Ref Sp Agents Instructions Authoring

Agents Instructions Authoring

Purpose

Provide portable defaults for designing maintainable repository instruction systems across the major agent entry points without duplicating the same workflow, policy, and routing text in every provider file.

When to use this skill

• Creating or refactoring repo instruction files.
• Deciding where the source of truth for instructions should live.
• Adding Gemini or Claude support to a repo that already has Copilot instructions, or the reverse.
• Reviewing whether top-level instruction files still match the codebase and the skill catalog.

Scope Boundaries

• Use this skill for the overall instruction architecture across providers.
• Use ./references/providers/copilot-instructions.md for what belongs specifically in .github/copilot-instructions.md.
• Use ./references/providers/gemini-instructions.md for GEMINI.md bridge or provider-specific decisions.
• Use ./references/providers/claude-instructions.md for .claude/CLAUDE.md bridge or provider-specific decisions.
• Use ./references/global-instructions.md for user-level/global instruction files (e.g. ~/.copilot/instructions/*.instructions.md, ~/.claude/CLAUDE.md) rather than repo-scoped files.
• Use ./references/agents-md-standard.md for the cross-provider root AGENTS.md file and how it fits the source-of-truth model.
• Use the repo's agent-persona skill (ref-sp-agents-mr-wolf-persona here) when the instruction system needs to preserve or refresh the repo's agent voice, interaction style, or escalation stance.
• Use the repo's skill-authoring skill (ref-sp-agents-skills-authoring here) for authoring skills rather than top-level instruction files.
• Use the guided instruction-maintenance skill (tool-sp-maintain-agents-instructions here) when the user wants a guided update workflow instead of just the reference guidance.

Major Provider References

• GitHub Copilot: ./references/providers/copilot-instructions.md
• Google Gemini: ./references/providers/gemini-instructions.md
• Anthropic Claude: ./references/providers/claude-instructions.md

Defaults

• Choose one source-of-truth instruction file for the repo.
• Prefer a root AGENTS.md as the repo source of truth by default — it is read natively by many agents; see ./references/agents-md-standard.md. Fall back to .github/copilot-instructions.md only when the repo is Copilot-centric or has a mature file already established there.
• Use thin provider bridge files for Gemini and Claude by default rather than duplicating the full instruction set.
• Keep always-on repo rules in the source-of-truth file and move domain-specific detail into skills.
• Inline the persona core in the source-of-truth file rather than routing to a persona skill — persona is the one category of guidance that must shape every turn, and a skill only loads when something triggers it. See "Persona placement" below.
• Add provider-specific exceptions only when a real platform behavior requires them.
• Keep cross-project personal defaults in user-level/global config, not duplicated into every repo; see ./references/global-instructions.md.
• For the global/home tier, as of today no tool documents an AGENTS.md equivalent, so keep Copilot (~/.copilot/instructions/*.instructions.md) as the recommended global source of truth; see ./references/global-instructions.md.

Task Framing

Core Rules

Source-of-truth model

• Make one file authoritative.
• Default to a root AGENTS.md as that authoritative file; fall back to .github/copilot-instructions.md only when the repo is Copilot-centric or already has a mature file established there (see ./references/agents-md-standard.md).
• Avoid parallel hand-maintained instruction bodies across several provider files.

Import bridge pattern

• In multi-provider repos, prefer a bridge pattern where the provider-specific entry files import or route back to the source-of-truth file.
• Keep bridge files minimal and readable.
• Use repo-root imports when the provider supports them so the bridge does not depend on folder depth.

Persona placement

Persona is the deliberate exception to "move detail into skills". Everything else in a skill can afford to load on demand; the agent's voice, directness, and escalation stance cannot, because they govern how the agent behaves on the very first turn — before any skill has been triggered.

• Inline the persona core in the source-of-truth instruction file (AGENTS.md or equivalent), where it loads on every task. Do not replace it with a pointer to a persona skill.
• Keep the persona skill as the canonical text the inline copy is refreshed against. The skill is the source; the instruction file is the always-loaded projection of it. When they disagree, the skill wins and the instruction file gets updated — not the reverse.
• Inline the core, not the whole skill. The instruction file carries voice, directness, pushing back, and escalation stance. Worked examples, rationale, and adoption guidance stay in the skill.
• Accept the duplication. It is real, and it is the correct trade: a persona that loads lazily is a persona that does not apply when it matters most. Treat it as a deliberate sync point, reviewed whenever either side changes.
• This is the one place a repo should tolerate a hand-maintained copy of skill text in an instruction file. It is not a licence to inline anything else.

In SwiftPost-opinionated setups, that persona is ref-sp-agents-mr-wolf-persona: the repo's AGENTS.md carries its Instructions section inline as the Personality block, and the skill remains the canonical source it is refreshed against. A repo adopting this instruction architecture without the SwiftPost persona applies the same pattern with whatever persona skill it owns.

Provider-specific exceptions

• Add provider-specific text only when the platform has a real bootstrap requirement, limitation, or routing constraint.
• Keep the provider-specific exception narrow and then route back to the shared instructions.
• Do not duplicate the full workflow, command list, or policy text in the provider bridge file when a reference is enough.

Maintenance

• Review top-level instructions when quick commands, workflow defaults, safety policy, the persona skill, or the skill catalog changes.
• When the persona skill changes, re-sync the inline persona block in the source-of-truth file in the same pass. This is the deliberate duplication from "Persona placement", and it only stays correct if it is refreshed deliberately.
• If the repo adds or removes important skills, update both the skill inventory and the routing hints in the source-of-truth file.
• Keep instruction files aligned with generated policy files and provider settings when the repo uses them.

Validation

• The repo has one clear instruction source of truth.
• Bridge files stay thin unless a provider-specific exception is genuinely required.
• Top-level instructions contain durable repo workflow and routing, not duplicated domain detail.
• The persona core is inline in the source-of-truth file, not merely pointed at, and it still matches the persona skill it is projected from.
• The instruction files still match the current skills, commands, and policy model.

References

• Read ./references/providers/copilot-instructions.md for Copilot-specific source-of-truth guidance.
• Read ./references/providers/gemini-instructions.md for GEMINI.md bridge guidance.
• Read ./references/providers/claude-instructions.md for .claude/CLAUDE.md bridge guidance.
• Read ./references/global-instructions.md for user-level/global instruction files and how the bridge pattern applies to personal defaults.
• Read ./references/agents-md-standard.md for the root AGENTS.md convention, nesting/monorepo behavior, and symlink migration.
• Read ./references/checklist.md for a quick multi-provider instruction review pass.
• Read ./references/import-bridge.md when choosing between thin stubs, bridge files, and rare split-source patterns.
• Read ./assets/trigger-eval-queries.example.json when testing trigger quality for instruction-authoring prompts.
• Review ./evals/evals.json when validating output quality for source-of-truth and bridge recommendations.