Most Claude Code skill repos are catalogs. This one is a system.
Every skill knows what comes after it. Judgment is protected from sycophancy by hard rules rather than good intentions. Ambiguous requests get decoded before a single file is touched. The graph that connects all of it is maintained by one of its own nodes.
/plugin marketplace add NickAme03/claude-second-brain
/plugin install second-brain-skills@claude-second-brain
A catalog is a folder you have to remember to open. A system routes you through itself. The distinction is not stylistic: it decides whether a session ends somewhere useful or ends nowhere.
The rule every skill obeys: finish by offering choosable consequences derived from the actual output, condition first and tool second, with an explicit exit. When a destination does not exist yet, it is recorded rather than silently dropped.
Every skill in this repo was admitted only if it passed all three.
The system works without constant intervention. Skills update themselves, registries track state, nothing depends on you remembering.
Skills are organs. They are born, they connect, they specialize. An isolated skill is dead tissue.
Every addition must increase the system's capacity to grow on its own, not just add one more file to a pile.
They work independently. They work better together, because the handoff sections reference each other. Start with the first two: they are the immune system.
decodifica-intentostart hereDecodes every ambiguous request before acting. Classifies intent, decides whether to ask or infer, and always declares the reading it chose. The most expensive moment of a session is an action launched from a misread request.
anti-compiacimentostart hereTotal ban on complacent lying. No unverified confirmations, no inflated praise, no minimized problems, no position changes under pressure without new facts. Every judgment split into what works and what does not, labeled as verified fact, opinion, or hypothesis.
verifica-artefattiartifact counterpartBefore any text reaches third parties, every named file, script, test or dataset is resolved against what the reader can reach. Three outcomes only: reachable, existing but unreachable (marked as a gray link, with the reason), or removed. Measures answer the other half of the rule: the command that regenerates them, or an explicit label saying they cannot be reproduced. A finished artifact that lives outside the repository is, for the reader, indistinguishable from one that never existed.
skill-linkerTurns every skill into a graph node: a standard "next skills" section, an edge registry, one skill per update. The reason this repo is a system and not a folder.
risparmio-tokenToken economy in three modes: operating rules during work, setup audit across CLAUDE.md and MCP and skills, and design of token-efficient artifacts.
orchestratore-pianificaTurns a concrete objective into an executable task sequence in PIANO.md.
orchestratore-eseguiExecutes tasks from the plan, resuming exactly where the last session stopped.
orchestratore-ripianificaUpdates a plan incrementally when reality diverges. Never rewrites from scratch.
orchestratore-cruscottoRead-only control tower over all active plans. Answers one question: where is the human actually needed?
percorsi-di-sviluppoTurns vague ambitions into 2 to 4 comparable paths built on a real profile. Each with structural cons that do not disappear with effort, the first three steps, and an explicit abandonment signal.
analista-progettiReviews a project, or a whole portfolio read as a network, the way whoever judges it next would. Three lenses, seven scored dimensions, ruthless prioritization, every criticism shipped with a fix.
cacciatore-skillThe procurement organ. Hunts and evaluates new skills against the three gates. Reports only, never installs.
cacciatore-citazioni-apaFinds real, verifiable academic citations through PubMed and Scholar, formats APA 7, extracts verbatim quotes. Absolute fabrication ban.
bussolaCold-start dispatcher for the "I don't even know where to begin" moment. Returns one direction plus a first concrete move, then hands off. A compass, not a planner.
Even if you never run a single skill from this repo, these transfer to any Claude Code setup. Take them and leave the rest.
End every process with 2 to 4 options derived from actual outputs, not from a generic menu. "If the report reveals stalled projects, go to project management." Condition first, tool second, and always an explicit way to stop.
A permanent memory rule points at the decoding skill, which guarantees the hook fires even when the ambiguous request resembles no skill description at all. Memory guarantees the trigger; the skill contains the method. Splitting those two is the whole trick.
"Be honest" does not survive contact with a user who insists. Hard prohibitions with a declared rationale do, and the rationale matters: insistence is data about mood, not about the object.
The graph maintainer is itself a skill, with non-negotiable rules: one skill at a time, diff before writing, registry updated in the same task. A system that cannot repair itself is documentation, not infrastructure.