light-research

Objective: reduce the token cost of a vague task over a large scope. The splitting is done by a deterministic script (zero model tokens). Haiku subagents read the batches. Opus writes the synthesis, on the findings only, never on the sources.

Procedure

1. Judge the prompt. Precise (explicit steps and deliverable): do not use this skill. Vague (open scope): continue.
2. Ask before acting. Write: "Do you want light research by Haiku subagents to reduce tokens?"
   • No: handle the request as best you can, without batches.
   • Yes: continue.
3. Measure the scope. Run python3 ~/.claude/skills/light-research/scripts/plan_batches.py <folder> --budget 25000. The JSON gives the number of files, the estimated tokens and the batches. If the scope is under 30 files and 50,000 tokens, do it yourself: do not launch subagents.
4. Specify the question. One sentence: what you are looking for in each batch (bugs, debt, inconsistencies, security...). Without a precise question, the findings will be vague.
5. Launch the batches. One researcher subagent (Haiku 5.5, medium) per batch, all in a single message (at most 10 at a time). Each brief contains: objective, exact list of the batch's files, question, output file (<scratchpad>/lot-<id>.json), output schema, stop condition. See the template below.
6. Synthesize. One analyst subagent (Opus 5.5, medium) reads only the findings files, deduplicates them, ranks them by severity and returns the report. It rereads the sources only to confirm a doubtful finding, backed by file:line.
7. Deliver the result: report, number of batches, estimated tokens, failed batches.

Per-batch brief template

Objective: read the files of batch <id> and answer: <question>.
Context: <one sentence about the project and what is already known>.
Inputs: files (exact paths): <list>. Read nothing else.
Scope and limits: no modification; do not reread files outside the batch.
Acceptance criteria: each finding cites file:line and evidence (excerpt of 1 to 3 lines).
Format: write a JSON to <output path>: [{"file": "...", "line": n, "severity": "blocking|major|minor", "finding": "...", "evidence": "..."}] ; empty list if nothing.
Stop: if a file is unreadable or binary, note it in the JSON and move on to the next one.

Safeguards

• Never launch the batches in series: parallelism (one message, several calls) is what saves time.
• A failed batch does not stop the others: flag it in the final report.
• The script ignores dependency folders and binaries, but not secrets: never include .env or *credentials* in a scope, check the list before launching.