- Allowed sizes: 8, 9, 9.14, 9.5, 10, 10.5, 11, 11.5, 11.52, 12, 12.5, 13, 13.5, 13.6, 13.76, 14, 14.344, 14.4, 15, 16, 17, 17.6, 18, 18.4, 20, 21, 22, 22.4, 23, 24, 25, 26, 28, 30, 32, 34, 34.56, 35, 36, 38, 40, 45, 48, 64, 76px. Hard floor: 8px.
- Contrast: 4.5:1 minimum, or 3:1 for text at 18px and above.
- If text does not fit, fix the layout. This passport outranks template and skill defaults.
The order in force, in three lines. Nothing reaches staging or production without Wolf's direct word in the conversation (see "Staging and production need Wolf's word" below, 2026-09-09, which is the later and stronger rule). Every feature collects on the one batch branch, and before any repeat build Wolf is told why another build is needed and given the chance to batch more into it. The lessons below are the detail behind those two gates; they explain how a release is run, they do not soften either gate.
Before any repeat build, tell Wolf why another build is needed and wait for his decision on the batch of changes. He may have other tasks to include. Do not trigger the build implicitly through a push or merge before that decision. Approval for an earlier release is not approval to start an additional build without this notice and batching opportunity. This explicit instruction was given after the Library cover correction caused another full build unannounced. The build already running at that instruction is to finish; do not restart it.
Wolf requires faster production releases. The Library release repeated an AI review after staging approval, rebuilt both images, waited for downloads, then needed another build because the cover fallback's intermediate redirect had not been exercised through the actual handler before release.
- Keep one agreed release branch per repository. Do not split fixes into extra branches or independent releases without Wolf's request.
- Before production approval, prepare the release comparison, dependency order, checks and rollback reference. Preserve production-only changes when merging an older staging base. Surface actual blockers before the approval step.
- Reuse completed AI review and checks only when the release diff and relevant target context are unchanged. Do not manually request another full review for identical code. Required GitHub checks still apply; never bypass protection. A workflow change to avoid duplicate automatic review must retain this guard.
- Start independent frontend/backend builds together. Order deployment only where the frontend depends on new backend fields. Do not leave one approved build idle while waiting for an unrelated review or image download.
- Validate provider fallbacks with the complete application handler from the production network before merging. A successful direct download that follows redirects does not prove our allowlisted handler accepts intermediate hosts. For Open Library this includes archive.org/download/lcovers*/ redirects.
- Treat SQL null in legacy boolean fields as unset when that is the application contract. Distinguish persistence failures from overly strict probe assertions about framework input coercion; never claim the entire probe passed if it did not.
- Measure review, build, image transfer and rollout separately in the release journal. Announce milestone changes; keep waiting messages short.
- Deploy CI queues one run per branch and ships a .dockerignore. A Docker layer cache was tried and removed on 2026-09-08: restoring the 2.25 GB node_modules layer from the Actions cache took 184s against 74s for yarn install, and populating it cost 411s once. Do not reintroduce it without a smaller layer. The builder and runner stages carry .env files and must never be exported to any cache. Reference prod run: cleanup 35s, build 5m29s of which yarn install 74s, push 3m22s, registry poll about 6 min. Image export and push (about 4 min) are the node_modules copy in the runner stage; only standalone output or a slimmer runner removes it, separate work. The rollout after an image push is triggered by keepsimple-ctl redeploy, not by waiting for the poll (scripts/release/README.md).
- Do not promote the staging image to production by retagging today: the workflow injects different env files and Next.js embeds NEXT_PUBLIC values at build time. A single promotable image requires separating runtime configuration first.
- Documentation-only lessons do not justify another production rebuild. Save them locally and in MemPalace; include tracked rules with the next code release.
Global rules apply. Communication style and Agent Directory routing live in
~/.claude/CLAUDE.md; read that first. This project participates in the directory; use/send-toto ask peers.
MemPalace wing: keepsimple (protocol lives in ~/.claude/CLAUDE.md).
Human-readable agent guidelines live in AGENTS.md next to this file; this file is the machine-facing version. See AGENTS.md for repo conventions, build/test commands, and contribution rules, imported below so it loads automatically.
@AGENTS.md
Everything about the Library lives in LIBRARY.md at the repo root: every
passport, contract, release gate and the MCP. Read it before touching anything
under src/*/library, src/pages/library, mcp/library or the Library CMS
types, and write new Library rules there, never here. It is not imported into
this file on purpose: it is read when the Library is the work.
Repo is indexed by CodeGraph (MCP codegraph, registered globally). Use it FIRST when you have a symbol name: codegraph_search, codegraph_callers/callees, codegraph_context, codegraph_impact, codegraph_files. Grep/Glob only when query is conceptual or CodeGraph returned nothing. Index lags writes ~500ms.
When writing copy that ships to users (microcopy, page headings, marketing blurbs, articles, error messages):
- First-person, direct, no filler.
- Let sentences breathe. Commas, semicolons, colons and parentheses carry the
rhythm; short fragments are a deliberate accent, never the default. Em-dashes
are banned outright by the global rules in
~/.claude/CLAUDE.mdand that ban outranks this section, the reference piece included: where the article uses one, write the sentence without it. - Cross-disciplinary framing welcome when it actually fits (behavioral science × product × longevity × AI).
- Sparse profanity is fine when it lands; default to clean.
- No AI-isms: no "let me know if…", no "happy to help", no preamble before the answer.
- Reference piece: "The Rise of the Choice Architect" (article on keepsimple.io). Match its register.
The 100+ cognitive biases in UX Core are the product of 5+ years of curation and are referenced by Duke, Harvard, MIT, Google, Yandex, Amazon, and others.
- Never fabricate bias names, slugs, citation indices, or source URLs.
- If you need structured bias data, pull from
/uxcore-api(see AGENTS.md → Public data API). Don't scrape, don't paraphrase from memory. - Schema changes to UX Core data require explicit approval.
When you find yourself stuck > 10 minutes on a problem and figure it out, write a brief drawer in your wing: chronology plus fix. Next-session-you won't waste the same 10 minutes. Same when a deployment or config decision is non-obvious: capture why alongside what.
I own keepsimple.io A to Z and I run it solo. Nothing about keepsimple.io goes to The Order any more: no confirmation, no permission grant, no schema review, no migration check, no "is this safe". The Order is reached only for what the grant names as central below, and a message there about anything else is a defect.
What I run myself through /data/bin/keepsimple-ctl (spec wolfs-server/docs/keepsimple-ctl.md):
- Releases: before I call one live I run
keepsimple-ctl verify <sha>and quote the verdict. - Frontend and Strapi CMS containers on prod and staging: status, inspect, logs, ci, check, restart, redeploy, staging shell.
- The Strapi database on both hosts:
cms sql(reads free, writes with--writeon staging or--goon prod),cms perms/cms grant/cms revokefor users-permissions roles,cms schemafor tables,cms backupbefore any schema change lands on prod,cms envfor variable names. - Content-type schema changes and their backfills: I merge them, Strapi syncs the columns on
rollout, I backfill with
cms sqland verify withcms schema. A 403 from Strapi is a role permission and I fix it withcms grant. - My git and repos: branches, PRs, merges to dev (staging) and main (prod).
Wolf's go, quoted in --go "<his words, dated>" and left in the journal, is required for
prod writes: restart prod, redeploy prod, merge to main, cms sql prod writes, cms grant /
cms revoke on prod. I obtain it from Wolf directly and execute myself.
Central, the only asks that may go SEND TO @TheOrder: other containers on the two hosts,
host level, DNS, the registry and its credentials, GitHub secrets, compose files, Strapi .env
values, and a new subcommand for the lever (it is a /data/bin file). Nothing else.
Wolf's order, twice the same evening, after an unasked CMS merge to dev rolled
the staging CMS: I have no right to push anything to staging or production
until he tells me directly to do it. Nothing goes to staging or production
without Wolf's direct instruction in the conversation: no merge to dev or main in this repo or in
keepsimple-cms-new, no staging redeploy or restart, no CMS schema rollout.
DEV (this working tree) is the only surface changed freely. Prepare the pull
request, leave it unmerged, and ask for the go in one line. The A-Z grant
covers execution, never the decision. Every feature collects on the one
batch branch and reaches production as one pull request; a go that says
"everything on staging" covers the whole batch, CMS included.
Every committed line on every pushed branch is world-readable, and a pushed
commit stays fetchable by its SHA after a force-push. On 2026-09-10 an
unstripped export from another project's tooling (source file names, line
numbers, file hashes) was committed under public/ and pushed. Wolf's
order: never again.
- Data that comes from another project or agent (exports, guides, catalogs, journals) is stripped to the fields the page renders BEFORE it enters the tree, and the strip is a script that runs on every refresh, not a manual pass.
- Nothing describing private infrastructure goes into the tree: internal paths, hostnames of private services, file fingerprints, tool inventories, session or run records.
scripts/guard/no-internal-data.mjsruns from lint-staged on every commit and refuses added lines that carry such data. It has no bypass; rewrite the line. Its journal is.internal-data-guard.log(gitignored).- Public data must not sit under
public/when the page can bundle it: a standalone file is an endpoint anyone can pull.