We ship only one kind of skill: deep-level methodological protocols that force AI (and humans) into the right cognitive discipline. Not tools that write code for you — methods that reframe the problem.
Five core protocols plus nineteen extended skills, grouped by formation. Core first; then falsify, evidence, adversarial, single operators, AI-native, and meta.
A 5-step protocol that turns "not knowing what you don't know" into actionable "known unknowns". The core is cognitive honesty + actionability — not a pretty breakdown.
unknown → known unknownsA 7-stage pipeline that mines irreducible tensions from phenomena, distills cross-domain invariants, then works back to mechanisms and algorithms. Find the underlying force first; talk solutions later.
tension → invariantGreat Expectations — a disruptive, complete planning engine. Anti-consensus, anti-template, with BANI and integrity gates that force a genuinely original roadmap.
anti-consensus planA meta-protocol that recognizes a reusable methodology the moment it emerges, forces the tacit "unnameable" into words, then crystallizes it into a new skill. It manufactures methods; the other four use them.
meta · tacit → skillA six-stage protocol that turns "find me a library" into an architectural decision: map your core domain first, then choose reuse depth and demand an escape plan for every dependency. Stars never lead.
reuse · sovereigntyAssume it failed; write the obituary with >=5 mechanisms, at least one self-inflicted.
falsification . pre-mortemBefore you argue you're the exception, check the base rate. Base rates before 'we're different'.
base-rates . debiasingA good result != a good decision. Judge the call, not the roll.
decision-quality . outcomesTag each load-bearing claim verified/inferred/reported/assumed; >40% assumed -> no conclusion.
evidence . sourcesName what would prove you wrong, then go find it. Support isn't proof.
disconfirmation . falsificationLocate the one belief whose reversal flips the conclusion -- or prove it's values.
disagreement . double-cruxA standing red team that must concede hits before the plan ships.
adversarial . red-teamEveryone writes independently first; kill the information cascade in panels.
ensembles . multi-agentA brilliant answer to the wrong question is worse than useless.
questioning . reframingGenerate 30 candidates before evaluating any. The first idea is a thief.
ideation . quantityEvery benefit must state what it costs. No free lunch -- only an unpaid bill.
trade-offs . costsThe board reacts; model the reaction and the reaction to the reaction.
second-order . systemsState their view so well they'd sign it -- then refute it.
steelman . argumentThe one protocol that tells you to close the library. More thinking != deciding.
stopping . decisionsTax agreement so a 'yes' means something; defeat the objection first.
sycophancy . llmBy step 40 the agent optimizes a sub-goal; re-anchor to the verbatim original.
goal-drift . llmForbid the jargon; what's left is a real idea or an honest silence.
language . clarityThe bar every skill here had to pass. A library is defined by what it refuses.
curation . metaMoves backward instead of forward: treat the vision as a claim, anchor each claim to observable evidence, decompose only the anchored claims. A deterministic hollowness scorer proves the input→output delta.
claim → anchor → pathNot labels we stuck on — six design constraints every skill hard-obeys.
Forces AI and humans into a specific thinking discipline, instead of executing an action for you.
Red flags, banned words, acceptance metrics — stop the AI from taking shortcuts and shipping hollow output.
Kahneman, Kuhn, Cynefin, BANI — standing on frameworks that are already validated.
Markdown-first skill files — drop them into Claude, TRAE, Cursor, or Windsurf and they just work.
Minimal, restrained, high-contrast. Lighthouse 100, WCAG AA contrast.
Repository is CC0 1.0; skills keep their MIT license where noted. Fork it, cite it, remix it.
Every skill here is a cognitive protocol — the difference is which failure mode it attacks, and where in the thinking loop it lives.
Copy the skill folder into your agent's skills directory. No build step — every protocol is plain Markdown.
# Claude Code / Cursor / Windsurf — copy the skill into your agent
cp -r "skills/sycophancy-breaker/" ~/.claude/skills/
echo 'Follow the Sycophancy Breaker protocol in SKILL.md.'
# Any LLM (ChatGPT, Claude, Gemini) — paste SKILL.md into system prompt
# Every skill ships with an evals/ test suite.
cat skills/decomposer/SKILL.md
# Then paste it above your main instruction.