My personal Coding Agent constraints, shared here for backup and reference.
This README was written entirely by hand, without AI assistance. Almost all other files in this repository are 100% AI-generated. The English version was translated by GPT 6.1 Sol.
- grill-me: Clarifies requirements without asking about product architecture.
- sectioned-feature-development: Breaks large features into sections for implementation. Includes a planning knowledge base.
- code-review: Reviews code after a feature change.
- quality-steward: Determines whether a project is due for an audit. Recommended for periodic automated execution.
- general-audit: Periodic project audits, including an audit knowledge base.
- deployment-preflight: Pre-deployment checks.
- adaptive-debugging: A debugging workflow to follow after the first fix attempt fails. Includes a debugging knowledge base.
- debug-notes-to-engineering-notes: Used only when official documentation is incomplete or incorrect. Works with adaptive-debugging, burning your own tokens to light the debugging path for those who come after you.
- feishu-message-sender: Sends periodic progress updates through a Feishu custom bot during long-running tasks.
AGENTS.md is the default filename used by Coding Agents such as Codex and ZCode for their instructions and constraints.
There are two levels of AGENTS.md:
~/.codex/AGENTS.md: Global instructions.AGENTS.mdin the project root: Project-specific instructions.
Main responsibilities:
[To be completed]
The hook scripts are included in this repository. After cloning, run ./scripts/install-git-hooks.sh from the repository root to enable them for the current repository. See git-hook.md for configuration and verification instructions.
- Every completed audit must have
Maintenance-Audit: truein its commit message. - The reminder scripts are located in
.githooks/and do not depend on absolute paths in the user's home directory. - On every commit, Git reminds you when the next audit is due or whether the configured audit interval has already been exceeded.
Users should be able to clearly describe what they want. The Agent uses this skill to ask only about requirements the user has not explained clearly, without asking architecture- or code-level questions.
This skill is an independently rewritten, requirements-focused adaptation inspired by Matt Pocock's
grill-me/grillingskills from https://github.com/mattpocock/skills. The upstream project is licensed under the MIT License. Upstream copyright: Copyright (c) 2026 Matt Pocock
For large features, a complete plan must be created before implementation begins. The feature is then implemented section by section.
The general workflow looks like this:
skill sectioned_feature_development() {
Main Agent creates a complete plan, divides it into sections/subsections,
and assigns models to each;
GPT 6 Astra (Xhigh) performs Plan Review;
for section in sections {
Extract the section into a separate file;
switch (section definition) {
case impl_std:
DeepSeek V4.1 Flash (high) implements the section;
case impl_large:
GPT 6.1 Sol (high) implements the section;
}
while (true) { // Code review loop
Invoke GLM 5.3 with $code-review;
Once issues are resolved, run another review with GPT 6.1 Sol;
if (loop count > 5) {
Package an advisor_pack for GPT 6 Astra Pro to adjudicate;
break;
}
}
}
}Written entirely by GPT 5.6 Sol Pro. I haven't actually read through the details myself, but it has worked quite well in practice.
Suitable for automated execution. Configure the repositories that need monitoring, and have the Agent check their commits at startup to determine whether an audit is due.
I recommend assigning this to a Flash-tier model. It doesn't require particularly strong reasoning capabilities.
An auditing methodology researched and developed by GPT 6 Astra Pro, with web research. Covers most major audit areas.
Checks deployment safety, including whether a deployment could cause database data loss or other issues.
Includes a debugging knowledge base.
Starts with an evidence inventory: what changed during the previous attempt, why it failed to resolve the issue, what the current failure looks like, and whether a runnable reproduction is available. The workflow then proceeds with the original debugging process.
New addition: records debugging attempts under .agent-work/debug so that, after multiple context compactions during particularly difficult debugging sessions, the Agent doesn't forget which unsuccessful approaches have already been tried.
After resolving an issue using adaptive-debugging, this skill consolidates the records from .agent-work/debug into docs/engineering-notes for future development and other contributors.
It documents correct approaches, failed attempts, discrepancies with official documentation, and version-specific findings.
It does not document project-specific or implementation conventions.
Recommended for long-running tasks. Don't have the Agent poll for status updates in the foreground.
This skill copies a template script, modifies it, and launches it in the background. It uses your Feishu bot to send:
- A notification when the task starts.
- Hourly progress reports.
- A final summary when the task finishes.
Bot configuration is stored at ~/.config/feishu-message-sender/secret.json:
{
"webhook": "https://open.feishu.cn/open-apis/bot/v2/hook/xxxxxxxxxx",
"secret": "signature secret",
"name": "name of this computer"
}Official Feishu documentation: Custom Bot Usage Guide (Markdown version, recommended when providing documentation to an Agent).