| name | webgpt-todo-response |
| description | Use after receiving WebGPT or other LLM feedback on an agent-infra roadmap/todo and before sending the todo back for another review. Produce a concise handoff response that says what we adopted, what we reject or question, and what WebGPT should specifically re-evaluate in the todo. |
WebGPT Todo Response
Purpose
Generate the message to send back to WebGPT after we have read its previous
analysis and drafted or updated docs/todolist.md.
Use this skill when the user asks:
- whether we have rebuttals or questions for WebGPT
- what to include when sending our todo back to WebGPT for another pass
- to prepare a response asking WebGPT to evaluate, improve, or challenge our todo
- to compare WebGPT's recommendations against our chosen implementation scope
Inputs To Inspect
Read only the files needed for the current handoff:
- WebGPT response, usually
repomix-output/webgpt.md or
repomix-output/llm/webgpt.md
- current todo, usually
docs/todolist.md
- relevant source-of-truth docs under
docs/source-of-truth/
- optional other LLM replies if the user asks for a multi-model synthesis
Do not re-run broad repository analysis unless the todo or WebGPT response
depends on code facts that are unclear.
Workflow
-
Identify WebGPT's strongest recommendations.
- Mark which ones are adopted in the todo.
- Mark which ones are intentionally deferred.
- Mark which ones are rejected or need clarification.
-
Check the todo against project boundaries.
- Shared semantics should live in
packages/* or source-of-truth docs.
- Playground UI should validate platform capabilities, not define them.
/chat should not become a quality-management surface.
- Trace/timeline data must not be treated as deterministic replay logs.
- Do not introduce shared user/org/tenant/billing models unless the task
explicitly requires them.
-
Find weak spots in the todo.
- Missing source-of-truth step
- Missing data/type/interface step before UI
- Metadata overwrite risk
- Auth/visibility ambiguity
- Scope creep into eval runner, experiment, exporter, prompt hub, or LLM judge
- Missing tests or review gates
- UI route boundary drift
-
Write a concise message for WebGPT.
- Assume WebGPT has no hidden context beyond the attached todo and bundle.
- Be explicit about decisions already made.
- Ask targeted questions instead of open-ended “any thoughts?”
- Request concrete todo edits, not generic feedback.
Output Shape
Produce a copy-ready Markdown response with these sections:
# Response To WebGPT
## What We Adopted
- ...
## Where We Differ / Pushback
- ...
## My Current Leaning
1. ...
## Highest-Value Review Points
1. ...
## Specific Questions For You
1. ...
## Please Review The Todo For
- ...
## Constraints To Preserve
- ...
Keep it short enough to paste into WebGPT with the todo. Prefer 5-10 specific
questions/checks over a long essay.
Use My Current Leaning to distinguish default decisions from genuinely open
questions. WebGPT may challenge these, but should not treat them as blank slate.
Use Highest-Value Review Points to focus WebGPT on the few risks most likely
to improve the todo. These should be sharper than the broader checklist.
Good Question Patterns
- “Does this todo accidentally move product semantics into playground code?”
- “Is
metadataJson.review sufficient for v1, or is there a concrete reason to
add top-level columns now?”
- “Are we missing any app-layer merge guards that prevent UI from overwriting
capture, feedback, or host metadata?”
- “Do the loops defer eval runner cleanly, or is eval scope leaking into review?”
- “Are the proposed tests enough to prove auth, dataset visibility, and source
lineage boundaries?”
Avoid
- Do not ask WebGPT to implement patches unless the user explicitly wants that.
- Do not ask WebGPT to run commands.
- Do not include local absolute paths.
- Do not send vague requests like “please improve this.”
- Do not restate the whole todo; reference it and ask for specific audit points.