Deep Study

A guided, user-paced walk through a research REPORT.md — either codebase research at .plans/research/<topic>/REPORT.md (citations are file:line locations in the repo) or topic research at .plans/research-with-instructions/<topic>/REPORT.md (citations are source URLs, page titles, or Context7 doc IDs). For each finding, opens the cited location, paraphrases what it says, and checks in before moving on. Read-only and stateless — every session is independent and writes nothing.

Core rules

1. Read-only. Never modify any file.
2. Stateless. Don't write STUDY.md, NOTES.md, or any state artefact. The conversation is the only progress tracker.
3. Cite while you walk. For every finding, open and paraphrase from the actual cited source — the file:line in the codebase, or the URL / doc for a topic-research report — never from REPORT.md alone. Reports drift; the cited source is the source of truth.
4. One section at a time. Walk top-level finding sections in order. Don't try to cover the whole report in one breath.
5. User-paced. After each section, check in via AskUserQuestion (continue / dig deeper / skip / ask). Don't barrel through.
6. Validate citations as you go. If a cited file is missing, or the lines no longer match what the report describes, surface it explicitly — don't silently paper over.
7. Don't produce a report inline. If no report exists, tell the user this skill studies an existing research report and one must exist first — don't generate one.
8. Walk in terse mode. Switch to terse output before the first section so paraphrases stay tight (one to two sentences per citation, no decorative formatting, no padding) — see step 0. High signal, low noise — the user is here to study code, not read prose.

Workflow

0. Enter terse mode