القائمة
Craft or refactor LLM instructions. Use when using the Agent or Task tool, when the user says "write a prompt", "prompt", "prompt an agent", or any similar request.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Break a whole into parts at its natural joints
Tile the space of possible futures and evaluate strategies across the tiles. Blends part-whole decomposition with tree and graph of thought. This skill should be used when the user asks "what if", "what could happen if", "evaluate possible futures", "map the scenarios", "explore how this could play out", "compare futures", "stress-test this plan against the future", or faces a decision whose outcome hinges on unresolved uncertainties. Ends with a single reasoned recommendation plus the conditions that would flip it.
Distributed navigation system for multi-file pipelines and processes. Creates compact, grep-friendly waypoint markers that orient readers within complex cross-file workflows — each file carries its own ID, role, adjacent nodes, and a pointer to the full map. Use when asked to "point me to", "where does X begin/end", "map this pipeline", "add waypoints", "trace this process", or when working with processes spanning many files that need discoverability without centralized documentation that drifts.
Ask a genuinely good question — or a composed set of them — in the moment, to discover, clarify, probe, or confirm, in any conversation or domain. Use when about to ask the user something or use the `AskUserQuestion` tool, when a question fell flat or one won't reach and you need a set, when the user's tone or messages signal you have drifted from what they need, or when the user wants questions phrased, generated, or sharpened — "read this spec and generate questions", "ask about the schema", "review this and ask questions", "help me understand what to ask", or to clarify a requirement or idea.
Communicate ideas and information to others with purpose, clarity, and integrity. Ensure that artifacts are structured and phrased in a way that is mindful of their audience while avoiding AI slop patterns. Use when the user requests assistance with writing, commentary, or communication; when they point out AI slop, when they ask things like "help me say", "write for [a specific audience or context]", "polish [comments, sentences, artifacts]", or the like.
Visualize data, concepts, relations, or diagrams. Produces browser-runnable HTML charts and markdown outputs that drop into pull requests, READMEs, tickets, and chat. Use whenever the user asks to visualize anything, make a chart or diagram, plot a trend, compare numbers, show a flow or timeline, or pick the right chart type.
| name | prompt |
| description | Craft or refactor LLM instructions. Use when using the Agent or Task tool, when the user says "write a prompt", "prompt", "prompt an agent", or any similar request. |
| metadata | {"user-invocable":true} |
the user asks for a prompt to be written, refactored, or polished, or pastes an existing prompt or CLAUDE.md asking for improvement.
The output artifact uses a fixed seven-section scaffold. Load assets/template.md and fill it. The procedure below covers detection, inventory, shaping, linting, and emit.
always aim for 80 lines or less in total when it is possible. Keep instructions and sections light and flexible.
identify what the input contains before writing anything.
in seed mode — gather signal from the conversation. Note the task verb, the domain, any files referenced, constraints stated, register requested, audience implied, and stakes. List ambiguities worth surfacing to the user.
in refactor mode — extract every instruction in the source as a flat list with stable IDs (I-01, I-02, ...). The list is the preservation contract. Truth-conditions are sacred — register changes, the underlying requirements do not. Every ID must end up in at least one section of the output, or it has been dropped.
load assets/template.md and fill its seven sections from the inventory.
stakes content sits at the close of the Task section. Name what is at risk and why this matters. There is no separate Why section anywhere in the artifact.
a section has no real content — write a one-line gap note that names the absence, for example: "No prior context identified — agent works from conversation signal alone." pad an empty section with filler.
in refactor mode — every extracted instruction ID lands in at least one section. Consolidations are explicit (note which IDs share a slot). An unassigned ID is a dropped instruction.
load references/eip-anti-patterns.md and read every phrasing in the draft against the patterns it lists.
For each hit, rewrite the phrasing only. The underlying requirement stays. The lint targets register. Truth-conditions are sacred. "You must write tests for every function" becomes "Every function needs a test. Missing tests in a payment system cost more than spurious tests."
writing the natural phrasing first, then checking it against the catalog. The rewrite changes phrasing — never structure, never the underlying instruction. Alignment is what remains after the lint passes; warmth pasted on top of an unaligned prompt does not produce it.
output only the artifact. No preamble, no postscript, no surrounding commentary, no closing question.
the user supplied a path — write the file directly. they did not — display the artifact inline.
fabricate file paths, scripts, or tools that the surrounding context has not established.
a required assumption was made — note it on the first line of the artifact, not in commentary around it.
in refactor mode and the source contains more than ~15 distinct instructions — show the coverage report (every input ID mapped to a destination section) and ask the user to confirm before writing the file. Smaller refactors and seeds emit without asking.
the user has indicated downstream subagent evaluation — keep section boundaries clear and place the coverage report in a fenced block.
the input is genuinely ambiguous between seed and refactor — default to seed. The assumption note on the first line of the artifact is mandatory.
the user pastes a prompt and asks for one specific tweak — run the full lint anyway, but emit a minimal diff. preserving choices the user did not ask to revisit.
the existing prompt is already aligned and would survive the lint untouched — say so plainly and skip the rewrite. Do not churn for theater.
the source prompt contradicts itself — surface the contradiction with AskUserQuestion. silently resolve it.
the user requests a register the lint would otherwise flag, e.g. terse command style for a one-shot formatter — honor it and name which patterns the requested style relaxes. Domain register wins over universal lint.
assets/template.md — the artifact scaffold. Load during shape; the output fills it.
references/eip-anti-patterns.md — the anti-pattern catalog with detection hints and rewrite patterns. Load every run.
references/eip-principles.md — the principles each anti-pattern derives from. Load when a rewrite needs grounding.
