deep-interview-planning-workflows
Use for Kanban-authoritative Deep Interview fact gathering and Native Swarm handoff.
- License 许可
- MIT
- 最近更新
- 2026-10-20
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- deep-interview-planning-workflows
- author
- Hermes Agent
- license
- MIT
- version
- 2.0.0
- metadata
- { "hermes": { "tags": [ "deep-interview", "planning", "kanban", "specification", "ultrawork" ], "related_skills": [ "plan", "durable-agent-orchestration", "deep-interview-for-memory-injection" ] } }
- platforms
- [ "linux", "macos", "windows" ]
- description
- Use for Kanban-authoritative Deep Interview fact gathering and Native Swarm handoff.
03 SKILL.md 正文
Deep Interview Planning Workflows
Contract
Deep Interview gathers explicit facts before implementation. kanban.db is the
only authority for the objective, feedback, hypotheses, decisions, future
goals, round snapshots, phase, and versions. Do not treat chat text,
kb_temp, Dynamic Context, or a specification file as a replacement authority.
The Host projects only goal and runtime state to
kb_temp(information_type="deep_interview") in three atomic sections:
current_focus, current_progress, and future_needs. AI may read this
projection through Dynamic Context but may not write it. The Host Projector is
the only writer.
Interview Turns
- Read the current Deep Interview block in Dynamic Context.
- Use
interview_searchwhen the complete Kanban fact snapshot is needed. - Ask one concrete, falsifiable hypothesis about the current objective.
- Do not create a plan, specification, implementation graph, or approval verdict during an interview round.
- Let the Host append the authoritative future-goals footer. Do not generate that footer in the model response.
Explicit feedback is classified by the AI Feedback Router as exactly one of:
confirm_hypothesis: record the confirmation and continue.reject_hypothesis: record the correction and ask a revised hypothesis.goal_conflict: keep the current goal and record the new goal as future work.stop_interview: stop immediately and generate the final specification.
There is no independent Judge, review round, approval step, next-round approval, or stop confirmation.
Specification And Handoff
Only stop_interview may create a specification. The canonical location is:
HERMES_HOME/lancedb-memory/resource/spec/<title>.md
The specification is generated from the final Kanban fact snapshot and uses a
monotonic spec_version. Deep Interview does not compute a content hash or
digest.
Run hermes ultrawork plan <spec-path> [repository] or
/ultrawork plan <spec-path> [repository] after the specification is
completed. Plan accepts either the completed specification path or a direct
goal, but a path-like missing value must fail instead of being reinterpreted as
a goal.
Read-Only Tools
interview_state: compact operational phase and version metadata.interview_search: complete session, round, all-session, or keyword-matched Kanban fact snapshots.kb_temp_get(information_type="deep_interview"): projected goal/runtime state only.
interview_search is independent from interview_state and never mutates the
Interview. No legacy alias is registered.
Verification Checklist
- Kanban facts commit before projection or visible continuation.
- The projection has exactly three sections and matches
goal_state_version. - Short feedback still receives strict Dynamic Context injection.
- Projection failure pauses the Interview and blocks the model turn.
- Future goals never replace the current goal.
- Every assistant response ends with the Host-generated future-goals footer.
- Importing Markdown creates a Kanban fact seed, not a specification.
- Only AI
stop_interviewcreates the final specification. - Native handoff uses the completed specification path/revision without a Deep Interview digest, and Execute consumes the resulting frozen plan path rather than the interview ID.
04 同仓库更多 Skills
web-retrieval-resilience
Web tools fail or search rules break: repair or fall back.
blocked-page-recovery
Use when a fetch fails: 403/429, paywall, WAF, bot wall.
verification-loop
Run a six-phase verification of a Claude Code session's work — build, type check, lint, tests with coverage, security grep, and diff review — then produce a PASS/FAIL verification report. Use when verifying work after completing a feature or refactor, before creating a PR, or when quality gates must pass.
unreal-world-builder
开放世界与环境专家——精通 UE5 World Partition、Landscape、程序化植被、HLOD 和大规模关卡流式加载,打造无缝开放世界体验
unreal-technical-artist
Unreal Engine 视觉管线专家——精通材质编辑器、Niagara 特效、程序化内容生成和 UE5 项目的美术到引擎管线
unreal-systems-engineer
性能与混合架构专家——精通 C++/Blueprint 边界、Nanite 几何体、Lumen GI 和 Gameplay Ability System,面向 AAA 级 Unreal Engine 项目