chat

star 2

Chat about the current status of the analysis, review progress, and request revisions to milestones. Use when the user wants to talk about what's working, what isn't, or what needs changing. Triggers on: chat, review status, what's the status, request revision, something is wrong, adjust.

stanford-ai4physics By stanford-ai4physics schedule Updated 4/2/2026

name: chat description: "Chat about the current status of the analysis, review progress, and request revisions to milestones. Use when the user wants to talk about what's working, what isn't, or what needs changing. Triggers on: chat, review status, what's the status, request revision, something is wrong, adjust." user-invocable: true

Discuss Analysis Status

Interactive discussion about the current state of the analysis project. Helps the user understand progress, diagnose issues, and request revisions that get written into project.json for the Overwatcher to apply.


The Job

  1. Read project.json and progress.txt to understand current state
  2. Present a concise status summary to the user
  3. Have a conversation about any concerns, issues, or desired changes
  4. If the user wants revisions: write them as a "revisions" array on the appropriate milestone in project.json

Step 1: Status Summary

Read the current project.json and progress.txt, then present:

  • Which milestones are done, which are pending, which are at a checkpoint
  • The current checkpoint (if any) and what it achieved
  • Any recent learnings or issues from progress.txt

Keep it concise — 5-10 lines max. Let the user drive the conversation.


Step 2: Conversation

The user may want to:

  • Understand what happened: Explain what the Overwatcher did, what tests passed or failed, what the subagents produced
  • Diagnose problems: Help figure out why something isn't working — read relevant source files, test outputs, or logs
  • Request changes: The user wants something adjusted — a different approach, a bug fixed, a naming convention changed, etc.

Listen carefully. Ask clarifying questions if the request is ambiguous. Read the relevant code files to understand the current implementation before suggesting changes.


Step 3: Writing Revisions

When the user confirms they want changes applied, write them as a "revisions" array on the appropriate milestone in project.json.

Rules for writing revisions

  • Each revision item must be a concrete, actionable instruction that the Overwatcher can execute without ambiguity
  • Only add "revisions" to milestones that have already been worked on (passes: true). For future milestones (passes: false), edit their description or other fields directly in project.json instead — there is no existing work to revise, and adding revisions would waste an extra iteration
  • If the milestone is at a checkpoint (passes: true, checkpoint_approved: false), add revisions directly — the Overwatcher will apply them on the next iteration
  • If the milestone already has checkpoint_approved: true, set it back to false and add the revisions
  • Do NOT set passes: false unless the user explicitly wants a full redo

Revision format in project.json

{
  "id": "M-003",
  "title": "ParticleNet-Lite model architecture",
  "is_checkpoint": true,
  "checkpoint_approved": false,
  "passes": true,
  "revisions": [
    "Rename knn() to k_nearest_neighbors() for clarity",
    "Add a test that checks output is invariant to particle ordering"
  ]
}

Translating user concerns into revisions

The user might say vague things like "the plots look wrong" or "this command doesn't work". Your job is to:

  1. Investigate — read the code, check the outputs, understand the issue
  2. Formulate concrete revision instructions the Overwatcher can act on
  3. Confirm with the user before writing to project.json

Good revisions:

  • "Fix the x-axis range in the mjj distribution plot to [3000, 4000] GeV"
  • "The SelectEvents task crashes with KeyError 'jet_pt' — rename the column access to match the dataset schema"
  • "Add a legend to all plots produced by PlotDistributions"

Bad revisions (too vague):

  • "Fix the plots"
  • "Make it work"
  • "Improve the code"

Step 4: Confirm and Save

Before writing revisions to project.json:

  1. Show the user the exact revisions you will write
  2. Show which milestone they will be added to
  3. Ask for confirmation
  4. Write to project.json
  5. Tell the user to re-run ./researcher.sh to apply the revisions

Important

  • Do NOT modify any source code, tests, or other files — only project.json
  • Do NOT run the Overwatcher or any law tasks — just discuss and write revisions
  • The Overwatcher will apply the revisions on its next iteration
  • If the user wants to approve a checkpoint, help them set checkpoint_approved: true in project.json instead of writing revisions
Install via CLI
npx skills add https://github.com/stanford-ai4physics/SHARP --skill chat
Repository Details
star Stars 2
call_split Forks 0
navigation Branch main
article Path SKILL.md
Occupations
More from Creator
stanford-ai4physics
stanford-ai4physics Explore all skills →