القائمة
Take something that works but could be better. Find the highest-leverage improvements and apply them. Not for broken things (/fix) — for things that work but underperform their potential.
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
Route any input through a branching question tree to narrow down the optimal response strategy before writing. Two stages — PERCEIVE (classify input) then ACT (select response). Covers all prompt types.
Generate exhaustive guesses about user input using ALL search methods with coverage tracking. Guessing is SEARCH through possibility space. Tracks space created vs space covered to ensure comprehensive exploration.
Systematically evaluate and select from a set of guesses, options, or possibilities. Combines ARAW analysis with prioritization to determine which guesses are strong, weak, actionable, or eliminable.
Extract hidden assumptions from any content. Surfaces what must be true for claims to hold, enabling deeper analysis.
Sub-orchestrator for analytical requests. Routes to decomposition, systems analysis, comparison, risk assessment, or synthesis based on what kind of analysis is needed.
Assume Right - Deep recursive rightness search. For every claim, assume it's right — find what must follow, then assume THOSE implications are right too. Recurse until bedrock. Track every claim found.
| name | improve - Improve It |
| description | Take something that works but could be better. Find the highest-leverage improvements and apply them. Not for broken things (/fix) — for things that work but underperform their potential. |
| output | {"format":"prose"} |
Input: $ARGUMENTS
Before executing, identify what's being improved:
Interpretation 1 — Code: Working code that could be cleaner, faster, more maintainable, or better designed. Interpretation 2 — Skill/document: A skill, plan, doc, or structured text that works but isn't good enough. Could be tighter, clearer, more effective. Interpretation 3 — Output: Something produced in this conversation that was adequate but not great. Raise the quality bar. Interpretation 4 — Process/approach: A way of doing things that works but leaves value on the table.
If the thing is BROKEN (errors, wrong output, fundamentally flawed), use /fix instead. /improve is for things that work but aren't as good as they could be.
Read the entire target. While reading, establish:
What is this trying to do? (purpose) How well does it do it? (current quality — rate 1-10 with one-line justification) What does excellent look like for this kind of thing? (quality bar — find or name a concrete reference point)
PURPOSE: [what it's for]
CURRENT QUALITY: [specific strengths and weaknesses] — VERDICT: [needs significant work / needs targeted improvements / good enough to ship / excellent]
QUALITY BAR: [what excellent looks like for this kind of thing]
GAP: [the distance between current and excellent]
Look for these in order. For each, state what's improvable and WHY it matters (not just that it could be different).
List all opportunities. Tag each:
IMPROVEMENTS:
[I1] [what to improve] — LEVERAGE: high/medium/low — EFFORT: high/medium/low
[I2] [what to improve] — LEVERAGE: high/medium/low — EFFORT: high/medium/low
...
Do first: high leverage + low effort Do next: high leverage + high effort Skip unless asked: low leverage + any effort
Execute the high-leverage improvements. Use Edit for targeted changes, Write for rewrites when >50% is changing.
For each improvement applied, the change should be visible in the diff — don't just describe what you'd change, CHANGE IT.
After applying improvements:
BEFORE: [N]/10
AFTER: [N]/10
DELTA: +[N]
IMPROVEMENTS APPLIED: [N]
[I1]: [one-line summary]
[I2]: [one-line summary]
...
REMAINING OPPORTUNITIES (diminishing returns):
[I-number]: [what and why it's not worth doing now]
QUALITY BAR MET: [yes/no — if no, what's still short]
| Depth | Read | Find | Apply |
|---|---|---|---|
| 1x | Skim | Top 1-2 high-leverage | Quick improvements |
| 2x | Full read | All opportunities, prioritized | Apply high-leverage |
| 4x | Full read + context + comparisons | All opportunities + benchmark against best examples | Apply all worthwhile |
| 8x | Full read + context + history + user patterns | Full quality audit + structural analysis | Rewrite if needed, verify against quality bar |
Default: 2x.
| Failure Mode | Signal | Fix |
|---|---|---|
| Taste-based changes | "I'd prefer..." without quality justification | Every change must have a WHY rooted in clarity, density, effectiveness, specificity, structure, or completeness — not preference. |
| Over-improvement | Rewriting working sections that are already good | If it's already at 8/10, leave it. Improve the 4/10 parts. |
| Feature creep | Adding new capabilities instead of improving existing ones | Improve means make the EXISTING thing better, not add to it. |
| Cosmetic improvement | Reformatting, renaming, restyling without substance change | Ask: does this make the thing MORE EFFECTIVE or just different? |
| Improvement without application | Listed what to improve but didn't do it | Step 4 exists. Apply the changes. |
| Wrong tool | The thing is actually broken, not underperforming | Use /fix for broken things. /improve assumes working baseline. |
/fix → for broken things. If /improve discovers the thing is fundamentally broken, route to /fix./araw → stress-test whether improvements actually make it better/cmp → compare before/after versions/cls → check improvements against specific quality criteria/iterate → for multiple rounds of improvementExecute now.