generate-agent-instructions
Generate or update AGENTS.md using evidence from the current repository.
Ignore CLAUDE.md entirely if present.
Workflow
- Locate existing
AGENTS.mdand treat it as the base when present. - Inspect the codebase deeply enough to understand:
- Architecture across multiple files (major components, boundaries, data flow, and structural intent).
- Real developer workflows (build, test, lint, run, debug, and release commands).
- Project-specific conventions and non-obvious patterns.
- Integrations, external services, and cross-component communication.
- Extract only discoverable facts from code, config, scripts, and docs.
- Merge intelligently with existing
AGENTS.md: keep still-valid guidance, remove stale guidance, fill important gaps. - Write concise, actionable instructions targeted to future coding agents.
Output Requirements
- Target roughly 20-50 lines in markdown.
- Prefer repository-specific guidance over generic best practices.
- Include concrete examples with key file or directory references.
- Document current-state behavior only (no aspirational rules).
- Preserve useful existing conventions from
AGENTS.mdwhen still accurate. - Keep wording imperative and easy to execute.
Quality Bar
- Avoid assumptions that are not supported by repository evidence.
- Prioritize information that prevents costly mistakes for new agents.
- Capture the "why" behind important structure when it is inferable.
Update AGENTS.md, then ask the user for feedback on unclear or incomplete sections and iterate.
More from sebkay/skills
audit-dead-code
Find dead code and cleanup candidates such as unused exports, unreachable branches, orphaned files, stale feature flags, dead registrations, and compatibility layers with no live callers. Use when auditing refactors, bundle-size cleanup, architecture simplification, pre-release cleanup, reviewing requests to find unused code or decide what can be deleted, or when deciding whether code can be safely removed or auto-fixed.
20commit
Create a focused git commit for the unit of work just completed. Use when the user asks to commit changes, record finished implementation work, save a fix or refactor to git, or prepare a clean commit from the current branch with a concise message.
16ask-plan-questions
Ask high-impact clarifying questions before implementation to reduce mistakes, rework, and hidden assumptions. Use when creating, reviewing, or updating a plan; when requirements are ambiguous; or when scope/constraints are not fully specified.
15add-plan-todos
Create and maintain granular todos in the active execution plan to reduce omissions, sequencing errors, and misunderstandings. Use when drafting a plan, updating scope, starting implementation, or recovering from uncertainty/blockers.
12audit-errors
Audit code for error-handling inconsistencies, anti-patterns, and silent failures such as empty catch blocks, ignored promise rejections, log-and-continue paths, fallback values that hide faults, broad catch clauses, and inconsistent error translation across layers. Use when reviewing controllers, services, jobs, API handlers, async workflows, UI actions, or any change where errors may be swallowed, downgraded, or surfaced unreliably.
9finalise-plan
Stress-test implementation, execution, rollout, migration, or task plans for missing steps, bad sequencing, weak assumptions, hidden dependencies, vague acceptance criteria, and inadequate validation. Use when a plan has just been drafted or revised, when inheriting someone else's plan, before starting implementation, or when a plan needs one last quality pass before execution.
9