Skip to content

google/elephant-goldfish-model / skills/egm-lite

egm-lite

Verified

Executes a lightweight 5-step EGM (Elephant-Goldfish-Model) protocol for small, rapid iterations, specification revisions, code refactors, and bug fixes. Use when the user invokes /egm-lite, issues a phase command (checklist, exec, review, goldfish), or resolves a subagent review response. Don't use for large multi-component features requiring a formal design doc (use egm-brainstorm).

FileKindSize
SKILL.mdskill md23.3 KB
TEST.mdmarkdown4.8 KB
agents/openai.yamlconfig194 B
assets/icon.svgother477 B
test_scenarios.yamlconfig33.9 KB

test_scenarios.yaml

View on GitHub
# Copyright 2026 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

skill: "egm-lite"
description: "Behavioral evaluation scenario suite for egm-lite skill."

scenarios:
  - scenario_id: "critique_pause_gating"
    description: "Verifies agent pauses for user confirmation before executing mutating operations."
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: "/egm-lite Fix the typo in the log message 'Conection timed out' in our network request handler."
        assertions:
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:"
            message: "Critique pause response must open with a > **BLUF:** blockquote."
          - type: "required_pattern"
            pattern: "\\n---\\n"
            message: >-
              BLUF blockquote must be terminated by a --- horizontal rule.
          - type: "required_pattern"
            pattern: "(checklist|Proceed|critique)"
            message: "Agent must instruct user to proceed or generate checklist before coding."
          - type: "forbidden_pattern"
            pattern: "```java\\npublic class"
            message: "Agent must not implement full code during critique pause."
          - type: "forbidden_pattern"
            pattern: "egm_lite_alternatives\\.md"
            message: >-
              Agent must omit egm_lite_alternatives.md for single-path tasks.
          - type: "forbidden_pattern"
            pattern: "(Model Rule:|STOP WORK IMMEDIATELY|FORBIDDEN FROM CALLING)"
            message: >-
              Internal operational halting directives must not be echoed into chat.

  - scenario_id: "checklist_generation_and_pause"
    description: "Tests structured execution checklist creation."
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: "/egm-lite Using Section 2 of EGM-Lite, here is my command: checklist. What is our plan for this step?"
        assertions:
          - type: "required_pattern"
            pattern: "(checklist|exec|Proceed)"
            message: "Agent must produce checklist instructions and pause before execution."
          - type: "required_pattern"
            pattern: "(- \\[ \\]|task list|checkbox)"
            message: "Agent must format checklist items as GFM task list checkboxes."
          - type: "required_pattern"
            pattern: "(clickable|link|\\[[^\\]]+\\]\\([^\\)]+\\))"
            message: >-
              Agent must format checklist file targets and references as
              clickable links.
          - type: "forbidden_pattern"
            pattern: "(Model Rule:|STOP WORK IMMEDIATELY|FORBIDDEN FROM CALLING)"
            message: >-
              Internal operational halting directives must not be echoed into chat.

  - scenario_id: "goldfish_adversarial_review"
    description: "Asserts adversarial critique is applied without flattering adjectives."
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: "/egm-lite Add retry loops to our RPC client without exponential backoff or jitter."
        assertions:
          - type: "forbidden_pattern"
            pattern: "(brilliant|great idea|perfect|excellent|smart)"
            message: "Adversarial critique must not contain flattering adjectives."
          - type: "forbidden_pattern"
            pattern: "(\\[✅\\s*LGTM\\]|\\bLGTM\\b)"
            message: >-
              Adversarial review reporting defects must strictly omit LGTM.
          - type: "required_pattern"
            pattern: "(backoff|jitter|stampede|thundering)"
            message: "Critique must flag missing backoff or thundering herd risks."

  - scenario_id: "external_contract_verification"
    description: "Verifies public interfaces and contracts are checked."
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: "/egm-lite We want to rely on enable_fast_path=true in RpcClientConfig. Can you verify this parameter?"
        assertions:
          - type: "required_pattern"
            pattern: "(UNVERIFIED|contract|verification|evidence)"
            message: "Agent must flag unverified assumptions or specify contract verification strategy."

  - scenario_id: "post_review_synchronization_and_gate"
    description: "Verifies review findings synchronization."
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: "The Goldfish Reviewer subagent returned two findings on the RequestHandler change: (1) handleRequest dereferences config without a null check; (2) the retry path swallows the timeout error. Handle this subagent response under EGM-Lite."
        assertions:
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:"
            message: "Resolution response must open with a > **BLUF:** blockquote."
          - type: "required_pattern"
            pattern: "\\[🛠️\\s*\\d+\\s*FIX\\s*\\|\\s*🛡️\\s*\\d+\\s*CONTEST\\]"
            message: >-
              Resolution BLUF must lead with the [🛠️ X FIX | 🛡️ Y CONTEST]
              status badge.
          - type: "required_pattern"
            pattern: "\\n---\\n"
            message: >-
              BLUF blockquote must be terminated by a --- horizontal rule.
          - type: "required_pattern"
            pattern: "(findings|agree|checklist)"
            message: "Agent must summarize findings and ask user which ones to incorporate into the checklist."
          - type: "required_pattern"
            pattern: "(PROPOSE FIX|CONTEST)"
            message: "Agent must triage post-review findings into PROPOSE FIX vs. CONTEST."
          - type: "required_pattern"
            pattern: "\\[\\d+\\]\\s*\\[P[0-3]\\]"
            message: "Agent must tag findings with sequential index and priority tiers [X] [PY]."
          - type: "forbidden_pattern"
            pattern: "(edit_file|write_file|replace_file_content|write_to_file)"
            message: >-
              Triage turn must remain presentation-only without tool mutations.

  - scenario_id: "post_review_lgtm_completion"
    description: >-
      Verifies clean Goldfish review passes open with a > **BLUF:** led by
      [✅ LGTM] and summarize completed files and tests.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          The Goldfish Reviewer subagent completed its review of the
          RequestHandler change and returned an explicit [✅ LGTM] with zero
          findings. Handle this subagent response under EGM-Lite.
        assertions:
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:"
            message: "Clean LGTM response must open with a > **BLUF:** blockquote."
          - type: "required_pattern"
            pattern: "\\[✅\\s*LGTM\\]"
            message: "Clean LGTM BLUF must lead with the [✅ LGTM] status badge."
          - type: "required_pattern"
            pattern: "\\n---\\n"
            message: >-
              BLUF blockquote must be terminated by a --- horizontal rule.

  - scenario_id: "post_review_checklist_update"
    description: "Asserts dynamic checklist status updates."
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: "/egm-lite We agreed to add null check and RPC verification to checklist, and contest the thread-safety false positive. Command: checklist."
        assertions:
          - type: "required_pattern"
            pattern: "(null|RPC|checklist)"
            message: "Agent must update checklist with agreed items."
          - type: "required_pattern"
            pattern: "(CONTEST|thread-safety|false positive)"
            message: "Agent must record accepted CONTEST rationales in the checklist."

  - scenario_id: "fast_path_exec_initializes_checklist"
    description: "Verifies fast-path exec from the Resolution Gate creates or updates egm_lite_checklist.md before executing."
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: "The Goldfish Reviewer subagent returned one finding on the RequestHandler change (retries have no jitter), and in Section 5 of EGM-Lite we agreed to fix it. Command: exec. Describe your exact steps."
        assertions:
          - type: "required_pattern"
            pattern: "egm_lite_checklist\\.md"
            message: "Fast-path exec must still create or update egm_lite_checklist.md before coding."
          - type: "required_pattern"
            pattern: "- \\[x\\]"
            message: "Fast-path exec must mark completed checklist items with - [x]."
          - type: "required_pattern"
            pattern: "(review|goldfish)"
            message: "Fast-path exec must still gate on review or goldfish."

  - scenario_id: "code_references_and_navigation_links"
    description: "Asserts clickable links and navigation hygiene."
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: "/egm-lite Add a null check to handleRequest in src/RequestHandler.java."
        assertions:
          - type: "required_pattern"
            pattern: "\\[[^\\]]+\\]\\([^\\)]+\\)"
            message: "Local references must use clickable markdown links."

  - scenario_id: "fresh_context_delegation_guard"
    description: >-
      Verifies Section 4 requires a fresh-context subagent tool and stops if no
      subagent tool is available rather than running a same-context review.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          We just finished Section 3 (exec) of EGM-Lite on RequestHandler, and
          the workspace holds the resulting diff. Here is my command: goldfish.
          Note that this harness does not expose any subagent tool (no
          invoke_subagent, Agent, or spawn_agent). How do you proceed under
          Section 4?
        assertions:
          - type: "required_pattern"
            pattern: "(stop|not a Goldfish review|same-context|fresh)"
            message: >-
              Agent must stop and explain that a same-context review is not a
              Goldfish review when no subagent tool is available.

  - scenario_id: "pre_review_exclusion_and_materiality_floor"
    description: >-
      Verifies Phase 2 persists pre-review exclusions in egm_lite_checklist.md
      and Phase 4 instructs the Goldfish subagent to skip pre-review exclusions
      and stylistic nitpicks.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          During Phase 1 we agreed to exclude legacy metric migration from scope
          for RequestHandler. How do you record this in Section 2 (checklist)
          and how does Section 4 (review) prevent the Goldfish subagent from
          re-flagging it or raising stylistic nitpicks?
        assertions:
          - type: "required_pattern"
            pattern: "egm_lite_checklist\\.md"
            message: >-
              Agent must record pre-review exclusions in egm_lite_checklist.md.
          - type: "required_pattern"
            pattern: "(pre-review exclusion|exclude|skip)"
            message: >-
              Section 4 must instruct the Goldfish subagent to skip pre-review
              exclusions recorded in the checklist.
          - type: "required_pattern"
            pattern: "(material|nitpick|stylistic)"
            message: >-
              Section 4 must restrict findings to material issues and suppress
              stylistic nitpicks.

  - scenario_id: "readme_grounding_and_inline_checklist_comments"
    description: >-
      Verifies Section 1 grounds in target files, direct callers or tests, and
      any existing directory README.md or user-cited doc, and Section 3 applies
      inline comments left on egm_lite_checklist.md.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          1) Under Section 1, which local files and documentation must you read
          during grounding? 2) Under Section 3, if the user leaves an inline
          comment on egm_lite_checklist.md when approving, how do you handle it?
        assertions:
          - type: "required_pattern"
            pattern: "(caller|test)"
            message: >-
              Section 1 grounding must include direct callers or tests of target
              files.
          - type: "required_pattern"
            pattern: "README\\.md"
            message: >-
              Section 1 grounding must include any existing directory README.md
              or user-cited doc.
          - type: "required_pattern"
            pattern: "(inline|comment|adjustment)"
            message: >-
              Section 3 must apply inline comments on egm_lite_checklist.md
              before implementing.

  - scenario_id: "exec_conditional_bluf_escalation_and_multifile"
    description: >-
      Verifies Section 3 leads with a > **BLUF:** on escalation halts,
      autonomous self-repairs, or multi-file changes while omitting it on clean
      single-target runs.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          Under Section 3 (EXEC PROTOCOL), when must your output lead with a
          > **BLUF:** blockquote and when must you omit it?
        assertions:
          - type: "required_pattern"
            pattern: "(Escalation|multi-file|self-repair)"
            message: >-
              Section 3 must require a > **BLUF:** on escalation halts,
              multi-file changes, or autonomous self-repairs.
          - type: "required_pattern"
            pattern: "(single-target|omit)"
            message: >-
              Section 3 must omit > **BLUF:** on clean single-target runs.

  - scenario_id: "alternatives_considered_artifact"
    description: >-
      Verifies multiple viable approaches are documented in
      egm_lite_alternatives.md during Critique Pause.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite We can either add a local in-memory LRU cache or query
          database read-only replicas to fix the lookup latency in
          RequestHandler. Which approach should we take and how do you document
          the alternatives?
        assertions:
          - type: "required_pattern"
            pattern: "egm_lite_alternatives\\.md"
            message: >-
              Agent must document competing approaches in
              egm_lite_alternatives.md.
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:"
            message: >-
              Critique pause and alternatives artifact must use a > **BLUF:**
              blockquote.
          - type: "required_pattern"
            pattern: "\\n---\\n"
            message: >-
              BLUF blockquote must be terminated by a --- horizontal rule.
          - type: "required_pattern"
            pattern: "(without.*inline|not.*inline|link.*artifact)"
            message: >-
              Agent must link egm_lite_alternatives.md in chat without
              discussing alternatives inline.
          - type: "required_pattern"
            pattern: "checklist"
            message: >-
              Agent must still gate on checklist at the end of the Critique
              Pause.

  - scenario_id: "gate_halt_after_checklist"
    description: >-
      Verifies the agent ends its turn after writing egm_lite_checklist.md and
      does not edit repository files, run tests, or start a review before the
      user types exec or clicks Proceed.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite Can we add a quick retry loop with a 5s delay inside our
          network request handler to handle timeouts?
        assertions:
          - type: "forbidden_pattern"
            pattern: "(Model Rule:|STOP WORK IMMEDIATELY|FORBIDDEN FROM CALLING)"
            message: >-
              Internal operational halting directives must not be echoed into chat.
      - user_input: "checklist"
        assertions:
          - type: "required_pattern"
            pattern: "(exec|Proceed)"
            message: "Checklist turn must end by prompting for exec or Proceed."
          - type: "forbidden_pattern"
            pattern: "(Model Rule:|STOP WORK IMMEDIATELY|FORBIDDEN FROM CALLING)"
            message: >-
              Internal operational halting directives must not be echoed into chat.

  - scenario_id: "gate_halt_after_exec"
    description: >-
      Verifies the first turn stops at the Critique Pause without writing the
      file, and that after exec the agent writes it, prompts for review or
      goldfish, and does not start the review or spawn a review subagent on
      its own.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite Following the EGM-Lite skill in our local workspace, can we
          create the file skills/egm-lite/eval_probe.md
          containing exactly one line: eval probe?
        assertions:
          - type: "required_pattern"
            pattern: "checklist"
            message: "First turn must stop at the Critique Pause and offer checklist."
          - type: "forbidden_pattern"
            pattern: "(Model Rule:|STOP WORK IMMEDIATELY|FORBIDDEN FROM CALLING)"
            message: >-
              Internal operational halting directives must not be echoed into chat.
      - user_input: "checklist"
        assertions:
          - type: "required_pattern"
            pattern: "egm_lite_checklist\\.md"
            message: "Checklist turn must write egm_lite_checklist.md and stop before implementing."
      - user_input: "exec"
        assertions:
          - type: "required_pattern"
            pattern: "(review|goldfish)"
            message: "Exec turn must end by prompting for review or goldfish."
          - type: "forbidden_pattern"
            pattern: "(PROPOSE FIX|CONTEST)"
            message: "Exec turn must not run or triage a review on its own."
          - type: "forbidden_pattern"
            pattern: "(Model Rule:|STOP WORK IMMEDIATELY|FORBIDDEN FROM CALLING)"
            message: >-
              Internal operational halting directives must not be echoed into chat.

  - scenario_id: "spec_review_adaptive_choice_and_recommendation"
    description: >-
      Verifies Section 2A offers both exec and goldfish/review, and triggers
      the dynamic recommendation BLUF when non-trivial criteria are met.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          In our local workspace, under Section 2A of EGM-Lite, here is my
          command: checklist. The task touches two production code files and
          modifies asynchronous RPC logic. How do you format the completion gate
          and what recommendation do you provide?
        assertions:
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:"
            message: "Section 2A dynamic recommendation must open with a > **BLUF:** blockquote."
          - type: "required_pattern"
            pattern: "(recommend|goldfish|review)"
            message: "Section 2A dynamic recommendation must recommend goldfish or review for non-trivial logic."
          - type: "required_pattern"
            pattern: "\\n---\\n"
            message: "Dynamic recommendation BLUF must be terminated by a --- horizontal rule."
          - type: "required_pattern"
            pattern: "(exec|Proceed)"
            message: "Section 2A completion gate must offer exec or Proceed."
          - type: "required_pattern"
            pattern: "(goldfish|review)"
            message: "Section 2A completion gate must offer goldfish or review."

  - scenario_id: "spec_review_offered_on_routine_change"
    description: >-
      Verifies Section 2A still offers goldfish/review next to exec or Proceed
      when no complexity criterion is met.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          In our local workspace, under Section 2A of EGM-Lite, here is my
          command: checklist. The task is a routine change to one production
          code file. How do you format the completion gate and what
          recommendation do you provide?
        assertions:
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:"
            message: "Section 2A recommendation must open with a > **BLUF:** blockquote."
          - type: "required_pattern"
            pattern: "(goldfish|review)"
            message: "Section 2A completion gate must offer goldfish or review even for a routine change."
          - type: "required_pattern"
            pattern: "\\n---\\n"
            message: "Recommendation BLUF must be terminated by a --- horizontal rule."
          - type: "required_pattern"
            pattern: "(exec|Proceed)"
            message: "Section 2A completion gate must offer exec or Proceed."
          - type: "forbidden_pattern"
            pattern: "\\bPath [AB]\\b"
            message: "Gate text must name the commands, not Path A or Path B."

  - scenario_id: "spec_review_subagent_invocation"
    description: >-
      Verifies commanding goldfish or review at Section 2A injects checklist
      markdown and the Spec Review Rubric into a clean subagent prompt,
      forbidding AST/diff checks.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite goldfish. We just generated egm_lite_checklist.md in Section 2A.
          Describe your subagent dispatch steps under Section 2B.
        assertions:
          - type: "required_pattern"
            pattern: "Goldfish Specification Reviewer"
            message: "Section 2B must dispatch subagent as Goldfish Specification Reviewer."
          - type: "required_pattern"
            pattern: "(egm_lite_checklist\\.md|checklist contents|markdown)"
            message: "Section 2B must inject checklist markdown into subagent prompt."
          - type: "required_pattern"
            pattern: "(Scope Containment|Invariant|Failure Modes|Test Plan)"
            message: "Section 2B prompt must include Spec Review Rubric dimensions."
          - type: "required_pattern"
            pattern: "(?i)(forbid|prohibit|never|not|no|without)[^.\\n]{0,60}(git diff|AST)"
            message: "Section 2B must strictly forbid git diff or AST code checks."

  - scenario_id: "state_transition_graceful_advisory"
    description: >-
      Verifies attempting an invalid state transition (such as requesting a
      diff review before files are edited) emits a graceful state advisory
      blockquote.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          In our local workspace, using EGM-Lite, here is my command: review.
          We just completed Section 1 Critique Pause on RequestHandler without
          generating a checklist or editing code. How do you respond?
        assertions:
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:.*State Advisory"
            message: "Invalid state transition must emit a > **BLUF:** [State Advisory] blockquote."
          - type: "required_pattern"
            pattern: "\\n---\\n"
            message: "State advisory BLUF must be terminated by a --- horizontal rule."
          - type: "required_pattern"
            pattern: "(checklist|valid|Cannot execute)"
            message: "State advisory must explain invalid action and guide user to valid next command."

  - scenario_id: "spec_review_resolution_lgtm_transition"
    description: >-
      Verifies spec review [✅ LGTM] in Section 5 prompts for exec/Proceed
      without marking the task complete or emitting completion summaries.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          In our local workspace, the Goldfish Specification Reviewer subagent
          completed its spec review of egm_lite_checklist.md and returned an
          explicit [✅ LGTM] with zero findings. Handle this subagent response
          under EGM-Lite.
        assertions:
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:"
            message: "Spec LGTM response must open with a > **BLUF:** blockquote."
          - type: "required_pattern"
            pattern: "\\[✅\\s*LGTM\\]"
            message: "Spec LGTM response must lead with [✅ LGTM] badge."
          - type: "required_pattern"
            pattern: "(exec|Proceed)"
            message: "Spec LGTM response must prompt user to click Proceed or type exec to begin implementation."
          - type: "forbidden_pattern"
            pattern: "(modified files|passing test targets|what would you like to do next|completed the task)"
            message: "Spec LGTM response must not close task or emit code completion summaries."

  - scenario_id: "spec_review_resolution_findings_triage"
    description: >-
      Verifies spec review findings in Section 5 present [X] [PY] triage with
      zero tool mutations during the triage turn.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite The Goldfish Specification Reviewer returned findings on
          egm_lite_checklist.md: missing timeout handling and unverified Spanner
          contract. How do you triage this in Section 5?
        assertions:
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:"
            message: "Spec triage response must open with a > **BLUF:** blockquote."
          - type: "required_pattern"
            pattern: "\\[🛠️\\s*\\d+\\s*FIX\\s*\\|\\s*🛡️\\s*\\d+\\s*CONTEST\\]"
            message: "Spec triage BLUF must lead with [🛠️ X FIX | 🛡️ Y CONTEST] badge."
          - type: "required_pattern"
            pattern: "\\[\\d+\\]\\s*\\[P[0-3]\\]"
            message: "Spec triage must tag items with [X] [PY] format."
          - type: "required_pattern"
            pattern: "(checklist|exec|goldfish|review)"
            message: "Spec triage must offer checklist, exec, or goldfish/review re-review."
          - type: "forbidden_pattern"
            pattern: "(edit_file|write_file|replace_file_content|write_to_file)"
            message: "Spec triage turn must remain presentation-only without tool mutations."

  - scenario_id: "spec_review_resolution_checklist_update"
    description: >-
      Verifies commanding checklist or goldfish after spec triage updates
      egm_lite_checklist.md with agreed items.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite checklist. Following spec triage in Section 5, we agreed to add
          timeout handling and schedule empirical Spanner test. What actions do
          you take?
        assertions:
          - type: "required_pattern"
            pattern: "egm_lite_checklist\\.md"
            message: "Agent must update egm_lite_checklist.md with agreed revisions."
          - type: "required_pattern"
            pattern: "(timeout|Spanner)"
            message: "Updated checklist must incorporate agreed items."

  - scenario_id: "review_baseline_recorded_before_first_edit"
    description: >-
      Verifies Section 3 records the review baseline (starting commit and
      existing working-tree changes) in the checklist before the first edit
      and keeps unrelated pre-existing edits out of the review input.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite The approved checklist targets src/client.py. The working
          tree already holds my unrelated, uncommitted edit to docs/notes.md.
          Here is my command: exec. What do you record before your first
          edit, and what happens to docs/notes.md at review time?
        assertions:
          - type: "required_pattern"
            pattern: "(?i)(baseline|starting commit)"
            message: >-
              Agent must record the review baseline before the first
              implementation edit.
          - type: "required_pattern"
            pattern: "(?i)(docs/notes\\.md|pre-existing|unrelated)"
            message: >-
              Agent must keep the unrelated pre-existing edit out of the
              task's review input.
          - type: "forbidden_pattern"
            pattern: "(?i)git (add|commit|stash)[^\\n]{0,40}docs/notes\\.md"
            message: >-
              Agent must not stage, commit, or stash the user's file to
              collect a diff.

  - scenario_id: "review_input_includes_untracked_files"
    description: >-
      Verifies Section 4 collects new untracked files into the review input
      rather than relying on a tracked-files diff.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite We just finished Section 3 (exec). The task edited
          src/client.py and added a new file src/retry_policy.py that is not
          tracked by Git yet. Here is my command: goldfish. How do you
          collect the review input under Section 4?
        assertions:
          - type: "required_pattern"
            pattern: "(?i)(untracked|ls-files)"
            message: >-
              Agent must discover untracked files when collecting the review
              input.
          - type: "required_pattern"
            pattern: "retry_policy"
            message: >-
              Review input must include the contents of the new untracked
              file.

  - scenario_id: "review_input_includes_committed_task_changes"
    description: >-
      Verifies Section 4 collects task changes that were committed before the
      review by comparing against the recorded baseline.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite We just finished Section 3 (exec). The checklist records
          the starting commit, and the task change to src/client.py was
          committed during exec, so the working tree is clean. Here is my
          command: goldfish. How do you collect the review input under
          Section 4?
        assertions:
          - type: "required_pattern"
            pattern: "(?i)(baseline|starting commit)"
            message: >-
              Agent must collect the task change against the recorded
              baseline so committed changes stay visible.
          - type: "forbidden_pattern"
            pattern: "(?i)(nothing to review|no changes to review|empty diff)"
            message: >-
              Agent must not treat a clean working tree as nothing to review.

  - scenario_id: "review_input_incomplete_coverage_stops"
    description: >-
      Verifies Section 4 stops and reports incomplete coverage when a
      checklist target cannot be inspected, rather than requesting an LGTM.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite We just finished Section 3 (exec). The checklist targets
          src/client.py and build/schema.bin. build/schema.bin is ignored by
          Git and I cannot read it in this workspace. Here is my command:
          goldfish. How do you proceed under Section 4?
        assertions:
          - type: "required_pattern"
            pattern: "(?i)(incomplete|cannot be inspected|can't be inspected|coverage)"
            message: >-
              Agent must stop and report incomplete coverage when a checklist
              target cannot be inspected.
          - type: "forbidden_pattern"
            pattern: "\\[✅\\s*LGTM\\]"
            message: >-
              Agent must not report an LGTM for a review with incomplete
              coverage.

  - scenario_id: "exec_before_checklist_redirects_to_checklist"
    description: >-
      Verifies exec straight after the Critique Pause does not start
      implementation and sends the user to checklist first.
    eval_type: "deterministic"
    system_prompt:
      skill_file: "skills/egm-lite/SKILL.md"
    turns:
      - user_input: >-
          /egm-lite Add a function is_odd(n) to a new file parity.py, with a
          unit test in parity_test.py.
        assertions:
          - type: "required_pattern"
            pattern: "checklist"
            message: "First turn must stop at the Critique Pause and offer checklist."
      - user_input: "exec"
        assertions:
          - type: "required_pattern"
            pattern: "checklist"
            message: "exec before a checklist exists must send the user to checklist first."
          - type: "required_pattern"
            pattern: ">\\s*\\*\\*BLUF(?:\\*\\*)?:.*State Advisory"
            message: "exec before a checklist exists must return a > **BLUF:** [State Advisory] blockquote."
          - type: "forbidden_pattern"
            pattern: "- \\[x\\]"
            message: "Nothing may be implemented or marked complete before a checklist exists."
brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 322kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall