Menu
Cross-repo research — formulate targeted research queries based on intent, run them against chomped repos via RLM, and write structured research output.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Use this skill whenever writing prose, essays, academic content, reports, or any long-form written output, even if the user doesn't explicitly mention writing quality. Trigger for any drafting, revising, or editing of written sections — including when the user pastes text and asks for feedback.
Use this skill when working with labeled thesis arguments — creating, referencing, or weaving definitions (D), views (V), arguments (A), alternative views (AV), counter-arguments (CA), rebuttals, or barriers (B) into the thesis. Trigger when the user mentions argument IDs, cross-references, or wants to build/extend the argument scaffold.
| name | bite |
| description | Cross-repo research — formulate targeted research queries based on intent, run them against chomped repos via RLM, and write structured research output. |
| user-invocable | true |
| allowed-tools | ["Read","Write","Edit","Grep","Glob","Bash","Agent","AskUserQuestion"] |
Formulates research questions driven by intent, runs them against chomped repos, and writes structured output.
/bite <chomp1,chomp2,...|all> <intent>
<chomp1,chomp2,...> or all (required): Comma-separated chomp names (matching filenames in chomp/), or all to use every .md file in chomp/.<intent> (required): What the user wants to accomplish. A phrase or sentence.If chomp names or intent are missing, ask for them before proceeding.
Read $ARGUMENTS. Extract the first token as the chomp selector and the rest as the intent string.
If the chomp selector is all, glob for all chomp/*.md files and use their basenames (without .md) as the chomp names. If no .md files exist in chomp/, stop and tell the user to run /chomp first.
Otherwise, split the first token on commas to get the chomp names.
For each chomp name:
chomp/<name>.md exists.If any chomp is missing, stop and tell the user. They may need to run /chomp first.
Based on the user's intent alone, draft 3-8 targeted research questions. These should be specific enough to answer from source code. Think about:
Then use AskUserQuestion to present the draft questions to the user and ask:
Refine the questions based on user feedback. This back-and-forth ensures the research targets what the user actually needs rather than guessing.
For each chomp name, init the RLM and write chunks to a chomp-specific directory:
python3 ~/.claude/skills/chomp-init/scripts/rlm_repl.py init chomp/<name>.md
python3 ~/.claude/skills/chomp-init/scripts/rlm_repl.py exec <<'PY'
paths = write_chunks('chomp/.rlm_state/chunks/<name>', size=200000, overlap=0)
print(len(paths))
PY
Repeat for each chomp. Collect all chunk file paths.
For each research question, invoke the rlm-subcall subagent against every chunk file across all chomps. Pass the research question as the query.
Collect all subcall results. Keep them structured (JSON).
If the first round reveals gaps or unanswered questions, formulate follow-up queries and run another round. One pass is often not enough for cross-repo integration questions.
Slugify the intent (e.g. "integrate streamcore into my working directory" -> integrate-streamcore).
From the collected subcall results, synthesize two files:
research.md — A narrative research document (1-3 pages). Should cover:
surface-map.md — A mapping table showing how APIs/modules in one codebase relate to the other. Format:
source_api -> target_location (action: replaces/augments/no equivalent)
Write both files to:
chomp/bites/<intent-slug>/
research.md
surface-map.md
Print a summary of what was found. Highlight the most important findings — especially mismatches or surprises. The user should be able to read the summary and decide whether to act on the bite, refine it, or shelve it.
chomp/.rlm_state/.