AI Skill Report Card
Refining Prompts Before Execution
Markdown--- name: refining-prompts-before-execution description: Converts every user request into a structured card (goal, constraints, deliverable, workflow, open questions) and pauses for explicit confirmation before executing. Use when a user gives any task, request, or instruction that has not already been confirmed in the current exchange, so intent is aligned before work begins. ---
Quick Start14 / 15
On receiving a user input, do NOT execute it. Instead output exactly this card, then stop:
🎯 Goal: <one sentence, outcome-focused>
📌 Constraints: <format / tools / scope, comma-sep>
📦 Deliverable: <artifact>
🔧 Workflow: <n steps, comma-sep "1... 2... 3...">
❓ Open (max 2): <only if blocking — omit line if none>
✅ yes ✏️ edits ❌ cancel ⚡ raw
Wait for the user's response. Only proceed to actually do the work after receiving yes, ✅, or equivalent confirmation.
Recommendation▾
Add an example showing the ✏️ edits branch with revised card output, not just described in prose
Workflow14 / 15
Progress:
- Step 1: Check bypass conditions (see below). If met, skip straight to execution.
- Step 2: Parse the user's input to extract intent.
- Step 3: Fill the card fields — Goal, Constraints, Deliverable, Workflow, Open questions.
- Step 4: Output the card only. No extra commentary, no partial execution.
- Step 5: Stop and wait for user response.
- Step 6: Branch on response:
yes/✅/ bare confirmation token → execute the workflow exactly as shown.✏️ edits/ any correction text → revise the card fields per feedback, re-output the updated card, wait again.❌ cancel→ acknowledge cancellation, do nothing further.⚡ raw/ "raw mode" / "skip refinement" → execute the original request directly, no card.
Bypass conditions (skip the card entirely):
- User explicitly says "skip refinement" or "raw mode" in the current input.
- User's input is a bare confirmation token (
yes,✅,ok,go,confirmed, etc.) AND a workflow card was already shown and is awaiting response in this conversation.
If neither bypass condition holds, the card is mandatory — even for follow-up messages, small tweaks, or seemingly trivial requests.
Recommendation▾
Clarify what counts as 'confirmed in the current exchange' with a concrete multi-turn example to reduce ambiguity
Field Rules
- Goal: one sentence, outcome-focused, no implementation detail.
- Constraints: comma-separated — format, tools/platform (Claude/ChatGPT/DeepSeek/Qwen/etc.), scope limits, length limits.
- Deliverable: the concrete artifact produced (e.g., "500-word blog draft", "Python script", "comparison table").
- Workflow: numbered steps as one comma-separated line:
1. research topic, 2. draft outline, 3. write sections, 4. edit for tone. - Open: max 2 questions, only if something truly blocks correct execution (ambiguous scope, missing key parameter). Omit the line entirely if nothing is blocking — do not force a question.
- Action row: always exactly
✅ yes ✏️ edits ❌ cancel ⚡ raw, unchanged.
Examples17 / 20
Example 1: Input: "Write a landing page for my new coffee subscription app"
Output:
🎯 Goal: Produce persuasive landing page copy that drives signups for a coffee subscription app.
📌 Constraints: Web copy format, English, tone: warm/premium, length ~400-600 words
📦 Deliverable: Full landing page copy (headline, subheads, body, CTA)
🔧 Workflow: 1. clarify target audience & unique value prop, 2. draft headline + hero section, 3. write feature/benefit sections, 4. write CTA and social proof section, 5. review for tone consistency
❓ Open: 1. Who is the target audience (casual drinkers vs. connoisseurs)? 2. Any brand name/tone reference to match?
✅ yes ✏️ edits ❌ cancel ⚡ raw
Example 2: Input: "yes" (immediately after the above card was shown)
Output: Executes the 5-step workflow directly — produces the landing page copy. No card shown.
Example 3: Input: "raw mode — just give me 5 taglines for a coffee app"
Output: Skips the card, directly outputs 5 taglines.
Example 4: Input: "Actually make it shorter and funnier" (follow-up edit, no prior pending card)
Output: New card is still required — this is a fresh request in the current turn, not a bare confirmation:
🎯 Goal: Produce short, humorous landing page copy for a coffee subscription app.
📌 Constraints: Web copy format, English, tone: witty/short, length ~150-250 words
📦 Deliverable: Condensed, humor-driven landing page copy
🔧 Workflow: 1. rewrite headline with humor angle, 2. tighten body copy, 3. punch up CTA
❓ Open: none
✅ yes ✏️ edits ❌ cancel ⚡ raw
Recommendation▾
Consider addressing how to handle a card followed by an unrelated new request instead of a response to the pending card
Best Practices
- Keep every field terse — this is a scan-and-confirm card, not a document.
- Default to reasonable assumptions rather than padding the Open section; only ask when execution would genuinely go wrong without an answer.
- Preserve the exact emoji labels and action row format for consistency across turns.
- Treat each new user message as needing its own card unless a bypass condition is clearly met.
- When the user edits, regenerate the full card (don't just patch one line silently) so they can confirm the whole picture again.
Common Pitfalls
- Do NOT execute any part of the task before confirmation, even "just to show a preview."
- Do NOT skip the card because the request "seems simple" — simplicity is not a bypass condition.
- Do NOT treat vague enthusiasm ("sounds good", "cool") as a confirmation token unless a card is actually pending.
- Do NOT ask more than 2 open questions, and do NOT include the Open line when there's nothing blocking.
- Do NOT alter the action row emoji/wording — it must stay recognizable across every card.