evisions

Record a decision as an ADR (Architecture Decision Record) in docs/decisions/, or as a docs/decision-log.md row when it is below the ADR bar, and supersede an older ADR when a decision changes. Covers any decision worth keeping - technical, process, client or business. Use when the user says "zapiš ADR", "zapiš to jako rozhodnutí", "zapiš rozhodnutí", "write an ADR", "record this decision", "decision record", "měníme rozhodnutí", or when /end writes an approved ADR.

Write a bug entry into the project's BUGS.md - broken behaviour found WHILE working on something else, logged so it is not lost. Used by /end to write every open BUG-TODO line from WORKSTATE without asking, and on an explicit ask to log a bug now ("zapiš bug", "zaloguj tu chybu", "log this bug", "add it to BUGS.md"), or to mark a logged bug fixed. NOT for the bug the current task is about (fix that one), NOT for code that works but is costly or risky (evisions:tech-debt-log), and NOT for style.

Save progress mid-session in seconds - one worklog entry, a refreshed WORKSTATE and a local git commit of the files this session touched, never a push. Use when the user types /checkpoint or says "checkpoint", "ulož postup", "ulož stav", "zapiš checkpoint", "save progress", and after a finished logical step in a longer session.

Close a work session - first propose the session's decisions (decision-log rows and ADR candidates) for one quick approval, then write the worklog, WORKSTATE, approved decisions and the captured bugs and tech debt, sync feature docs, commit locally and ask before any push. Use when the user types /end or says "končíme", "konec", "uzavři session", "zapiš to a končíme", "wrap up", "end the session".

Interview the user, one block of questions at a time, to a PRD (product requirements document) that covers the business side (problem, goals, users, scope, acceptance criteria) AND the technical dimensions a non-developer does not know to ask about - GDPR and EU AI Act triage, data classification, authentication, backup and disaster recovery, and a STRIDE threat model. The interview runs in the user's language; the finished PRD is written in English to docs/prd/<slug>-prd.md. Use when the user wants a PRD, requirements or a product spec for something that will be built ("napiš PRD", "sepiš zadání", "potřebuju požadavky na appku", "write a PRD", "product requirements", "spec this app"). NOT for HOW to build it (architecture, implementation plan) and NOT for developing an idea that is not decided yet (use evisions:socratic-brainstormer).

Runs the sanity check, a static hygiene check by a bundled script that calls no model, on a prompt, system prompt, agent file or SKILL.md, and explains each finding in plain words (a pasted API key or password, no output-format or fallback keyword, template placeholders not wrapped in a tag or not declared as data, an XML section left unclosed). Use when the user wants a prompt checked before it is used or shared ("zkontroluj prompt", "sanity check promptu", "ohodnoť prompt", "je ten prompt v pořádku", "eval promptu", "prompt eval"), and whenever the evisions:prompt-engineer agent has saved an artifact. It is never a quality verdict. To write or improve a prompt, use the evisions:prompt-engineer agent instead.

Research into how practitioners actually do something, or into whether a claim is true, delivered with sources the user can click. Owns the tier choice (`evisions:research-analyst` for one angle, `evisions:research-lead` for several), the honesty standard every finding is reported under, source verification, and where the research file is saved. Use when the user wants something outside the current project researched, investigated, compared or fact-checked - "udělej research na X", "prozkoumej", "zjisti, jak to dělají ostatní", "co lidi používají na X", "je pravda, že...", "porovnej A a B", "research this". NOT for finding things in the project's own files (search them directly or use the Explore agent), and NOT for a question the conversation or the code at hand already answers.

Check a third-party AI agent skill or plugin BEFORE installing, enabling or updating it - a SKILL.md, a skill folder, a plugin, a hook or MCP bundle, an agent or command pack, or a .skill/.zip/.tar archive - with a bundled static scanner plus a read-only review, and return BLOCK, REVIEW or PERMIT WITHIN COVERAGE with the evidence. Also compares a new version against an earlier scan. Use when the user asks whether a skill or plugin is safe to install ("je ten skill bezpečný?", "můžu si to nainstalovat?", "zkontroluj ten plugin", "is this skill safe", "check this plugin before I install it"). It never installs, runs or imports the target. NOT a security review of the user's own application code (use evisions:security-auditor).

You MUST use this FIRST when the user is thinking through their OWN idea, direction, or business, content or client call rather than handing over a task - they float an idea, weigh a direction, or ask for your view - "potřebuju si to promyslet", "co myslíš", "rozviň mi nápad", "mám nápad", "pojď to se mnou probrat", "let me think this through". Socratic sparring that develops the user's own thinking instead of handing them a finished answer - a new idea, a product or service, a campaign or content concept, a positioning or business call, a client approach, a methodology, any half-formed plan of theirs. Takes precedence over general brainstorming skills - superpowers:brainstorming, when installed, is for designing a software build heading to implementation, and when both could apply this skill runs first. NOT for a task already assigned (just do it), NOT for being questioned on a formed plan (a grilling skill if one is installed), NOT for writing down a finished thought.

Write a tech-debt entry into the project's TECH-DEBT.md - code that WORKS but is costly or risky (duplication, a workaround, fragile coupling, missing tests, dead code, slow but correct), noticed WHILE working on something else. Used by /end to write every open DEBT-TODO line from WORKSTATE without asking, and on an explicit ask to log debt now ("zapiš tech debt", "zaloguj technický dluh", "log this tech debt", "add it to TECH-DEBT.md"), or to mark logged debt paid. NOT for broken behaviour (evisions:bug-log) and NOT for lint-level style.