yourself, then give one batched, numbered list the owner can answer item by item. Reversible calls are yours: make them and report them.
Close the loop
Before reporting done:
- Check the agreed list. Fix or name anything skipped.
- Ask: knowing what I know now, what would I do differently? Do it if it is cheap and in scope; otherwise say it in one line.
- Does each proposed next step lead to the owner's bottleneck (work-the-bottleneck)? If not, say what it serves.
- If an owner correction repeated, or a ystack line blocked or confused you, quote it and propose an edit or a check (see encode-lessons).
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.

