deep-understanding

name	deep-understanding
description	Require a thorough deep-read of the relevant codebase before planning or implementing; write findings to a persistent markdown file (e.g. research.md) so the user can review and correct before any work proceeds.

Deep understanding

Every meaningful task that touches non-trivial code should start with a deep-read directive: thoroughly understand the relevant part of the codebase before doing anything else. Findings must be written into a persistent markdown file, never only a verbal summary in chat. That file is the user’s review surface — they can verify understanding and correct misunderstandings before any planning or implementation.

Apply intelligently. Use this skill for non-trivial work (multiple files, new area, complex behavior); do not use it for simple, self-contained requests (e.g. creating a regex, one-line fix, single known file) to avoid wasting tokens and overloading context.

Trivial tasks: for very small, single-file edits (e.g. typo, one-line fix), a short in-chat confirmation of understanding in the form of a question is acceptable instead of a full research report. Confirm in chat before editing code.

Mindset

You are a research scientist. Form no hypotheses before gathering evidence. Write everything down. Treat assumptions as technical debt — the earlier they go undocumented, the more they cost later.

When to use this skill

• The task involves a folder, flow, or system the agent may not already understand in depth
• The user asks to "understand deeply", "research first", "study before implementing", or to produce a research/report before planning
• The work is non-trivial (multiple files, non-obvious behavior, possible bugs) and surface-level reading would be risky
• The user wants a written artifact they can review and correct before implementation

When NOT to use

• Simple, self-contained requests (e.g. creating a regex pattern, one-line fix, single known file, quick question)
• Trivial, single-file changes where a quick read is enough
• The user explicitly says to skip research or already has a detailed spec
• The relevant code is already well understood and documented in context

How to invoke

Use explicit depth language so surface-level reading is not acceptable. Examples:

• "Read this folder in depth, understand how it works deeply and all its specificities. When done, write a detailed report of your learnings in research.md."
• "Study the [system/flow/folder] in great detail; understand its intricacies and write a detailed research.md with everything there is to know about how it works."
• "Go through the [flow] deeply and look for potential bugs. Keep researching until you find them. When done, write a detailed report of your findings in research.md."

Phrases that signal depth: "deeply", "in great detail", "intricacies", "go through everything", "don’t stop until". Without them, the agent may skim (signature-level reading and move on).

Workflow

1. Scope the research

• Identify the folder, flow, or system to understand (from the user’s request or by asking once if unclear).
• Agree on the output file path (default: research.md in the relevant directory or repo root; user can specify another path).

2. Deep read before acting

• Read the relevant code in depth: control flow, data flow, edge cases, dependencies, and how pieces interact.
• Do not rely on signatures or file names alone; follow calls, state, and side effects as needed.
• Treat "understand deeply" as: no planning or implementation until the read is done and written up.

3. Write a persistent report

• Write a detailed report to the agreed markdown file (e.g. research.md).
• Include: what the system does, how it works, main components, important details, quirks, and (if requested) bugs or risks.
• Do not substitute a chat summary for the file. The file is the primary artifact.

4. Pause for review

• Present the report to the user (e.g. "I’ve written the report to research.md. Please review and correct any misunderstandings before we plan or implement.").
• Do not proceed to planning or implementation until the user has had a chance to review (or explicitly approves).

5. Proceed only after validation

• If the user corrects the report, update the file and re-pause if needed.
• Only after the research is validated (or approved) should you move on to planning and implementation.

Why the written artifact matters

• Review surface: The user can read the file, verify that the agent actually understood the system, and fix errors before any plan is made.
• Understanding: If the research is wrong, the plan will be wrong and the implementation will be wrong. Correct understanding first.
• Traceability: The report stays in the repo; later sessions or humans can see what was assumed and what was learned.

Output file

• Default: research.md (in the folder under study, or repo root if scope is broad).
• User can specify: e.g. docs/notifications-research.md, CONTRIBUTOR-DOCS/.../findings.md.
• Format: Markdown with clear headings (overview, how it works, components, specifics, bugs/risks if applicable, open questions).