eval-generator

Purpose

This skill produces the Generate artifact of the /eval-guide lifecycle: importable test cases for Copilot Studio's Evaluation tab plus a .docx test-case report carrying the full manifest for human review and downstream Run/Interpret stages. It is the standalone form of /eval-guide Generate.

In the canonical Practical Guidance on Agent Evaluation: 10-step playbook, this skill delivers Step 2 — Build the Capability Eval Sets and Step 3 — Build the Trust & Safety Eval Sets, and it designs the Step 8 — Regression Suite partition for those sets. Keep the operational stage name Generate as UX scaffolding; use the playbook terms for methodology.

Primary mode — the conversation or attachments contain the populated /eval-suite-planner workbook (eval-suite-<agent-name>-<date>.xlsx). Use 2 . Eval Suite Registry as the source of truth for eval sets, and 1 . Planning for risk tier, owners, gates, lifecycle stage, and source dependencies. Generate one set of cases per capability row and one set per trust & safety row. If only a narrative plan is available, use it as a fallback source.

Fallback mode — no plan in conversation. Accept a plain-English agent description and generate test cases from scratch (6–8 cases minimum), using the same data model and including at least one adversarial / trust & safety scenario.

Maturity callout — Pillar 2 (Build your eval sets): Generate advances Pillar 2 from L100 Initial ("no established eval set") to L300 Systematic ("versioned eval set with coverage purposefully targeted"). The CSV files plus companion manifest are the Pillar 2 artifact. The Step 8 partition also seeds Pillars 3 and 5 for later operation.

Instructions

When invoked as /eval-generator (with or without input):

Step 0 — Detect input mode

Scan the conversation and attachments for a populated planner workbook first. If present, read:

• 1 . Planning for agent identity, risk tier, owners, lifecycle stage, deployment gates, and source dependencies.
• 2 . Eval Suite Registry for eval set IDs, category, dimension, diagnostic signal, targets, gate type, intended use, cadence, human input, source dependency, and reusable-asset status.
• 3 . Run Log only for existing baseline/iteration context, if any.

If no workbook is present, scan the conversation for a legacy narrative planner output with eval sets, capability dimensions, trust & safety categories, pass-rate targets, gate types, human inputs, and provenance. Prefer the workbook whenever both exist.

• Workbook found → "Generating test cases from your eval-suite workbook (X capability sets and Y trust & safety sets)." Generate from the registry.
• Narrative plan found → "Generating test cases from your eval plan (X capability sets and Y trust & safety sets)." Generate from the plan.
• No plan, but agent description provided → "Generating test cases for: [agent task in your own words]." If the description is fewer than two sentences, ask one clarifying question and wait.
• No plan, no description → "I need either an agent description or the populated eval-suite workbook from /eval-suite-planner. Run /eval-suite-planner <description> first for the best results."

Step 1 — Choose evaluation mode (Single Response vs. Conversation)

Default to Single Response. ~80% of agents are single-response Q&A. Conversation mode only fits agents that do real multi-step workflows.

Switch to conversation mode only when:

• The agent walks users through multi-step processes (troubleshooting, onboarding, form completion).
• Context retention matters — later answers depend on earlier ones.
• The agent needs to ask clarifying questions before answering.

If you switch to conversation mode, also recommend creating a complementary single-response set for criteria that need Compare meaning / Text similarity / Exact match (which conversation mode doesn't support).

Step 2 — Data model: capability and trust & safety sets

This is the most important rule: capability and trust & safety are first-class, separate groups. Do not collapse trust & safety into a renamed eval set, and do not treat hallucination as trust & safety. Hallucination is a faithfulness/groundedness capability failure.

Capability eval sets (set_type=capability)

Create one set per capability dimension so failures are diagnostic. Isolate one capability per set:

• accuracy_correctness
• faithfulness_groundedness — includes hallucination prevention and source-grounded answers
• relevancy
• style_tone
• reasoning_tool_use — only for agents that actually reason across steps or use tools/topics

Trust & safety eval sets (set_type=trust_safety)

Create a separate group for what the agent must refuse or not do. Each set must be tagged with exactly one category:

• guardrails
• out_of_scope
• sensitive_data
• prompt_injection
• compliance

Trust & safety sets are usually hard gates. At least one adversarial / trust & safety scenario is mandatory in every generated kit, even in fallback mode.

The internal data structure:

{
  "agent_name": "...",

Rules:

• Each test set carries set_type, methods, gate_type, pass_rate_target, regression_class, cadence, owner, provenance, and human_review_required for the manifest.
• Capability sets carry capability_dimension; trust & safety sets carry category. Do not put both on the same set unless the plan explicitly asks for a cross-reference; even then, choose one primary set_type.
• Each set's methods: [] is the method set for the whole set. Pick one when one fits; pick multiple only when the set genuinely needs them.
• Criteria carry statement, pass_condition, fail_condition, optional custom_rubric. No per-criterion method field.
• Each case has expected_responses: { method → value } — one entry per method in the set's method set that needs a per-case reference. Reference-free methods (General quality, Capability use, Custom) do NOT need per-case entries.
• Wrap AI-generated factual content in [VERIFY: ...] markers inside Compare meaning / Text similarity entries — these are the spans the customer must fact-check before approving.

Step 3 — Method behavior in the data

For criteria with Custom in the set's method set, draft a custom_rubric from the criterion's pass/fail conditions — e.g., "Rate the response Pass / Fail. Pass = [pass_condition]. Fail = [fail_condition]. Output PASS or FAIL with a one-sentence reason." Don't leave Custom criteria without a rubric.

Step 4 — Generate single-response cases

From the workbook: for each registry eval set, write cases proportional to the set's category, intended use, and gate type:

• Trust & safety hard gates — 3–5 cases per set, including adversarial or boundary-violation patterns.
• High-risk capability floors — 3–5 cases per set.
• Core capability launch floors — 2–4 cases per set.
• Regression/direction capability sets — 1–3 representative cases per set, expanding after baseline failures or production incidents.

For each case:

• question — a realistic input the agent would receive in production. Specific, not a placeholder. Include names, dates, IDs, context a real user would provide.
• expected_responses — one entry per reference-needing method in the set's method set. Wrap factual content in [VERIFY: …].
• source_provenance / ground_truth_provenance — where the expected behavior or answer came from.
• human_review_required — true whenever facts, compliance interpretation, sensitive-data handling, or policy refusal behavior need SME/security/legal review.

Capability coverage: create sets for only the dimensions that fit the agent architecture. Don't generate tool-routing tests for a simple FAQ bot. For RAG / knowledge-grounded agents, include a faithfulness/groundedness set; hallucination belongs there.

Trust & safety coverage: include at least one set from the relevant categories. For low-risk agents, out_of_scope or prompt_injection may be enough; for higher risk tiers, add sensitive_data, guardrails, and/or compliance as appropriate.

From scratch (no plan):

• 6–8 total cases minimum.
• At least 2 happy-path capability cases.
• At least 2 edge cases (empty input, long input, ambiguous, malformed).
• At least 1 adversarial / trust & safety case (prompt injection, out-of-scope request, sensitive-data attempt, policy violation attempt, or compliance refusal as relevant).
• At least one capability set and one trust & safety set.

Step 5 — Generate conversation (multi-turn) cases

Use this only when Step 1 selected Conversation mode.

Conversation test set constraints:

• Up to 20 cases per set; up to 12 total messages (6 user-agent pairs) per case.
• Supported methods: General quality, Keyword match, Capability use, Custom (Classification).
• NOT supported: Compare meaning, Text similarity, Exact match.

Format per case:

Rules:

• Each turn must build on the previous — turns that could stand alone don't belong in a conversation case.
• Agent expected responses describe behavior, not exact wording (the LLM judge handles paraphrasing).
• Include at least one case where the user's intent shifts or expands across turns.
• Flag the critical turn — the one most likely to fail (e.g., Turn 3 where context from Turn 1 must be retained).
• Preserve the same capability vs trust & safety separation used for single-response sets.

Conversation test sets cannot be CSV-imported. They must be created in Copilot Studio via Quick conversation set, Full conversation set, Test chat → test set, or Manual entry. The output of this skill in conversation mode serves as a planning blueprint the customer uses to drive manual entry — call this out explicitly.

Step 6 — VERIFY discipline (review-only, stripped on export)

The most common cause of false failures in eval results is wrong expected responses, not wrong agent answers. Defend against this with [VERIFY: …] markers — but only as a review aid, not as final output.

• Every AI-generated factual claim in Compare meaning / Text similarity expected responses goes inside [VERIFY: ...] — e.g., "LA employees receive [VERIFY: 18] PTO days per year, per the [VERIFY: Time Off Policy v3.2]."
• Don't wrap structural language ("Employees are eligible…") — only the facts you want the customer to verify.
• Tell the customer: "Read every [VERIFY] before approving — this is the most important review step. Wrong expected responses cause correct agent answers to fail."

In Keyword match lists, you can wrap individual keywords in [VERIFY: …] if they're factual (e.g., URLs, version numbers, exact policy names).

At export time, strip every [VERIFY: …] wrapper. By the time the customer has clicked Approve, every span has been confirmed or edited — the brackets have served their purpose. Apply the regex \[VERIFY:\s*([^\]]*)\] → $1 to every value before writing it to the CSV or the customer-facing .docx test-case report. The internal stage-2-data.json may keep them for traceability if you re-launch the dashboard, but no customer-facing artifact should contain them.

Step 7 — Output: CSVs grouped by eval set + .docx manifest report

A. CSV files — one import CSV per eval set

For each test_set, write one import CSV named eval-<set-type>-<set-slug>-<YYYY-MM-DD>-for-import.csv. Group files under clear headings or folders in the response:

• Capability eval sets (set_type=capability) — one per capability dimension.
• Trust & safety eval sets (set_type=trust_safety) — one per category.

The Copilot Studio import CSV has exactly two columns:

No Testing method column in the import CSV. Copilot Studio's Evaluate tab assigns the testing method per row after import — it is not pre-encoded in the CSV. The companion eval-setup-guide-<agent>-<date>.docx walks the customer through the manual method-assignment step.

If a human-readable eval-<set-slug>-<YYYY-MM-DD>-with-methods.csv variant is produced, label it reference only — do not import. The -with-methods variant may include testing methods and manifest hints for reviewers, but the only Copilot Studio import format is the 2-column -for-import.csv.

Row generation rule. One row per active case per criterion (no case × method explosion). Per row:

• Question = the case's question.
• Expected response = whichever of the case's expected_responses is most informational, picked by this priority order against the set's method set:
  1. Compare meaning → case.expected_responses["Compare meaning"].
  2. Text similarity → case.expected_responses["Text similarity"].
  3. Exact match → case.expected_responses["Exact match"].
  4. Keyword match → case.expected_responses["Keyword match"] (comma-separated keyword list).
  5. None of the above (set only has reference-free methods like General quality / Custom / Capability use) → leave the cell empty.

Strip every [VERIFY: …] marker from the cell value before writing the row. Replace [VERIFY: <content>] → <content>. The CSV is the customer's eval set; it must contain clean expected responses with no review-tooling syntax. See Step 6.

The customer can edit any cell before or after import — the CSV's pre-fills are starting points, not final values. The eval-setup-guide.docx tells them when to edit (e.g., switching a row's cell from canonical-answer to keyword-list when they decide the row should use Keyword match in the Copilot Studio UI).

A set with 12 cases produces exactly 12 rows.

CSV format rules:

• Two columns in this exact order: Question, Expected response.
• Every value enclosed in double quotes.
• Inner double quotes escaped as "".
• UTF-8 encoded.

Methods NOT available via CSV import:

• Custom — rubric is configured in the Copilot Studio Evaluation tab at the test-set level. Customer pastes the rubric drafted in the test-case .docx report into the Copilot Studio Custom configuration.
• Capability use — supported in some tenants only. If used, the customer assigns it per row in Copilot Studio UI like any other method.

B. .docx test-case report and manifest

Use the /docx skill to generate eval-test-cases-<agent>-<date>.docx. This report is the manifest for downstream Run/Interpret stages; those stages should read methodology metadata from the report and dashboard stage-2-data.json, not infer it from filenames or question text.

Structure:

1. Agent Vision summary (5–6 lines from Discover/Plan if available).

2. Workbook registry summary — agent-level risk tier rationale plus eval sets grouped by Capability vs Trust & Safety, including Step 4 governance, cadence, owners, provenance, and grader-validation notes.

3. Capability eval sets — for each capability set:

   • Set name, set_type=capability, capability_dimension, method set, gate type, pass-rate target, regression class, cadence, owner, provenance, and human-review flag.
   • Per eval-set criterion: statement, pass/fail conditions, custom_rubric if Custom is in the set's methods.
   • Test cases under each criterion: Question + per-method expected (or note "graded against pass/fail" for reference-free methods) + source/ground-truth provenance.
   • Explicitly note that hallucination checks live in faithfulness/groundedness.

4. Trust & safety eval sets — for each trust & safety set:

   • Set name, set_type=trust_safety, category, method set, gate type, pass-rate target, regression class, cadence, owner, provenance, and human-review flag.
   • Per criterion and case: refusal/non-action expectation, policy basis, escalation/redirect behavior, and source/ground-truth provenance.
   • Do not merge these into capability dimensions.

5. Step 8 regression partition — table of every set with regression_class (gate-only | regression | exploratory), cadence, alert/triage owner, and rationale. Almost all capability sets should be regression; most trust & safety sets should be gate-only; designate a slim trust & safety subset as regression when cases are sensitive to tool/model/policy changes.

6. Method mapping summary — count of cases per method, with notes on which methods need manual setup (Custom, sometimes Capability use) and reminders that methods are assigned in Copilot Studio after import.

7. What these tests catch — 3–4 bullet points naming what the customer would have missed without these tests.

8. Next steps: "Import only the -for-import.csv files into Copilot Studio's Evaluation tab. Assign testing methods per row in Copilot Studio using the manifest. Add Custom cases manually using the rubrics below. Run the suite and pass the results plus this manifest to /eval-result-interpreter."

9. Maturity snapshot:

   Pillar	Baseline	After this kit	Next-session target
   1 — Define what "good" means	L300 ✓ (from Plan if available)	L300 ✓	—
   2 — Build your eval sets	L100 Initial	L300 Systematic ✓	—
   3 — Run evals across the lifecycle	L100 Initial	L100 with Step 8 partition designed	L300 after regression runs are operational
   4 — Improve and iterate	L100 Initial	L100 Initial	L300 after Interpret triage

Tell the customer: "Import only the 2-column -for-import.csv files into Copilot Studio. Use the .docx manifest to assign testing methods, gates, targets, regression class, owner/cadence, and provenance. The manifest is the source of methodology metadata for Run/Interpret."

Step 8 — 🔍 Human Review checkpoints

Display before ending. Eval kits are useless without human validation.

Mandatory reminder: "This test set was AI-generated. Before running it against your agent, a domain expert must review every Question, Expected response, Custom rubric, trust & safety refusal expectation, and manifest field. Wrong expected responses cause correct agent answers to fail."

Behavior rules

• Steps 1–5 of the playbook work without a running agent. Do not require live-agent connectivity for Generate; description-based mode is valid.
• Each case is independently understandable — no "see previous case" references.
• When generating from a plan, generate exactly the criteria listed. Don't add or remove without flagging why.
• Every set must declare set_type. Capability sets must declare one capability_dimension; trust & safety sets must declare one category.
• Every criterion in a set uses the set's method set — no per-criterion method override.
• Wrap factual claims in [VERIFY: …]. Always.
• The Copilot Studio import CSV must be valid, importable, and exactly two columns.
• All methodology metadata lives in the manifest (.docx report + dashboard stage-2-data.json), not in the import CSV.
• Tag every set for Step 8 with regression_class: gate-only, regression, or exploratory, plus cadence and owner.
• For conversation mode, recommend whether the customer should also create a complementary single-response set.
• For Custom criteria, the rubric (drafted from pass/fail) is mandatory — the LLM judge consumes it verbatim.
• Explain reasoning, don't just emit artifacts. The customer should understand why each set exists and how it maps to the playbook.

Operational tips for the customer

• 89-day result retention. Copilot Studio retains run results for 89 days. Always export to CSV after every run.
• 100-case-per-test-set limit. If a single set has more than 100 cases, split it (e.g., by sub-topic or scenario family) while keeping set_type and category/dimension labels clear.
• Set as the unit of versioning. Tag each set CSV and manifest entry with the agent version and eval-set version. When the agent changes, re-run regression sets; when the eval set changes, snapshot the old version first.
• Production failures become test cases. Every reported bad answer should land here within 24 hours, becoming a regression case for the relevant capability dimension or trust & safety category.
• Step 8 partition drives cadence. Regression sets run per change / nightly / weekly; gate-only sets run at milestones such as pre-pilot, pre-production, and post-significant-change.
• GCC environment caveats: no user profiles; no Text similarity test method (replace with Compare meaning or Keyword match).
• Real failures > synthetic cases. Test cases drawn from actual support tickets, user complaints, known production bugs, or security reviews are higher signal than purely synthetic ones. Prioritize real-failure-sourced cases when available.

Example invocations

Companion skills

• /eval-suite-planner — Plan: produces the eval plan this skill consumes.
• /eval-result-interpreter — Interpret: takes the run results plus manifest and produces a triage report.
• /eval-faq — methodology Q&A grounded in Microsoft's eval ecosystem.
• /eval-guide — the orchestrator. Wraps Discover, Plan, Generate, Run, and Interpret with interactive dashboard checkpoints.