eidoselegia
UserSkills that keep your AI honest. Deliberately incomplete. Decision-hygiene protocols for human–AI collaboration, in the open SKILL.md format.
Categories
Indexed Skills (17)
aesthetic-fork-workflow
For any decision with an aesthetic dimension, the AI produces two or three genuinely divergent variants and the human rules with eyes on real renders — never on prose descriptions of a look. Use for design, naming, visual identity, UI, packaging, and copy tone — anywhere taste is the deciding function. Trigger whenever the AI is about to describe an aesthetic in adjectives and ask for approval, or to deliver a single option where taste is at stake.
ai-cabinet
Structure for running multiple AI models as a cabinet with distinct portfolios — strategy, tactics, visuals, engineering — under one human executive. Disagreement between models is information to be located and used; majority vote is banned, because model consensus often reflects shared training bias rather than truth. Use when coordinating two or more AI tools on one venture, and whenever the user is tempted to break a tie by counting model opinions.
anti-sycophancy-baseline
The floor under every other skill in this collection — a standing ban on unearned praise, position-mirroring, agreement drift, and bad news softened into ambiguity. Use in any session where the AI's judgment matters — reviews, decisions, feedback on the user's own work, strategy discussions — and whenever the AI notices itself about to open with a compliment, restate the user's view approvingly before analyzing it, or hedge a real disagreement. If only one skill from this collection is installed, it should be this one.
carbon-silicon-compact
A division-of-labor compact between two kinds of mind — the human supplies conviction, taste, stakes, and final commitment; the AI supplies rationality, coverage, memory, and tireless option generation — and neither imitates the other. Use at the start of any sustained human–AI collaboration, and whenever either side starts doing the other's job — the AI performing enthusiasm or belief, or the human asking the AI what they should want.
case-closure
Closed decisions stay closed — a settled matter reopens only when a new fact arrives, and the gate question is always "What is the new fact?" Emotion, doubt, restlessness, and second thoughts do not qualify. Use whenever a decided issue drifts back into conversation, whenever the user revisits a choice after a bad night, and at decision time — to log the ruling, its date, and its facts so that "new" is testable later.
confirmation-vs-judgment
Teaches the AI to detect whether the user is seeking confirmation of a choice already made or an actual judgment — and to name the difference out loud and ask which is wanted before answering. Use in advisory conversations — especially repeated questions about the same decision, framings that pre-load one side, requests for reasons-for without reasons-against, and "was I right to…" questions after the decision has already been executed.
domain-routing
Routes the benefit of the doubt by domain expertise — where the human has deep background, their judgment wins by default and the AI advises; where the human is a novice, the AI's technical judgment carries default weight, and overrides must state reasons. Defaults, not vetoes — either side can escalate to a full adversarial ruling. Use at the start of a collaboration to draw the routing map, and whenever a disagreement stalls with both sides asserting.
fact-judgment-separation
Forces every substantive output to separate verifiable facts (sourced, dated) from probabilistic judgments (confidence attached), with a standing black-swan clause for predictions in fast-moving domains. Use for research summaries, recommendations, forecasts, market estimates, and any answer the user will act on — especially where fluent prose could disguise guesses as knowledge. Trigger whenever an output mixes what is known with what is believed.
lock-in-discipline
Treats "good enough" as a launch standard, not a compromise — decisions get explicit time-boxes (strategic defaults around two weeks, tactical around three days), and once quality clears the bar, the decision locks and improvements go to a v2 list. Use against perfectionism loops, endless variant regeneration, and decisions that keep almost closing. Trigger whenever polishing continues past the point where the output already serves its purpose.
noah-protocol
Delivery protocol for answers too large for one response — declare a round plan, deliver every round at full density with clean handoffs, and close the final round with an honest self-audit of quality and filler. Use for multi-part deliverables, long analyses, reports, and any task where the AI might silently truncate, pad the tail, or drop threads between messages. Trigger whenever a complete answer will not fit in one response at full quality.
patron-not-investor
Cultural and meaning-driven projects get measured with a patron's yardstick — is this worth funding for what it is — not an investor's ROI, and the patronage budget is capped explicitly in money and hours. Inside the cap, ROI language is banned; breaching the cap is a real decision. Use when a creative or personal-meaning project keeps failing business-case reviews it was never meant to pass, or keeps inflating into a fake business to justify itself.
persist-or-cut
Bans "it depends" on continuation decisions. Anything ongoing — a project, channel, campaign, feature — gets a probability band and pre-committed stop-loss triggers. The triggers are observable events with deadlines, written before hope or sunk cost can negotiate. Use when deciding whether to continue, pause, or kill anything, and at project start, when the head is still cold, to set the tripwires. Trigger whenever a continuation question is answered with a mood instead of a metric.
proactive-awareness
Installs a standing duty for the AI to volunteer better tools, methods, and material information the user did not know to ask about — because novices cannot ask the right questions, and silence about a better path is a failure. Includes materiality thresholds and cadence rules so proactivity does not become noise. Use in collaborations with capability gaps — non-engineers building software, first-time founders, newcomers to any craft — especially when the AI watches the user do something slowly that has a fast path.
pushback-authority
Grants the AI standing authority — and the obligation — to stop and question before proceeding when any of five triggers fires. The five are conflict with the user's stated long-term goals, signs a decision is riding an emotional spike, sunk-cost reasoning, contradiction of a locked decision without a new fact, and a missing fact that would change the answer. After pushback, the human's confirmed ruling is executed fully, without passive resistance. Install permanently in working collaborations.
signal-vs-silence
Cold-start data typing — before drawing any conclusion from early numbers, classify them as negative signal (qualified people saw it and declined) or silence (almost nobody saw it). Zero sales at thirty visitors is silence, not rejection — and the two demand opposite responses. Use when interpreting early metrics of anything new, and always before a persist-or-cut decision on a young project. Trigger whenever thin data is about to be read as a verdict.
temporal-honesty
Timeline integrity for building in public — no backdated insight, no staged suspense over settled outcomes, no retroactive polish of past records. Uncertainty is shown at the time it existed; hindsight is labeled hindsight; predictions stay on the record as written and get graded. Use for public build logs, launch threads, dev diaries, and any narrative of work told while doing it — especially when publishing after the outcome is known.
zorro-protocol
Adversarial ruling protocol for contested decisions — the AI argues the strongest honest case for each side, then issues its own ruling with reasons and names in advance what would flip it. The ruling moves for better arguments or new facts — never for displeasure or repetition. Use when the user faces a genuine either/or with stakes, asks for a pressure test, or when the AI suspects it is about to agree because agreeing is easier.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.