refining-problem

Refining Problem

Sharpen a rough problem statement into a precise, research-ready formulation through autonomous self-refinement backed by literature research. Produce problem.md as the source of truth for all downstream steps.

CRITICAL: This skill is fully autonomous. You MUST NOT ask questions, wait for user input, or request clarification. You have all the information you need: the problem slug and the initial description. Infer everything else from the description and from literature findings. Proceed through every step without stopping.

The skill writes an initial draft from the provided description, uses literature findings to assess and improve it, and auto-determines novelty requirements based on the literature landscape.

Arguments

• Problem slug -- kebab-case directory name (required)
• Initial description -- rough problem statement (required on first run, optional when resuming)

Workflow

Step 0: Setup

Create the problem directory and references subdirectory:

bash -c "mkdir -p <problem-slug>/references"

If problem.md already exists, read it and skip to Step 3 for assessment. The file is the checkpoint.

Step 1: Launch background research agents

Immediately launch 2-3 background Task agents (run_in_background: true) to map the problem space. Focus each on a different angle:

• Scope agent: Search for canonical definitions, existing frameworks, or established formulations of this problem. Run /researching-literature with the topic and problem directory.
• Prior work agent: Search for existing solutions, known attempts, and related work. Run /researching-literature.
• Boundary agent: Search for edge cases, known failure modes, and adjacent problems. Run /researching-literature.

Launch all three in a single response.

Step 2: Write initial problem.md draft

Using the initial description, write <problem-slug>/problem.md using the structure in references/problem-template.md. Fill in each section based on what can be inferred from the initial description:

• Problem Statement: expand the initial description into 1-3 precise paragraphs
• Scope: infer in-scope and out-of-scope boundaries from the description
• Key Unknowns: identify the core unknowns implied by the description
• Success Criteria: derive testable success criteria directly from the initial description
• Novelty Required: set to TBD (determined in Step 4)
• Constraints: note any constraints implied by the description
• Background: set to Pending literature review (filled in Step 3)

Step 3: Autonomous refinement loop (max 3 iterations)

Wait for background literature agents from Step 1 to complete. Read references.md for their findings.

For each refinement iteration (up to 3):

3a. Update background section: Synthesize literature findings into the ## Background section of problem.md using Edit. Incorporate relevant definitions, prior work, and known boundary conditions discovered by the research agents.

3b. Assess problem definition quality: Evaluate problem.md against three criteria:

• Clear research question: Does the Problem Statement articulate a specific, answerable question -- not vague, compound, or circular?
• Testable success criteria: Are the Success Criteria specific enough that an experiment could confirm or refute them?
• Appropriate scope: Is the scope neither too broad (would take unbounded effort) nor too narrow (trivially answerable)?

3c. Decision:

• If all three criteria pass: proceed to Step 4.
• If any criterion fails and iterations remain: identify the specific deficiency, refine problem.md to address it, launch 1-2 additional targeted /researching-literature Tasks to fill knowledge gaps related to the deficiency, wait for results, and loop back to 3a.
• If 3 iterations pass without all criteria met: proceed to Step 4 with the best available formulation. Note any remaining deficiencies in the Background section.

Step 4: Auto-determine novelty and finalize

Assess the novelty landscape from literature findings in references.md:

• If existing published solutions substantially cover the problem space (multiple papers addressing the same question with confirmed results): set Novelty required: yes -- replications of known work will not satisfy the research goal.
• If the space is largely unexplored (few or no papers directly addressing the question, or existing work is preliminary/inconclusive): set Novelty required: no -- confirming or replicating findings is a valid outcome.

Update the Novelty required: field in problem.md from TBD to yes or no using Edit. Once set, this field is immutable for the duration of the research session.

Output the path to problem.md and confirm the autonomous loop is ready to proceed.

References

• references/problem-template.md -- Structure template for problem.md output