candidate-kb
Manage candidate knowledge base. Use when ingesting CVs, adding profile information, or querying candidate details for resume/cover letter generation.
Collect user feedback about the Agentic Meta Project and file it. Use when the user says "feedback", "report a bug", "file an issue", or wants to report friction, suggestions, or feature requests. Files as a GitHub issue if gh CLI is available, otherwise writes a local report.
---
name: feedback
description: |
Collect user feedback about the Agentic Meta Project and file it. Use when the
user says "feedback", "report a bug", "file an issue", or wants to report
friction, suggestions, or feature requests. Files as a GitHub issue if gh CLI
is available, otherwise writes a local report.
---
# feedback
## Purpose
Provide a structured way for users to report bugs, friction, suggestions, or feature requests back to the meta-project maintainers. This closes the feedback loop between real-world usage and template improvements.
## Preconditions
- The user wants to report an issue, suggestion, or friction point.
## Procedure
1. **Ask the user for feedback type.** Options: bug, friction, suggestion, feature request.
2. **Collect a description.** Ask the user to describe the issue. Include:
- What happened (or what they want).
- Steps to reproduce (if applicable).
- What they expected instead.
3. **Gather context.** Note the following automatically (ask the user before including project-specific details):
- Client project name (anonymize if the user requests).
- Amplify version from `.agents/project.json`.
- Which skills/workflows were involved (if applicable).
4. **Check if `gh` CLI is available.** Run `gh auth status` to verify.
5. **If `gh` is available**: create a GitHub issue on the meta-project repository with:
- Title: `[<feedback-type>] <short summary>`
- Labels: `feedback`, `<feedback-type>`
- Body:
```
## Feedback Type
<bug | friction | suggestion | feature request>
## Description
<user's description>
## Reproduction Steps
<steps, or "N/A">
## Context
- Amplify version: <version>
- Project type: <type>
- Skills involved: <list or "none">
## Expected Behavior
<what the user expected>
---
Filed via `/feedback` skill.
```
- Confirm the issue URL with the user.
6. **If `gh` is not available**: write the report to `.untracked/feedback/<YYYY-MM-DD>-<summary-slug>.md` using the same format. Tell the user where the file is and how to file it manually.
7. **Update feedback index.** Append a row to `.untracked/feedback/feedback-index.md`:
`| <YYYY-MM-DD> | <short summary> | <feedback-type> | open | <filename or issue-URL> |`
If `feedback-index.md` does not exist, create it with the header row first.
## Output
- **If `gh` available**: a GitHub issue URL.
- **If `gh` unavailable**: a local file at `.untracked/feedback/<YYYY-MM-DD>-<summary-slug>.md`.
## Completion Criteria
- Feedback type and description collected from user.
- Issue filed (GitHub) or report written (local).
- User informed of the result.
## Error Handling
- If `gh auth status` fails, fall back to local file without asking the user to authenticate.
- If the GitHub issue creation fails, fall back to local file and report the error.
- If `.untracked/feedback/` does not exist, create it.
---
name: feedback
description: |
Collect user feedback about the Agentic Meta Project and file it. Use when the
user says "feedback", "report a bug", "file an issue", or wants to report
friction, suggestions, or feature requests. Files as a GitHub issue if gh CLI
is available, otherwise writes a local report.
---
# feedback
## Purpose
Provide a structured way for users to report bugs, friction, suggestions, or feature requests back to the meta-project maintainers. This closes the feedback loop between real-world usage and template improvements.
## Preconditions
- The user wants to report an issue, suggestion, or friction point.
## Procedure
1. **Ask the user for feedback type.** Options: bug, friction, suggestion, feature request.
2. **Collect a description.** Ask the user to describe the issue. Include:
- What happened (or what they want).
- Steps to reproduce (if applicable).
- What they expected instead.
3. **Gather context.** Note the following automatically (ask the user before including project-specific details):
- Client project name (anonymize if the user requests).
- Amplify version from `.agents/project.json`.
- Which skills/workflows were involved (if applicable).
4. **Check if `gh` CLI is available.** Run `gh auth status` to verify.
5. **If `gh` is available**: create a GitHub issue on the meta-project repository with:
- Title: `[<feedback-type>] <short summary>`
- Labels: `feedback`, `<feedback-type>`
- Body:
```
## Feedback Type
<bug | friction | suggestion | feature request>
## Description
<user's description>
## Reproduction Steps
<steps, or "N/A">
## Context
- Amplify version: <version>
- Project type: <type>
- Skills involved: <list or "none">
## Expected Behavior
<what the user expected>
---
Filed via `/feedback` skill.
```
- Confirm the issue URL with the user.
6. **If `gh` is not available**: write the report to `.untracked/feedback/<YYYY-MM-DD>-<summary-slug>.md` using the same format. Tell the user where the file is and how to file it manually.
7. **Update feedback index.** Append a row to `.untracked/feedback/feedback-index.md`:
`| <YYYY-MM-DD> | <short summary> | <feedback-type> | open | <filename or issue-URL> |`
If `feedback-index.md` does not exist, create it with the header row first.
## Output
- **If `gh` available**: a GitHub issue URL.
- **If `gh` unavailable**: a local file at `.untracked/feedback/<YYYY-MM-DD>-<summary-slug>.md`.
## Completion Criteria
- Feedback type and description collected from user.
- Issue filed (GitHub) or report written (local).
- User informed of the result.
## Error Handling
- If `gh auth status` fails, fall back to local file without asking the user to authenticate.
- If the GitHub issue creation fails, fall back to local file and report the error.
- If `.untracked/feedback/` does not exist, create it.
Imported from the listed AgentSkill.sh category. Complete supplied package files and notices are retained. Required external tools and credentials are not installed. No bundled code was executed during import. Individual agent task behavior has not been tested.