Agent Skills: Files To Requirements

Files To Requirements

Overview

Use this skill when the input is not one clean document but a local folder full of notes, docs, configs, drafts, logs, markdown files, or mixed working materials. The goal is not RAG-style Q&A. The goal is to turn scattered files into a clear requirement statement with explicit gaps and follow-up questions.

When To Use

• The user points to a local directory instead of a single source file.
• The source material is spread across multiple files or nested folders.
• The task is to clarify what is being asked, what is constrained, and what is still missing.
• The user wants a repeatable preprocessing workflow before summarization or requirement analysis.

Do not use this skill when:

• The user already provided one clean document and only wants a short Requirements and Target extraction.
• The task is general knowledge retrieval or open-ended Q&A over a knowledge base.

Workflow

1. Collect files

• Start from the user-provided directory or file set.
• Filter by:
  • directory scope
  • file extension
  • modified time
  • optional filename keywords
• Prefer high-signal text sources first, such as md, txt, rst, json, yaml, yml, csv, source code comments, and lightweight docs.
• Exclude obvious low-value files such as binaries, lockfiles, build artifacts, vendor folders, caches, and generated bundles unless the user explicitly wants them.
• Record the collection rule you used so the result is auditable.

2. Extract text and metadata

• For each selected file, keep:
  • absolute or repo-relative path
  • title or inferred topic
  • modified time if available
  • short summary
  • key signals such as goals, constraints, decisions, TODOs, open questions, or interfaces
• Preserve evidence anchors by referencing the source file for important claims.
• If the file is too large, summarize only the sections that affect scope, constraints, or expected outcomes.

3. Identify themes and overlap

• Group files by topic, system, feature, project phase, or stakeholder concern.
• Merge duplicates and near-duplicates into one normalized statement.
• Separate:
  • repeated facts
  • conflicting statements
  • stale or superseded material
• Prefer the clearest and most recent source when two files say nearly the same thing.
• If two sources conflict and recency does not resolve it, keep both and mark the conflict explicitly.

4. Generate requirement candidates

• Convert the consolidated knowledge into candidate requirements under these lenses:
  • user goal
  • functional need
  • technical constraint
  • delivery or workflow expectation
  • missing dependency or missing decision
• Express each candidate as a short, reviewable statement.
• Mark whether each item is:
  • explicit
  • inferred
  • unresolved
• Do not present implementation ideas as requirements unless the source makes them mandatory.

5. Generate clarification questions

• Turn ambiguity into a short list of concrete follow-up questions.
• Focus questions on decisions that materially change scope, architecture, output, or sequencing.
• Prefer questions that can be answered with one sentence or a small option set.
• Group questions by topic if there are many, but keep them brief.

6. Produce the final requirement brief

• Produce a final structured summary that a teammate can use without re-reading the source files.
• Keep the brief stable and template-driven.
• Distinguish clearly between what is known, what is inferred, and what still needs confirmation.

Default Output

Use this structure by default:

## Collection Scope
- Path:
- Filters:
- Files reviewed:

## Source Inventory
- `path` | topic | modified | short summary

## Consolidated Knowledge
- Theme:
- Confirmed facts:
- Conflicts or duplicates:

## Requirement Candidates
- [Explicit] ...
- [Inferred] ...
- [Unresolved] ...

## Clarification Questions
- ...
- ...

## Final Requirement Brief
### Goal
...

### Requirements
- ...

### Constraints
- ...

### Missing Information
- ...

Working Rules

• Stay anchored to files, not guesses.
• Prefer fewer, stronger requirement statements over long paraphrases.
• Keep provenance visible for anything important or controversial.
• If the source set is noisy, say which files drove the final conclusion.
• If the source set is incomplete, say so directly instead of filling gaps with assumptions.

Relationship To Other Skills

• Use this skill before extract-requirements-target when the real problem is scattered source collection and normalization.
• Use extract-requirements-target alone when the input is already a single document or a clean pasted block of text.

Example Triggers

• "Scan this local folder and tell me the real requirements."
• "Collect these markdown and yaml files, merge overlapping notes, then clarify the ask."
• "I have a pile of local docs. Turn them into structured knowledge and then output requirement questions."
• "先整理本地目录里的文件，再帮我把需求澄清出来。"