Maintaining AI Context (Supersubset)
Supersubset keeps agent/playbook context in .github/ and docs/ (not a separate .ai/ tree). This skill merges the “how to edit” and “how to audit” responsibilities into one place.
Scope (what this skill owns)
| Artifact | Role |
|---|---|
features/*.feature.yaml |
Acai-compatible product requirements (ADR-011) |
features/terms/*.term.yaml |
Glossary terms for requirement prose |
.github/skills/*/SKILL.md |
How-to playbooks |
.github/agents/*.agent.md |
Subagent definitions + tool allowlists |
AGENTS.md |
Orchestrator + recovery + pointers |
.github/copilot-instructions.md |
Lean always-on Copilot summary |
.github/skills/github-cli/SKILL.md |
Canonical gh patterns for GitHub operations |
docs/bootstrap.md |
Fresh-session orientation |
docs/status/* |
Master plan, risks, phase summaries, checkpoints |
docs/testing/* |
Verification strategy, QA lists, browser plans |
When to add a new skill
- A repeatable procedure is being reinvented across issues (orchestration already points to patterns — extract when stable).
- Triggers belong in the YAML
description:(one line, “Use when …”). - Keep body focused; link to
docs/for long reference material. - Add a See also back-link from related skills or
AGENTS.mdif discoverability is low.
When to add or change an agent file
- Package boundaries and tool lists stay accurate (
tools:, MCP tool names). - Handoffs must reference agents that exist in
.github/agents/. - Keep
user-invocableaccurate for Copilot Agents UX.
Audit passes (targeted “ai-context-review”)
Run these when context feels bloated or agents misfire.
Pass A — Structural
- Every skill referenced from
AGENTS.md/orchestration/bootstrapexists. - Agent
handoffs→ existing agent files. -
docs/bootstrap.mdfile tree matches repo (update counts/paths when packages move).
Pass B — Signal vs noise
-
copilot-instructions.mdstays short; deep content lives in skills ordocs/guides/. - Skills avoid duplicating
initial-spec.mdinvariants — reference instead. - No stale ports/URLs/commands — compare to
package.jsonscripts andplaywright.config.ts.
Pass C — Duplication
- One source of truth per workflow (e.g. verification commands not copy-pasted in five skills — cross-link
branch-ci-promotion).
Pass D — Discoverability
- New skills linked from
bootstrap.md“read first” or orchestration “See also” as appropriate.
Size guidance
| Artifact | Heuristic |
|---|---|
copilot-instructions.md |
Keep roughly under ~80 lines of dense lists |
| Single skill | Split beyond ~300 lines; add docs/guides/ for narrative |
AGENTS.md |
Index + protocols; move long prose into skills |
Anti-patterns
- Duplicating the same verification checklist in every skill — use
branch-ci-promotion. - Putting architecture essays in
copilot-instructions.md— use ADRs + schema docs. - Orphan skills never linked from bootstrap, AGENTS, or another skill.
See also
.github/skills/github-cli/SKILL.md.github/skills/orchestration/SKILL.md.github/skills/branch-ci-promotion/SKILL.mddocs/bootstrap.md
Source: wandir-tech/supersubset — distributed by TomeVault.