Applies user feedback to prompts with precision - incorporates feedback intent without over-fitting, maintains information density, prevents regressions. Use when asked to apply feedback, incorporate suggestions, or update a prompt based on user input.
Applies user feedback to prompts with precision - incorporates feedback intent without over-fitting, maintains information density, prevents regressions. Use when asked to apply feedback, incorporate suggestions, or update a prompt based on user input.
Apply Prompt Feedback
Apply user feedback to prompts with calibrated precision—addressing feedback intent fully while avoiding over-fitting, maintaining information density, and preventing regressions.
Overview
This skill applies feedback through:
Initial Application - Apply feedback to prompt using Application Techniques
Verification - prompt-feedback-verifier checks for issues (over-fitting, regression, density loss)
Refinement - Fix issues based on verifier findings, iterate if needed (max 5 iterations)
Output - Atomic replacement only after verification passes
Verifier validates: Verifier catches over-fitting, under-fitting, and regressions
Required tools: This skill requires Task tool to launch the verifier agent. If Task is unavailable, report error: "Task tool required for verification loop." This skill uses TodoWrite to track progress. If TodoWrite is unavailable, track progress internally.
Workflow
Phase 0: Create Todo List (TodoWrite immediately)
Create todos tracking workflow phases:
- [ ] Input validation; done when prompt + feedback files read successfully
- [ ] Initial application→verify; done when verifier returns VERIFIED or issues fixed
- [ ] (Expand: refinement iterations on ISSUES_FOUND)
- [ ] Output updated prompt; done when updated prompt displayed + optionally saved
Phase 1: Input Validation
Mark "Input validation" todo in_progress.
Step 1.1: Parse arguments
Extract from $ARGUMENTS:
Prompt (file path or inline)
Feedback (may be in arguments or need to ask)
Step 1.2: Determine prompt source
Input Type
Detection
Action
File path
Contains / or \, OR ends with .md, .txt, .yaml, .yml
Read file content
Inline prompt
Does not match file path criteria
Write to /tmp/prompt-{timestamp}.md
Priority: If ambiguous, treat as file path first; if not found, treat as inline.
Step 1.3: Gather feedback
If feedback not in arguments, use AskUserQuestion:
questions: [
{
question: "What feedback do you want to apply to this prompt?",
header: "Feedback",
options: [
{ label: "I'll type it", description: "Enter feedback in the text field" }
],
multiSelect: false
}
]
Step 1.4: Validate inputs
If prompt file exists and has content: Use file content
If prompt file empty: Error: "Cannot apply feedback to empty prompt: {path}"
If file not found: Re-classify as inline, write to temp file
If no prompt: Error: "Usage: /apply-prompt-feedback [feedback] OR provide feedback when prompted"
If no feedback after asking: Error: "Feedback required to proceed"
Step 1.5: Store metadata
source_path: Source file path (or temp path for inline)
is_inline: Boolean
original_content: Full prompt text
feedback: User's feedback text
original_path: /tmp/feedback-original-{YYYYMMDDHHMMSS}-{4-lowercase-alphanumeric}.md (copy of original for verifier comparison)
prompt: "Verify feedback application. Original: {original_path}. Modified: {working_path}. Feedback: {feedback}. Check for: feedback not addressed, partial incorporation, over-fitting, over-specification, regression, information density loss. Report VERIFIED or ISSUES_FOUND with specific details."
Step 2.4: Handle verifier response
If "VERIFIED": Mark todo completed, proceed to Phase 4
If "ISSUES_FOUND": Mark todo completed, save issues, add "Refinement iteration 1" todo, proceed to Phase 3
If verifier fails: Retry once. If retry fails, proceed to Phase 4 with warning: "Verification failed - manual review recommended."
Step 2.5: Display findings
If issues found:
Feedback Summary: {summary from verifier}
Verifier found {count} issues. Proceeding with refinement...
Mark "Initial application" todo completed.
Phase 3: Refinement Loop (Verifier-Driven)
Mark "Refinement iteration 1" todo in_progress.
Key principle: All refinements driven by verifier feedback. Only fix issues the verifier reported.
For each iteration from 1 to 5:
Apply fixes from verifier feedback: For each issue, apply the Suggested Fix using Application Techniques. Write to working_path.
Only address issues the verifier identified
If Write tool fails: display error, proceed to Phase 4 with most recent version
Re-verify: Launch prompt-feedback-verifier agent via Task tool:
prompt: "Verify feedback application. Original: {original_path}. Modified: {working_path}. Feedback: {feedback}. Check for: feedback not addressed, partial incorporation, over-fitting, over-specification, regression, information density loss. Report VERIFIED or ISSUES_FOUND with specific details."
Handle response:
If "VERIFIED": mark todo completed, exit loop, proceed to Phase 4
If "ISSUES_FOUND" and iteration < 5: mark todo completed, save issues, add "Refinement iteration {next}" todo, continue
If "ISSUES_FOUND" and iteration = 5: mark todo completed with note, proceed to Phase 4 with warning
If verifier fails: display error, retry once. If retry fails, proceed to Phase 4 with warning: "Verification incomplete - manual review recommended."
Application Techniques
Apply feedback using these calibrated techniques:
Incorporation Techniques
Issue Type
Technique
Example
Feedback Not Addressed
Add minimal instruction that addresses feedback intent
Add missing aspects only, don't expand what's already there
Feedback: "handle A, B, C" (A done) → Add B and C handling only
Calibration Techniques
Issue Type
Technique
Example
Over-Fitting
Remove unrelated additions, keep only feedback-relevant changes
Remove new sections that feedback didn't request
Over-Specification
Condense verbose implementation to minimal form
"When timeout occurs after 30 seconds of no response from the server, the system should log the error, notify the user, and retry up to 3 times" → "Timeout after 30s; retry up to 3x"
Preservation Techniques
Issue Type
Technique
Example
Regression
Restore removed/changed behavior, integrate with new additions
Re-add removed instruction, resolve any conflict with new content
When applying feedback, maximize information density:
Minimal additions: Add only text necessary to address feedback
Concise phrasing: Use fewest words that convey meaning precisely
No duplication: Don't repeat what prompt already says
Consolidate: Merge related additions into existing sections when possible
Delete redundancy: If adding X makes existing text Y redundant, remove Y
Density checklist before writing:
Can this be said in fewer words?
Does existing text already partially cover this?
Can this be merged with an existing section?
Am I adding explanation where a rule suffices?
Resolution Strategy
For each issue:
Issue from verifier
│
▼
Is the fix text provided in Suggested Fix?
│
├─ YES → Apply exact fix text
│
└─ NO → Use Application Techniques
│
▼
Apply minimal change that:
1. Addresses the issue
2. Uses fewest possible words
3. Preserves existing content
4. Doesn't introduce new issues
Phase 4: Output
Mark "Output updated prompt" todo in_progress.
Step 4.1: Apply changes
After application complete:
For file input: mv {working_path} {source_path} (atomic replacement)
For inline input: Keep at working_path, report location
Step 4.2: Display results
If verification passed:
Updated: {path}
Iterations: {count}
Status: Feedback applied
Feedback: {original feedback}
Changes applied:
- {List each change as bullet, categorized: "Incorporation:", "Calibration:", or "Preservation:"}
If verification failed after 5 iterations:
Updated with warnings: {path}
Iterations: 5
Status: Some issues may remain
Feedback: {original feedback}
Unresolved issues:
- {list from last verification}
Review the changes manually.
Mark "Output updated prompt" todo completed. Mark all todos complete.
Key Principles
Principle
Rule
Apply then verify
Apply feedback first, verifier catches issues
Verifier-driven refinement
Refinements only fix issues verifier identifies
Just-right application
Address ALL feedback, ONLY feedback
Information density
Minimal text to achieve the change; no bloat
Regression prevention
Preserve everything feedback didn't mention
Track progress
TodoWrite for phases; expand on iteration
Atomic output
Original untouched until application complete
Edge Cases
Scenario
Handling
No prompt provided
Error: "Usage: /apply-prompt-feedback [feedback]"
No feedback provided
Use AskUserQuestion to gather feedback
File not found
Re-classify as inline prompt, write to temp file
Empty file
Error: "Cannot apply feedback to empty prompt: {path}"
Vague feedback
Apply reasonable interpretation; verifier will catch over/under-fitting
Conflicting feedback
Apply most recent/prominent request; note conflict in output
Verifier fails
Retry once; if fails, output with warning: "Verification failed - manual review recommended."
Task tool unavailable
Error: "Task tool required for verification loop."
Example Usage
# Apply feedback to a prompt file
/apply-prompt-feedback prompts/assistant.md "Make it handle empty input gracefully"# Apply feedback to inline prompt (will be prompted for feedback)
/apply-prompt-feedback "You are a helpful assistant. Answer questions concisely."# Apply multiple feedback points
/apply-prompt-feedback skills/reviewer/SKILL.md "Add timeout handling and clarify what 'significant' means"
Example Output
Updated: prompts/assistant.md
Iterations: 2
Status: Feedback applied
Feedback: Make it handle empty input gracefully
Changes applied:
- Incorporation: Added "If input is empty, respond with 'Please provide a question or topic.'"
- Preservation: Existing response format unchanged
- Calibration: Kept addition minimal (single sentence vs. elaborate error handling)
