Al Agent Instructions

Claude Code adapter — generated by scripts/sync-plugin-support.js, do not edit. Canonical source: prompts/al-agent.instructions.prompt.md (sha256 1080481dc7f897b6a67f06ab8da8260d8bd67dddfdf34e3091a351711be64044). This workflow was written for @workspace use … in GitHub Copilot Chat. Apply these harness mappings and treat everything else as binding:

Copilot surface in the contract	Claude Code equivalent
#runSubagent / runSubagent / agents:	The Task tool with subagent_type: "aldc:<agent>"; one invocation carries everything inline (subagents are stateless)
#askQuestions / vscode/askQuestions	AskUserQuestion (main session only; a subagent returns its questions in its structured output)
handoffs: entries, and send: false on one	No equivalent — the gate is yours to keep. In Copilot a handoff is a button the human clicks, and send: false additionally hands them the prompt to review before it is sent; the host supplies the approval. Claude Code delegates through the Task tool, with no click and no review step. So never auto-delegate: present your output, get explicit approval, and only then delegate or route to another role. The handoff's routing is carried by the plugin's role entry skills; its human gate is carried by this rule
@al-architect, @al-spec-agent, @al-conductor, @al-developer, @al-developer-reviewer, @al-triage, @al-presales, @al-agent-builder, @dredd	/aldc:architect, /aldc:spec, /aldc:conduct, /aldc:develop, /aldc:review, /aldc:triage, /aldc:presales, /aldc:agent-builder, /aldc:audit (or claude --agent aldc:<agent>)
@workspace use <workflow>	/aldc:<workflow> with dots replaced by dashes (al-spec.create -> /aldc:al-spec-create); already rewritten below
${input:Name} prompt variables	Take the value from $ARGUMENTS; ask through AskUserQuestion when missing. Already rewritten below as <Name from $ARGUMENTS>
execute / runInTerminal, AL compilation or symbol restore	Developer/Implementer: discovered official AL MCP mcp__al__al_compile / mcp__al__al_build / mcp__al__al_downloadsymbols, or an installed ALTool/AL-Go runner through Bash. Use the correct App/Test project; unavailable capabilities are never simulated. Publishing retains its separate human/CI gate and is not granted by this adapter
edit, read/readFile, search, #codebase	Edit/Write, Read, Grep/Glob
#changes, changes, search/changes	Read-only git status / git diff through Bash only where granted; otherwise read the caller-supplied diff artifact or return the missing evidence request
read/problems, al_get_diagnostics, #testFailure	Discovered mcp__al__al_getdiagnostics for the explicit project, or actual compiler/test output. Diagnostics are not evidence of a fresh compile or executed tests. No result: record not-run / unavailable
al-symbols-mcp/*, upstash/context7/*, microsoft-learn/*, microsoft-docs/*	Tools named mcp__al-symbols-mcp__*, mcp__context7__*, mcp__microsoft-docs__* (or the mcp__plugin_aldc_… form). A server that is not loaded is unavailable; do not simulate it
sshadowsdk.al-lsp-for-agents/*, #usages, semantic definition/reference queries	Claude LSP with the existing AL LSP for Agents companion; discover the actual operations and project context first. VS Code tool identifiers are not Claude tool names
ms-dynamics-smb.al/*, vscode.mermaid-chat-features/*	VS Code tool identifiers do not exist here. Use only a separately discovered Claude capability; never rename community tools as Microsoft tools
vscode/memory, vscode/*	Not available; keep state in the canonical .claude/plans/ artifacts of the project
todo	TodoWrite for in-session tracking; durable state stays in .claude/plans/
"load skill-x"	The Skill tool (aldc:<skill-name>) or read ${CLAUDE_PROJECT_DIR}/.claude/skills/<skill-name>/SKILL.md
.github/instructions/*.instructions.md (applyTo)	${CLAUDE_PROJECT_DIR}/.claude/rules/*.md (paths:); /aldc:al-initialize copies them to the project's .claude/rules/ for auto-apply in the main session; inject inline for subagents exactly as the contract says
.github/skills/, .github/plans/…, docs/templates/	${CLAUDE_PROJECT_DIR}/.claude/skills/, .claude/plans/… (project-owned, aldc.yaml -> plans.root), ${CLAUDE_PROJECT_DIR}/claude-plugin/docs/templates/; already rewritten below
python … runtime helpers, --aldc-root	Paths already point at ${CLAUDE_PLUGIN_ROOT}; use python or python3, whichever exists (stdlib only)
aldc.yaml	${CLAUDE_PROJECT_DIR}/aldc.yaml when the project has one, otherwise ${CLAUDE_PROJECT_DIR}/claude-plugin/aldc.yaml

Read ${CLAUDE_PROJECT_DIR}/claude-plugin/docs/claude-al-tooling.md for this surface’s LSP/MCP bindings, then ${CLAUDE_PROJECT_DIR}/.claude/skills/skill-migrate/references/cli-al-tools.md for the shared role/evidence contract. The Claude guide supersedes older capability-availability examples, never role responsibilities or human gates.

Workflow: Generate Agent Instructions

Applies the Responsibilities-Guidelines-Instructions framework from skill-agent-instructions to produce a runtime-ready instruction file. This prompt does not contain the framework — it executes it.

Load skill first: skill-agent-instructions (framework, keywords table, validation checklist, anti-patterns).

Output modes

• Designer mode → text ready to paste into the Agent Designer wizard
• SDK mode → text stored in .resources/Instructions/InstructionsV1.txt, loaded via NavApp.GetResourceAsText() returning SecretText, applied via Agent.SetInstructions(UserSecurityId, InstructionsText)

Workflow

1. Gather inputs

From the developer:

1. Agent purpose — one-sentence business goal
2. Pages in scope — pages included in the agent's profile
3. Fields to read/write — exact field names from those pages
4. Actions to invoke — exact action captions on those pages
5. Decision criteria — business rules for branching (if/then)
6. Output format — how should the agent report results?
7. Safety gates — which actions need user intervention?

2. Draft RESPONSIBILITY

One sentence with business domain + outcome. No paragraphs.

3. Draft GUIDELINES

3–7 cross-task rules. Required: at least one ALWAYS, one DO NOT, one safety gate for critical actions, one data-access boundary (read-only vs read-write).

4. Draft INSTRUCTIONS per task

For each task:

1. Start with navigation (how does the agent reach the data?)
2. Read and MEMORIZE context before decisions
3. Apply business logic with explicit if/then branches
4. Execute actions using proper keywords (see skill-agent-instructions keywords table)
5. Report results with structured Reply format
6. Include error handling at the end of the task

5. Validate

Run through the validation checklist from skill-agent-instructions. Key items:

• Page/field/action names match agent profile exactly
• MEMORIZE placed BEFORE value is needed in later steps
• MEMORIZE includes example format
• Critical actions gated by user intervention or review
• Written in English
• Environment-agnostic (no hardcoded company names, URLs, user IDs)
• Error handling section per task
• No references to "Tell Me"
• Guidelines count is 3–7
• Stored in .resources/Instructions/InstructionsV1.txt (SDK mode)

6. Test and iterate

Deploy → observe timeline → refine. Apply the "less is more" principle.

Reference

For the framework details, keyword tables, anti-patterns, and complete examples, consult skill-agent-instructions and ${CLAUDE_PROJECT_DIR}/.claude/skills/skill-agent-instructions/examples/.

Skills Evidencing

End with:

Copy code
**Skills loaded**: skill-agent-instructions
**Patterns applied**:
- RGI framework — Responsibilities/Guidelines/Instructions structure
- Keywords applied: {list keywords used: Navigate to, Memorize, Reply, etc.}
- Safety gates: {list user intervention / review points}