Skip to content

copilot.md

Latest commit

 

History

History
90 lines (63 loc) · 7.19 KB

copilot.md

File metadata and controls

90 lines (63 loc) · 7.19 KB
Raw

SpecKit with GitHub Copilot

GitHub Copilot reads its configuration from .github/. Running specify init --integration copilot wires SpecKit into Copilot by projecting everything — commands, orientation, skills, rules, and memory — under that one root. The wrinkle unique to this harness: each /speckit.* command is projected twice — once as a custom chat-mode under .github/agents/ and once as a slash-prompt under .github/prompts/ — so you can invoke a command either way in Copilot Chat.

Initialize

specify init my-app --integration copilot
cd my-app

You get the nine /speckit.* commands (as both agents and prompts), a .github/copilot-instructions.md orientation file, the process skills and rules under .github/, a seed memory index, and the shared .speckit/ runtime. --here initializes the current directory instead of creating my-app; --force merges into a non-empty directory; re-running init never clobbers accumulated memory (the seed MEMORY.md is written skip-if-exists).

What init projects

Artifact Path
/speckit.* commands (chat-modes) .github/agents/speckit.<cmd>.agent.md
/speckit.* commands (slash-prompts) .github/prompts/speckit.<cmd>.prompt.md
Orientation file .github/copilot-instructions.md
Process skills .github/skills/<skill>/SKILL.md
Rules .github/rules/<rule>.md
Memory .github/memory/MEMORY.md (seed index)
Review subagents none (Claude-only — see below)

Alongside these, init writes the shared .speckit/ runtime (the constitution, the spec/plan/tasks/checklist templates, extensions.yml, and — after a green verify — the per-spec lock under .speckit/lock/<target>/<spec-id>.json). The runtime is identical for every harness and holds no shell scripts; the command logic lives in the specify binary.

The orientation file

Copilot's orientation file is .github/copilot-instructions.md. Unlike Claude's CLAUDE.md, it does not auto-import rules and memory — it uses a read-at-start directive that points the agent at .github/memory/MEMORY.md, and it references the rule files as prose (not @import lines). At the start of a task, Copilot reads the memory index and any topic files it points to, and follows the named rules.

## Rules

Follow the conventions in `.github/rules/`: `code-quality.md`,
`commit-discipline.md`, and `spec-conventions.md`;
`enforcement-hierarchy.md` governs where a new convention lives.

## Project memory

At the start of a task, read the project memory index —
`.github/memory/MEMORY.md` — and any topic files it points to. It's
durable, agent-owned working knowledge about this repo (not spec truth; the engine
never reads it). Maintain it with the `managing-memory` skill.

Driving the /speckit.* commands

Each command is projected two ways, so you have two ways to invoke it in Copilot Chat:

  • Slash-prompts (.github/prompts/speckit.<cmd>.prompt.md) are invoked as slash commands — type /speckit.specify, /speckit.plan, and so on directly in chat.
  • Custom agents / chat-modes (.github/agents/speckit.<cmd>.agent.md) are the same commands as selectable Copilot agents (chat-modes), with frontmatter name: speckit.<cmd>.

Both forms carry the identical, agent-neutral prompt body — the projection just exposes it through both of Copilot's invocation surfaces. Run the specify engine (scan, verify, drift, parity) in your terminal.

Command What it does
/speckit.specify Author a new feature folder from a description (narrative, prioritized stories, models, view-models, errors).
/speckit.clarify Resolve [NEEDS CLARIFICATION] markers by asking up to five targeted questions and writing the answers back into the specs.
/speckit.analyze Read-only cross-artifact consistency/quality analysis of a feature; reports by severity, never edits.
/speckit.checklist Generate a requirements-quality checklist ("unit tests for the spec").
/speckit.constitution Create or update the project constitution and propagate the change.
/speckit.plan Produce a per-target technical plan for a feature.
/speckit.tasks Build an ordered, story-prioritized task list for a target; each task maps to a spec ID.
/speckit.implement Implement on a target: failing tests first, layered review, adversarial pass, then verify-and-lock.
/speckit.taskstoissues File a feature's task list as GitHub issues on the repo matching the git remote. (The one command that touches GitHub.)

Skills, rules, and memory

Process skills (.github/skills/<skill>/SKILL.md) — eight discipline skills the commands lean on: test-driven-development (RED/GREEN), verification-before-completion, adversarial-review, systematic-debugging, implementing-a-spec, brainstorming-feature, writing-user-stories, and managing-memory.

Rules (.github/rules/<rule>.md) — four always-relevant conventions, referenced from the orientation file: code-quality, commit-discipline, spec-conventions, and enforcement-hierarchy (the standard for deciding where a new convention should live).

Memory (.github/memory/MEMORY.md) — committed, repo-local, agent-owned working knowledge about this project: not spec truth, never verified or gated, and the engine never reads it. MEMORY.md is a one-line-per-topic index loaded each session; topic files are read on demand. Maintain it with the managing-memory skill. The test: a statement of required behavior belongs in the spec library; knowledge that helps an agent work here but isn't the behavior under test belongs in memory; anything already in code, git, or the specs belongs in neither. See the agent-memory design.

No review subagents. Copilot gets none. The .github/agents/ directory here holds the command chat-modes, not dispatched reviewers — the review subagents (spec-reviewer, test-gap-finder, drift-hunter, handoff-builder, visual-verifier) are a Claude Code dispatch concept and are projected only for the claude integration.

What's different for Copilot

  • One root. Everything lands under .github/ — Copilot's native config root — rather than a tool-specific top-level dir.
  • Dual command projection. Every /speckit.* command exists as both a .github/agents/*.agent.md chat-mode and a .github/prompts/*.prompt.md slash-prompt. No other harness double-projects.
  • .github/agents/ is for commands, not reviewers. It holds the command chat-modes; the five review subagents are Claude-only.
  • Read-at-start orientation, not native import. .github/copilot-instructions.md directs the agent to read .github/memory/MEMORY.md and references the rules as prose — unlike CLAUDE.md, which auto-imports both every session.

Next