minh-toolkit

Opt-in only — do not trigger automatically from general conversation about process, CLAUDE.md, or multi-agent architecture. Invoke only when the user explicitly runs /bootstrap-eagd-pattern or directly asks by name (e.g. "set up EAGD for this repo", "give this repo a mechanism to spawn an advisor agent"). Installs a standing, repo-wide mechanism — live instructions in the project's anchor doc — that let any future agent session in this repo autonomously spawn an Advise/Grade/Dream role agent (via the harness's sub-agent tool, with a probe-verified named model per tool) when it hits a matching trigger, without the human setting it up again. Also handles re-runs, where existing model bindings are read and kept rather than re-asked. Runs once per setup or re-calibration; does not run continuously.

Opt-in only — do not trigger automatically from general conversation about process, CLAUDE.md, or project setup. Invoke only when the user explicitly runs /bootstrap-way-of-working or directly asks by name. Sets up a CLAUDE.md/AGENTS.md and companion docs (bug log, decision log, retrospective log, task ledger) encoding a disciplined way of working — task approval via Definition-of-Done + Test Plan, bug/decision/retro logging with distinct triggers for each, and a proportionality check against overengineering — calibrated to the target project's actual size, team, and risk profile rather than transplanted wholesale from the source project.

Coaches the user to turn a vague goal into working code themselves — the agent reviews but never writes the implementation. Use when the user is learning a new language, framework, or concept and says things like "I don't know where to start," "my mind goes blank," "I don't know how to begin," "give me an exercise," or wants hands-on practice rather than a delivered solution. Also use for step-by-step learning exercises, building coding intuition, or references to "the unblock skill" / "the goal-to-code skill" from a prior session. Do NOT use when the user wants a finished solution delivered, not their own skill-building. Scope: one function or single-sitting problem, not a multi-week curriculum — for a full roadmap prefer learn-technology-by-building, and hand off to this skill only when the learner freezes on one specific line.

Mentors the user through learning a new programming language, framework, platform, developer tool, system design/architecture topic, or other technology by building one cumulative project and visualizing structural/runtime knowledge with diagrams. Use when the user asks for a learning plan, starts/continues a hands-on project, says "next", "continue", or "done" during guided study, wants progress tracked, needs a course/video mapped to a practical gap, wants architecture-diagram practice, or wants a concept explained within their project. Preserves the user's coding-first, weekly-checkpoint, copy-disturb-transfer-and-visualize process across subjects. Scope: the multi-session arc, not one-line freezes — hand off to goal-to-code-unblock when the learner blanks on one line, then resume here. Technical/software learning only, not language, finance, or instrument skills.

Runs a short, near-passive learning session for a user whose brain is tired, overloaded, or unable to think, so the learning habit survives without demanding generation effort. Use when the user says things like "my brain is fried", "too tired to code", "can't think today", "I feel mindless", "brain is mush", "no energy but I don't want to skip learning", "give me something easy", or wants to keep a streak alive on a low-energy day. Also use, even without an explicit ask, when the user is mid-technology and reports fatigue, burnout, or overwhelm. Fully standalone; cooperates with learn-technology-by-building and goal-to-code-unblock if installed. Do NOT use when the user has normal energy and wants to practice or build — active learning or direct help fits better.

Create realistic, execution-ready personal plans or harden plans produced by other skills without overriding their domain logic. Use when the user asks to plan, schedule, roadmap, organize, execute, or make a goal actionable; when another planning skill needs feasibility, capacity, buffer, risk, fallback, rollback, or review support; or when a proposed plan looks vague, overloaded, calendar-driven, or falsely precise. Operate in enhancer mode when another domain skill owns the plan, and standalone mode otherwise. Prefer evidence, dependency-aware sequencing, conservative commitments, explicit uncertainty, and adaptive replanning over wishlists or rigid schedules.

Help build a personal brand — positioning, content strategy, writing posts, choosing channels, measuring progress — with the explicit goal of producing real-world outcomes (job opportunities, inbound leads, recognition that converts into something), not the illusion of a brand made of vanity metrics. Use whenever the user wants to build a personal brand, grow a professional presence, plan content strategy for LinkedIn/X/other platforms, write posts about their work or projects for that purpose, or asks to evaluate whether their branding effort is actually working. Also use when the user wants to turn a project or work experience into public content as part of building recognition.

Create and run thoughtful quizzes for almost any subject from a simple request such as “Test me on SQL” or “Help me prepare for an interview.” Use when someone wants a quiz, practice test, interview practice, or questions that check real understanding. Work out the likely level, important topics, difficulty, common mistakes, and question mix. Say what you assumed, ask only when the topic is truly unclear, adjust the quiz based on answers, check facts when needed, and review every question so there is one clearly best answer without obvious guessing clues.

Runs a design or strategy decision through a lead-with-question loop instead of answering directly — the user proposes a rough or "crazy" idea first, gets probed with follow-up questions to stress-test it themselves, and only then receives the agent's direct feedback or missing information. Covers both technical design decisions (architecture, data model, which tool to use) and non-technical strategy decisions (positioning, USP, product direction). OPT-IN ONLY — do not trigger this automatically just because a design or strategy question appears. Use this only when the user explicitly asks for it by name ("dùng socratic-brainstorm", "brainstorm kiểu Socratic", "hỏi ngược tôi trước", "dùng skill hỏi-đáp") or explicitly says they want to be questioned before getting an answer. If the user asks a design question without invoking this skill, just answer normally — do not silently start interrogating them.

Makes any prose the agent writes read like a real person wrote it, not an AI — the agent's own output first, the user's pasted drafts second. Apply this by default before delivering any substantial piece of writing on the user's behalf: emails, LinkedIn posts, blog drafts, reports, proposals, messages, summaries with real prose in them. Also use whenever the user shares their own text and asks for it to sound more human or natural — triggers include "too AI," "too ChatGPT," "too robotic," "sounds generic," "humanize this," "make it sound like a real person," "give it more voice," or any request to edit a draft for authenticity and tone. Do not skip this just because the user didn't ask for it by name — AI-sounding prose is a default failure mode to actively avoid, not an opt-in fix.

Coach the user to genuinely improve their own writing ability through deliberate practice — not to produce a finished piece for them. Use when the user says they want to practice writing, get better at writing, are stuck or blocked on turning real material into prose, feel overwhelmed by open questions about their own experience, or explicitly ask to train their writing skill. Also use when the user has raw facts or a story they can recall but can't shape into sentences. Works for any kind of writing and any domain — technical, creative, business, academic, personal — the skill being built is writing itself, not writing about any one subject.

Evaluate a YouTube video with critical thinking to decide if it is worth the viewer's time. Use whenever the user shares a youtube.com or youtu.be link and asks things like "is this worth watching", "should I watch this", "đánh giá video này", "video này có đáng xem không", "review video giúp tôi", or asks for a critical/objective opinion on a YouTube video. Also trigger when the user asks to judge, rate, or assess the value of a YouTube video, even without using the word "evaluate". Requires the youtube-mcp-cli tools (https://github.com/johncegom/go-youtube-mcp-cli) for metadata and transcript access — check for these tools before starting.