Eight critique skills. One router. Zero generic reviews.
ACIDMIND
Stop asking AI coding agents for generic reviews. AcidMind routes every critique to the one specialist that can answer the actual question — then hands it back with a hardening plan.
The skill family
Pick the lens, not the vibe.
Each specialist owns a separate critique axis, cites the rule that produced the finding, and stays read-only unless you explicitly ask for implementation.
review request --> decision tree --> skills/ruthless-critic/SKILL.md
persona: The acid test — three registers: ROAST, AUTOPSY, HONEST.
question: Where does this artifact break, specifically?
scope: Default lens for any review, argument, or plan. Whole-repo autopsy on demand.
finding
[RC-04] Vague praise is not a finding.
Gate: FIX FIRST
§ mechanism
One router. Exactly one lens per request.
AcidMind.md holds no rules — only a decision tree that points your agent at exactly one skills/<name>/SKILL.md. Skills load on demand, so a review costs one skill's worth of context, not eight.
Skills diagnose and prescribe fixes. None rewrites unless you ask.
Each skill names what it does not cover and which sibling does.
Technique is fine. Technique without a purpose is a finding.
Live router
Type it. Watch it route.
The same signal matching the decision tree uses — deterministic, no model call. Every routing is written to Postgres and replayed below.
review request --> decision tree --> awaiting input
the router picks exactly one lens, never a committee.
Gate: _
Nothing routed yet. Be the first.
Quick start
Run from GitHub. Start small.
No clone, no global install, no npm account. The default core edition installs the daily lenses plus the automatic secondthought critic.
npx github:ragajatsuma-cmd/acidmind init
Ask for a review, roast, pentest, feature audit, or full panel.
Receive findings with cited rules and one mechanical gate.
npx github:ragajatsuma-cmd/acidmind init --edition security npx github:ragajatsuma-cmd/acidmind init --all npx github:ragajatsuma-cmd/acidmind add security-critic
ruthless · design · feature · badass · secondthought
Daily review
core + security-critic
Production teams
all eight skills
Pre-launch audits
Why it works
Scoped critics beat soft paragraphs.
Read-only by default. Skills diagnose and prescribe fixes. They do not rewrite code unless you ask.
Scoped, not overlapping. The router picks one lens, so eight installed skills do not become eight duplicate opinions.
Purpose-gated, not style-gated. Technique is allowed when it has a reason. Technique without purpose becomes a finding.