Coding Higher Mind: the AI system behind the work
Two AI tools, hardened prompts, automatic checks at every step, AI reviewers split by speciality, and a loop that learns from failures. Counted, not estimated. Now open source.
The skills I use every day, open source. Each one is a folder with a SKILL.md: Claude reads the description and loads it when a task matches. The cards are a snapshot of the public repository's skills.json: the list comes from the skill files themselves, and the short descriptions are written in both languages next to them.
Install
Just the skills, as a Claude Code plugin (no hooks, no settings changes):
claude plugin marketplace add kamiljan11/coding-higher-mind
claude plugin install higher-mind-skills@coding-higher-mindThe whole system with hooks, gates and reviewer agents:
git clone https://github.com/kamiljan11/coding-higher-mind
cd coding-higher-mind
node install.mjs --dry-runOne skill only: copy its folder from the repository into ~/.claude/skills/.
Architecture and decisions
pg-council
A group decision as a protocol, not a chat: each department gathers facts and positions independently, a devil's advocate is mandatory, a script aggregates, and the decision record (ADR) keeps the dissent.
architecture-advisor
Architecture decisions through trade-offs and real context, not hype: choosing an architecture for a new project, auditing an existing one and proposing changes.
divergent-thinking
Instead of the first solution that works: several options, free and creative alternatives, then a choice checked against the rest.
diagrams
Creates and updates architecture diagrams whenever a system or a flow is being designed, discussed or built.
Review and quality
pg-review
Code review the way a software house does it: the risk tier comes from the diff, several specialist reviewers check the change in parallel, a script merges their findings and a verifier tries to disprove them.
anti-sycophancy
A counterweight to agreeing too easily: Claude checks before it agrees, names false premises and never reports success without evidence.
verify-audit
A verification protocol that treats a claim as false until the source proves it true.
Debugging
systematic-debugging
Four-phase debugging down to the root cause, instead of guessing and patching symptoms.
Agent loops
ultra-loop
An “ask once, finish the job” loop: a definition of done, then cycles of implement, verify and score until every dimension meets its threshold.
gauntlet-build
A multi-agent build: the spec split into units, a builder and a critic with fresh context, deterministic gates and a verifier, until the bar is met.
caveman-code
Caveman mode for coding work: replies and subagent prompts cut to the fewest tokens without losing meaning.
model-router
Picks the right Claude model (Opus, Sonnet or Haiku) for each task, so limits aren't spent on work that doesn't need them.
rate-limit-auto-recovery-templates
Adds rate-limit recovery templates, with checkpoints and resume logic, to task prompts.
Frontend
landing-design-patterns
A catalogue of proven effects for B2B landing pages (Next.js + Tailwind 4) with ready code: buttons, menus, switchable panels, cards and backgrounds.
mobile-optimization
Audits a site's mobile version with real measurements (touch targets, font sizes, Core Web Vitals, overflow, images) and applies fixes, including app-like patterns.
web-bajery
An encyclopedia of about 230 web effects: the name, a search phrase, the library and a mobile rating for each one.
Writing
humanizer
A pass over every text that goes out (posts, emails, articles, website copy) so it reads like a person wrote it, not a model.
book-typesetting
Turns raw text or markdown into a professionally typeset, navigable book PDF without changing a word of the author's content.
Reviewer agents
Read-only reviewers with a fresh context, one speciality each. They are part of the full install and run through the pg-review and pg-council skills.
catfish
The devil's advocate in a pg-council decision: finds the strongest argument against the leading option and the status quo, plus facts nobody raised.
code-reviewer
Reviews a change for readability, correctness, tests and observability, read-only and with a fresh context.
data-reviewer
Migrations, schema, indexes and constraints, RLS coverage of new tables, N+1 queries, transactions, backward compatibility and restore.
landscape-reader
Reads untrusted content from other people's repositories with web access only: no files and no shell, so hostile text has nothing to reach secrets with.
ops-reviewer
Deploy path, environment flags, timeouts and retries, logs with context but no personal data, rollback, cost, and whether CI gates really run.
product-reviewer
Whether a change solves the named problem, stays in scope, has a metric and can be handed over to a stranger; what the client will change should be data, not code.
qa-reviewer
Runs the app instead of reading only the diff: replays the acceptance criterion by clicking and checks critical paths on parallel, isolated instances.
security-reviewer
STRIDE, OWASP ASVS and lessons from past incidents (RLS, IDOR, race conditions, secrets) for changes to auth, payments, migrations and webhooks.
ux-reviewer
UI states, wording in three languages, currencies and dates, accessibility, mobile, and confirmations for irreversible actions.
verifier
After aggregation it tries to disprove each finding: reruns the evidence command and returns reproduced, not reproduced or cannot run.