Never re-fight a settled battle. Discipline for recording and RETRIEVING past failures — gotchas, incidents, postmortems, dead ends, reverted fixes — as symptom → root cause → evidence → status. Use before debugging anything that smells familiar, when an incident closes, when a fix gets reverted, when an approach is abandoned, or when asked to document a bug just solved, write a postmortem, do failure archaeology, incident memory, or "did we hit this before". Not for capturing general knowledge (use brain_capture / brain_write) or live task tracking.
Write research prompts that come back decision-ready — zero-prior-knowledge framing, evidence hierarchy, fact/inference separation, and a mandatory gap round. Use when formulating a research prompt, deep research request, competitive research, market research, due diligence question, or when delegating research to exa_search, an agent, Codex, or an external researcher. Not for running the research itself end-to-end (deep-research, autoresearch, gemini-deep-research, exa_search EXECUTE research — this skill writes the brief that feeds them), not for the reference deployment strategic deliberation (use council/punta-de-flecha), and not for Brain lookups (brain_search first, always).
Transfer full working context to a fresh agent session with zero memory — end of session, before compaction, switching models/agents mid-project, or pausing multi-day work. Use when asked for a handoff, handoff doc, session transfer, "deja esto listo para la próxima sesión", context dump for another agent, or when a long session is about to hit context limits. Not for delegating a bounded task (use codex-handoff) or for task cards (use handoff-task / brain_task tools).
Evaluate whether a the reference deployment skill actually works — blind runner + separate judge against a bar, rubric scoring, pass-rate on re-runs, and defect diagnosis. Use when asked to evaluate a skill, eval skills, run a blind eval, score against a rubric, test a skill, audit skill quality, grade a skill, review the skill library, or before promoting a new skill to the company master prompt or the whole swarm. Not for writing skills (use skill-authoring-standards) or deciding which to create (use suggest-skills).
Turn a big or ambiguous technical task into a spec of independently verifiable slices, implement slice-by-slice with verification gates, and close the spec with a claim audit. Use when asked to write a spec, implement a spec, plan a build, break down a feature, structure a multi-day or multi-week technical project or migration, or when a Codex/agent delegation is too big for one self-contained prompt. Not for small tasks (single codex-handoff prompt suffices) or non-technical plans.
Verify visual work with evidence instead of vibes — screenshot critique by a fresh judge, before/after comparison, crops of the disputed detail, and a distance-not-verdict mindset. Use when reviewing a UI, landing, email template, dashboard, canvas, or design change; when doing a screenshot critique, comparing screenshots, doing visual QA or UI QA on a landing page, checking a clone against a reference, or before declaring any visual deliverable done. Not for choosing a design direction or running a dedicated accessibility audit.
Cruza foot traffic Store A/Store B con Meta geo spend, GA4 city sessions y Shopify POS. Estima halo effect por tienda y recomienda redistribucion spend.
the reference deployment standard for writing and editing canonical skills (SKILL.md) in the Brain. Use whenever you create a skill, write a skill, edit a skill, review a skill draft, or convert a runbook/gotcha/process into a skill. Covers frontmatter, trigger quality, description keywords, structure, when-not-to-use, provenance, and the failure-mode taxonomy. Not for deciding WHICH skills to create (use suggest-skills) or for grading existing skills (use skill-evaluation).