Personal AI Skill Cheat Sheet — When to Use Each Skill

Updated 2026-05-24: incorporates the May 23 install batch — CodeGraph MCP, react-doctor, the full openai/skills curated catalog (39 skills incl. figma/notion/deploy stacks), the PaperSpine paper-writing suite (1 skill on Codex via dist/codex/, 7 skills on Claude Code via dist/claude/skills/), and the complete voidful/academic-skills set (paper-reading/review/writing + experiment-design + idea-generation + proof-writer + professor-fit-analyser).

This is Q’s personal router for the Codex and Claude Code skill stack. A skill is a task-specific playbook: it tells the agent when to slow down, what workflow to follow, what files or tools to inspect, and what quality bar to satisfy.

The reliable pattern is to name the skill explicitly, with invocation adjusted to the runtime:

• Codex: Use diagnose..., Use $tdd..., or Use kicad...
• Claude Code: use /skill-name for any user-invocable skill or plugin command (both skill folders AND plugins can register slash commands); otherwise use plain language (“use tdd”) for skills the agent auto-routes to.

Operational legend

Not every item in the skill stack behaves the same way.

Before first use of any third-party skill, review SKILL.md, scripts, hooks, and MCP/plugin manifests. Re-verify this page after installing, removing, or replacing skills because the stack will drift.

The mental model

Use skills as phase controls, not as a giant menu. Pick the skill that matches the current kind of work.

Unclear goal
    |
    v
grill-me / grill-with-docs
    |
    v
create-plan / to-prd / to-issues
    |
    v
tdd / diagnose / prototype
    |
    v
webapp-testing / semgrep / codeql / fuzzing
    |
    v
differential-review / insecure-defaults / supply-chain-risk-auditor
    |
    v
changelog-generator / content-research-writer / handoff

Three rules make the stack usable:

1. Name the skill explicitly when you care. Do not hope the agent guesses the right workflow.
2. Use one skill per phase. zoom-out, diagnose, and differential-review are sequential tools, not one combined prompt.
3. Treat third-party skills as operational instructions. They can tell agents to run commands, use tools, or trust sources. Keep project rules and human review in charge.

Daily default stack

Use this smaller set before reaching for the full catalog.

Quick chooser

Normal coding workflow

Use these for everyday software work.

For a normal feature, the default flow is:

grill-with-docs
    -> create-plan
        -> tdd
            -> webapp-testing
                -> differential-review

For a bug, use:

zoom-out
    -> diagnose
        -> tdd
            -> differential-review

Security and review

Security skills are review aids, not proof of safety. Static analysis, fuzzing, and scanners need build context, dependencies, and human validation.

Good prompt:

Use differential-review on my current git diff. Prioritize auth, crypto, validation, external calls, missing tests, and blast radius.

Fuzzing and test depth

Use fuzzing skills when ordinary unit tests are too example-driven.

Typical loop:

harness-writing
    -> aflpp/libfuzzer/cargo-fuzz/atheris
        -> coverage-analysis
            -> fuzzing-obstacles
                -> fp-check

Crypto, protocols, and blockchain

Code graph and audit context

Use these when review quality depends on structure, not just text diff.

CodeGraph rule of thumb (May 23 install)

CodeGraph is the only manually-configured stdio MCP server currently wired into both ~/.claude.json and ~/.codex/config.toml (plugins or connectors may add their own tools separately). Steering text lives in ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md. Use it for structural questions, not literal text search.

Indexed projects as of 2026-05-24 (verified by ls ~/Desktop/*/.codegraph/): learnai-3d-studio only. Run codegraph init -i in any other project to enable. If .codegraph/ doesn’t exist, the MCP returns “not initialized” — for literal text search use rg instead, for structural queries ask the user whether to initialize CodeGraph.

Financial modeling, spreadsheets, and decks

These came from Anthropic’s financial-services skill folders. Some are finance-specific; others are general spreadsheet or presentation utilities.

Finance-specific modeling:

General spreadsheet and deck utilities:

Example:

Use audit-xls to check this model for formula errors, balance sheet balance, cash tie-out, hardcodes, and logic sanity.

Figma and UI design assistance

These came from the May 23 openai/skills/.curated/ install. Use them when the artifact is a Figma file, a design system, or screens generated from spec.

Pair with brand-guidelines, theme-factory, or canvas-design for visual polish.

Hardware, KiCad, PCB, and BOM

Use the KiCad stack when the artifact is electrical, not software.

Example:

Use kicad, bom, datasheets, and emc to review this PCB before fabrication. Start with schematic/PCB consistency, then BOM risk, then EMC risks.

Writing, research, docs, and media

Academic and research stack (May 23 expansion)

Three different academic pipelines now coexist, with different shapes:

• academic-research-suite (Imbad0202): one Codex-native adapter / root skill that internally routes through “select-topic → literature → write → review → revise → finalize” workflows. Best when you want a single skill to dispatch the full pipeline.
• PaperSpine (WUBING2023): orchestrator + 6 specialized sub-skills (Claude side only — Codex only got the single paper-spine orchestrator from dist/codex/). Best when you want explicit motivation-driven scaffolding.
• voidful/academic-skills: 7 standalone persona-heavy skills (some default to Traditional Chinese with character personas). Best when you want a specific narrow tool, e.g. “explain this paper” or “draft a proof.”

The academic stack is best when the output needs research structure or formal venue conventions. The general writing stack below is best when the output needs audience fit, citations, or polish without paper-format constraints.

General writing, comms, and media

Ops, integrations, and production debugging

Use these only when the relevant CLI, credentials, or service access exists. Important: a skill folder by itself does NOT grant OAuth/API access — vercel-deploy, netlify-deploy, linear, sentry, etc. only tell the agent how to use the corresponding CLI; you must still install the CLI and authenticate it (vercel login, gh auth login, etc.) before the agent can act.

Utility skills

What Q should use most

For teaching and LearnAI work, the highest-leverage pattern is:

content-research-writer
    -> theme-factory / canvas-design
        -> webapp-testing
            -> differential-review

For research and paper work:

academic-research-suite
    -> content-research-writer
        -> proof-writer or experiment-design if needed

For security coursework:

zoom-out
    -> trailmark-summary
        -> semgrep/codeql
            -> fp-check
                -> differential-review

Deferred or special-case skills

Some useful items were intentionally not treated as ordinary skill-folder installs:

• Composio app automations: the composio-skills/* set claims 1000+ app integrations / actions (per upstream). Install only when you need a specific app.
• Claude native plugins: some repos add slash commands, hooks, MCP servers, or agents. Ask before installing those because they change more than skill discovery.
• awesome-codex-cli (RoggeOhta): reference list only; curates 150+ tools/skills/subagents/plugins but does not itself ship installable skill folders.
• awesome-codex-skills (ComposioHQ): NOT just a reference list — it ships installable skill folders and a skill-installer/ Python script. The Q installation already uses this installer at ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py. Browse the README and python3 install-skill-from-github.py --repo ... --path ... --name ... to add individual skills.
• Existing pdf in Claude Code: left untouched on the Claude side because it already existed and did not match the pinned OpenAI curated copy. The Codex side uses the OpenAI curated pdf.
• rohitg00/agentmemory: explicitly skipped. The MCP-plus-12-hooks footprint conflicted with the file-based auto-memory system already in place. See CodeGraph entry for the parallel decision on a more focused MCP install.
• aklofas/kicad-happy (full Claude/Codex plugin + GitHub Action): the skill folders (kicad, bom, datasheets, spice, emc, jlcpcb, pcbway, digikey, mouser, lcsc, element14, kidoc) ARE installed and usable — visible in the Hardware/KiCad section above. The kicad-happy native plugin (with hooks + GitHub Action) is intentionally NOT enabled, pending real PCB work.
• Asen pack mapping to currently-loaded equivalents:

So 5 of 6 map to OMC; prompt-optimizer maps to a separate prompt-master skill. Asen originals were not separately installed pending a side-by-side comparison.

MCP servers currently active

User-level (loads in every session):

Project-level Claude MCPs (per ~/.claude.json → projects.<path>.mcpServers, as of 2026-05-24):

To disable CodeGraph: run codegraph uninstall (verify the CODEGRAPH_START/END blocks are gone from ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md after — CLI removes the MCP entries reliably; the steering-block cleanup is worth diffing).

How to ask the agent

Use direct invocation language:

Use diagnose. Reproduce the bug with the fastest deterministic loop, rank hypotheses, and only fix after the loop proves the failure.

Use grill-with-docs. Challenge this plan against the repo's domain language and write down the decisions that should become docs.

Use differential-review on the current git diff. Prioritize auth, crypto, validation, external calls, missing tests, and blast radius.

Use kicad and bom on this KiCad project. Check schematic/PCB consistency, BOM completeness, datasheet gaps, and manufacturing risks.

Use academic-research-suite. Goal: turn this rough topic into a literature-review plan with missing-evidence checklist.

The pattern is simple: name the skill, state the artifact, state the quality bar, and state whether edits are allowed.