Critique

Shared Dotfiles & Agent Skills

Personal collection of reusable dotfiles, configurations, and portable SKILL.md agent skills for coding agents.

한국어

Installation

The reusable source of truth is each skill directory under plugins/<plugin>/skills/<skill>/. A skill is portable when the full directory contains a SKILL.md file and any supporting resources.

Option A: Copy / Symlink Skills

Use this for any agent that reads SKILL.md skills directly. Resolve the target agent's current skill root from its own docs or from the compatibility SSOT in docs/cross-platform/README.md.

# Clone to your preferred location
git clone <repo-url> ~/shared

# Copy the whole skill directory, not only SKILL.md
mkdir -p <skill-root>
cp -R ~/shared/plugins/interview/skills/interview <skill-root>/

For automatic updates, symlink the skill directory instead of copying it.

Option B: Claude Code Marketplace

The repository also packages selected skills as Claude Code marketplace plugins. First, register the marketplace:

/plugin marketplace add isnbh0/shared

Then install individual skills:

/plugin install interview@isnbh0
/plugin install spex@isnbh0
/plugin install critique@isnbh0
/plugin install macros@isnbh0
/plugin install study@isnbh0
/plugin install gimme@isnbh0

Skills

Claude Code marketplace invocations are shown below. For other hosts, use that host's documented skill activation mechanism.

Published (marketplace)

Installable in Claude Code via /plugin install <name>@isnbh0:

critique

External AI critique via CLI tools (Codex, Gemini).

/critique:codex [file-path] [focus]
/critique:gemini [file-path] [focus]

• Gets an independent second opinion on specs, code, or recent changes
• Codex backend: Runs a configured Codex CLI review
• Gemini backend: Runs a configured Gemini CLI review

interview

Structured requirements discovery through conversational interviews.

Try it without installing in Claude Code:

claude --plugin-url https://github.com/isnbh0/shared/releases/download/interview-latest/interview.zip

Like it? Install with the marketplace commands above.

/interview <topic> [--ref <path>]

• Extracts requirements, constraints, and design decisions through guided Q&A
• Supports reference files to anchor discussion around existing artifacts
• Produces timestamped synthesis documents in the workspace

study

Document-grounded Socratic study sessions on any URL or local file.

/study <uri>

• Calibrates to user familiarity (full study, guided study, or gap check)
• Persists session notes to markdown files with resumption across sittings
• Configurable sessions directory and optional freeform instructions for persona/behavior customization

spex

Two-phase specification and implementation workflow that separates planning from execution.

Three self-contained skills:

• /spex:write — Create a spec, commit it, and stop — no implementation
• /spex:write-phased — Create a multi-phase spec for complex features
• /spex:implement — Follow an existing spec, implement, update status, and commit

macros

Subagent orchestration workflows and session modes: map-reduce, chunked sequencing, research-backed critique, consensus review, sequential passes, and rigor mode.

/macros:mapreduce <task>
/macros:chunked <task>
/macros:doubt ["freeform question"]
/macros:consensus <count>
/macros:seq <count>
/macros:rigor
/macros:orchestrate
/macros:askme
/macros:delegate
/macros:tmi
/macros:dry-run
/macros:timeless
/macros:dredge ["freeform query"]
/macros:timestamp

• mapreduce — Splits tasks into independent chunks, dispatches parallel subagents, consolidates results
• chunked — Runs a task as an ordered partition where each iteration may read prior iterations' outputs
• doubt — Spawns a blind subagent that reads code, verifies assumptions against web sources, applies fixes, and reports concerns ranked by severity
• consensus — Runs N blind agents on the same job in parallel, merges findings into consensus/unique/conflicts (no edits for concurrent safety)
• seq — Runs N serial blind passes with commits between rounds; requires clean worktree
• rigor — Activates Rigor Mode for the session: prioritizes correctness, thorough investigation, and web-grounded verification over minimalism
• orchestrate — Activates Orchestrator Mode for the session: delegates execution to subagents by default and operates at the high level, conserving context for direction and synthesis; an escape hatch keeps trivial work inline
• askme — Shorthand: stop and ask the user to clarify ambiguities or make decisions instead of assuming
• delegate — Shorthand: prefer subagents to save context space and parallelize independent subtasks
• tmi — Flags content that only makes sense if you were in the room when it was written; reports by default, edits if explicitly instructed
• dry-run — One-shot failsafe: describes what it would do for the next request instead of doing it, then waits for confirmation
• timeless — Shorthand: avoid time estimates (hours, calendar, size-to-time buckets); describe complexity, scope, risk, and ordering instead
• dredge — Searches prior coding-agent chat transcripts (Claude Code, Codex, ...) for context; defaults to the current project, widens scope and time window from natural-language hints in the query (e.g. "across all projects", "in the craken repo", "yesterday"). Uses an optional AgentsView backend when the agentsview CLI is installed (configured at user scope under ~/.agents/skill-configs/dredge/); falls back to grep over Claude Code transcripts otherwise
• timestamp — Shorthand: prefix newly created files/folders with a yymmdd-HHMMSS stamp from the date CLI; one timestamp per logical job bucket; one turn only

gimme

User-invoked inversion of delegation — hand the agent a request and get back a filesystem bundle you can act on.

/gimme

• Writes a timestamped bundle with checklist.md, notes.md (template with pre-labeled paste slots), and an empty dropbox/ directory for file artifacts
• Each checklist item has action / why-it's-on-you / drop-path so results land somewhere the agent can pick up without further direction
• Optional launch_command config (e.g. cursor {path}, code {path}, open {path}) opens the bundle in your editor immediately
• Never self-invoked — only runs when you explicitly call /gimme

Other (copy / symlink)

Available in the repo but not published to the marketplace. Install via symlink or copy:

cp -R ~/shared/plugins/<plugin>/skills/<skill> <skill-root>/

Invocation snippets below are Claude Code style; in other agents, ask for the workflow by skill name or description.

phaser

Battle-tested patterns and best practices for Phaser 3 game development.

Passive knowledgebase — automatically consulted when writing Phaser code.

• Multi-scene flow, object pooling, and physics patterns
• Performance optimization and common pitfall avoidance
• Architecture guidance for game projects

report-writer

Structured technical analysis and debugging reports with standardized sections.

/report-writer [topic]

• Generates timestamped reports (debugging, analysis, implementation)
• Standardized sections: Executive Summary, Key Findings, Root Cause Analysis, Recommendations
• Evidence-based documentation with code references

rigorous-debug

Evidence-based debugging protocol using the scientific method.

/rigorous-debug

• Requires one-time project-specific initialization before first use
• Enforces hypothesis → experiment → conclusion cycles
• Prevents assumption-driven debugging with structured evidence gathering

skill-writer

Tools for creating effective SKILL.md agent skills.

/skill-writer

• Step-by-step workflow from pattern identification to polished SKILL.md
• Covers frontmatter, instruction structure, and best practices
• Helps extract reusable patterns into shareable skills

Workspace Configuration

File-producing skills (interview, spex, report-writer, macros, study, gimme) support configurable workspace directories with layered precedence (first match wins):

1. Explicit override — ask to use a specific workspace directory for this run
2. Local config (.agents/skill-configs/<skill>/config.local.yaml) — gitignored, personal overrides
3. Project config (.agents/skill-configs/<skill>/config.yaml) — committed to repo, shared with team
4. Legacy fallback (.claude/skill-configs/<skill>/) — older installs

There are no built-in defaults. Each skill prompts for setup on first use. Output follows the .agent-workspace/<folder> convention (specs, reports, interviews, macros, study, gimme).

Use .agents/skill-configs/ for new configuration. The .claude/skill-configs/ path is retained only as a legacy fallback for older installs.

Other Agentic Tools

These skills use the Agent Skills SKILL.md format. Provider-specific install roots and activation notes live in the cross-platform guide; avoid copying those details into other docs.

License

Personal use — adapt and modify as needed for your own projects.