Add agent skills for AI coding #2804
Conversation
used opencode to read devGuide and generated relevant skills
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## master #2804 +/- ##
=======================================
Coverage 72.09% 72.09%
=======================================
Files 134 134
Lines 7411 7411
Branches 1587 1525 -62
=======================================
Hits 5343 5343
- Misses 1940 2022 +82
+ Partials 128 46 -82 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
based on feedback from skill use
|
The references would definitely be useful for an AI agent to immediately have a good grasp of the codebase. However, I am worried about the maintainability of the references.
How do we find the balance of keeping a abstract and high level explanation of the design in the references enough to not have to add too much maintainability burden, but have it detailed enough to justify the token count and usefulness. This is on top of the current developer documentation that has to be updated as well. wdyt? |
|
Would it be useful if we created a subagent #2822 that updates the skills accordingly when changes are made to the devGuide? the solution to AI is... more AI... |
Some skills are converted to subagents
|
closes #2822 |
Harjun751
left a comment
Harjun751
left a comment
There was a problem hiding this comment.
Some minor nits. Only problems are the prev/next page navigation buttons.
Also let's discuss on reverting the update-docs skill deletion
There was a problem hiding this comment.
I see that the update-docs was converted from a skill to a subagent. Maybe a good time to open up a dialogue on when to use a skill vs a subagent because I'm also not too sure haha.
My impression was to use a skill when there is a structured, repeatable task that an agent would see added benefit from having some "recipe" to follow, and to use subagents when there are tasks that we can modularize FOR the agent (e.g. abstract away some part of a task for an agent).
If my impression is correct then, I think the subagent route would be good for agents who when asked to perform a task (e.g. change behaviour of a particular function) will subsequently and continuously update documentation as-required. For example, agent goes okay, i finished the task now let me delegate to the update docs subagent to update docs as required.
But personally, I prefer to update documentation on my changes at the very end, so I can "take stock" right before I create a PR. Therefore I usually prevent my agents from updating documentation as they are working on the codebase.
I think both routes are acceptable (incrementally update docs vs big-bang update), so should we still keep the update-docs skill for developers such as myself? Correct me if any of my assumptions/impressions were wrong.
|
|
||
| - Files scheduled for deletion/major refactor | ||
| - External patches (keep matching upstream) | ||
| - Files with minimal logic |
There was a problem hiding this comment.
Hmm why should they avoid migrating this? Seems to cut down on some of the busywork LOL
| | CommonJS | TypeScript Equivalent | ES6 | | ||
| | --------------------------- | --------------------- | ------------------------------- | | ||
| | `module.exports = X` | `export = X` | ❌ `export default` (don't use) | | ||
| | `module.exports = { a, b }` | `export = { a, b }` | ✅ `export { a, b }` | |
There was a problem hiding this comment.
I think @gerteck mentioned avoiding emojis in instruction files. No whimsy allowed 😭
| * [Contribute to Documentation]({{baseUrl}}/devGuide/bootcamp/contributeToDocs.html) | ||
| * [Fix a Bug]({{baseUrl}}/devGuide/bootcamp/fixABug.html) | ||
| * [Implement a New Feature]({{baseUrl}}/devGuide/bootcamp/implementAFeature.html) | ||
| * AI Use |
There was a problem hiding this comment.
Perhaps a bit nitpicky, but should the AI use section go under Development? As I assume readers would want to read about the general development workflow first.
| Subagents are specialized AI agents designed to handle specific types of tasks autonomously. For detailed information about available subagents and how to use them, see the [Subagents](subagents.html) page. | ||
|
|
||
| {% from "njk/common.njk" import previous_next %} | ||
| {{ previous_next('../devGuide', 'skills') }} |
There was a problem hiding this comment.
Should the previous page be pointing to the "implement a new feature" subsection of the onboarding header? Currently it's skipping back to the "contributing" section.
If moving the AI Use section below Development, we should remember to take this into account.
Yea this is pretty annoying 😓
| * AI Use | ||
| * [Overview]({{baseUrl}}/devGuide/aiUse/index.html) | ||
| * [Skills]({{baseUrl}}/devGuide/aiUse/skills.html) | ||
| * [Agents]({{baseUrl}}/devGuide/aiUse/subagents.html) |
There was a problem hiding this comment.
Nit: Page title in frontmatter is "Subagents" but navigation just lists it as "Agent". Should we standardize to "Subagents"?
3 participants
used opencode to read devGuide and generated relevant skills
What is the purpose of this pull request?
Overview of changes:
This PR serves as part of the effort to develop an AI workflow for development.
We opted to use the open source SKILL.md standard as it is provider agnostic and has good community support.
Therefore we could adopt mainstream and popular workflows developed by others without being restricted to a specific provider (e.g. ClaudeCode).
Useful references in creating skills:
Anything you'd like to highlight/discuss:
Testing instructions:
Use AI Coding tools like OpenCode/ClaudeCode to start writing code. Check if the skills are being picked up
correctly and performs as desired
Proposed commit message: (wrap lines at 72 characters)
Add Agent Skills.
Checklist: ☑️
Reviewer checklist:
Indicate the SEMVER impact of the PR:
At the end of the review, please label the PR with the appropriate label:
r.Major,r.Minor,r.Patch.Breaking change release note preparation (if applicable):