conon.ai

conon.ai/Tips

Send Noisy Research To A Subagent

Any side-investigation that would flood your session — 'which of these fifty files mentions the old pricing?' — goes to a subagent. It reads everything inside its own separate memory and returns three lines; your session's context stays clean for the actual work.

AdvancedAGENTSClaude Code

Why it works

Some questions are cheap to ask and expensive to host: 'which of these fifty files mentions the old pricing?' floods your session with fifty files' worth of context that outlives the answer. A subagent runs the investigation in its OWN separate memory and returns three lines; your session keeps the conclusion and none of the reading.

How to do it

  1. Spot the pattern: a side-question whose ANSWER is small but whose research is large.
  2. Phrase the delegation: 'use a subagent to investigate X and report back briefly.'
  3. Specify the return shape: 'a list of affected files and one line on what each needs.'
  4. The subagent reads the pile in its own context; your window stays clean.
  5. Receive the summary; act on it in your main session.
  6. Chain them for breadth: 'one subagent per directory, then combine their findings.'

Copy this

prompt
Use a subagent to investigate which files in this project still reference the old pricing. I want back only: the file list, one line each on what needs changing. Don't paste file contents here.

What to watch for

  • Without a return-shape instruction, some subagents report like enthusiastic interns — pages of process. The 'report back only X' clause is load-bearing.
  • Don't subagent the MAIN task — delegation is for side-research; the work you're actually steering belongs where you can steer it.

Sources

Original: conon.aiDocumentation: code.claude.com

Read next

Scheduled Agents Need An Open Lid
Local scheduled tasks run only while the app is open and the machine is awake — a missed slot is skipped, not postponed. Turn on keep-awake, leave the laptop lid open, and write drift-tolerant instructions: 'if it's already past 5pm, skip today and report that you skipped'. Need lid-closed reliability? That's a cloud task — which can't see your local files.
Claude Code
Typed It Three Times? Make A Skill
The third time you type the same procedure is the signal: say 'create a skill that does this' and the agent writes the skill file itself. Then improve it in a loop — run it, note what's off, have it edit its own skill. House rules: 2nd repeated mistake → CLAUDE.md line; 3rd repeated prompt → skill.
Claude Code
Writer Session, Reviewer Session
Run two parallel sessions on one version-controlled project: one builds, the other — with fresh context — reviews. The docs are explicit about why it works: a fresh session isn't attached to code it didn't write, so it critiques honestly where the author session defends.
Claude Code
/context Names What Eats The Window
When the agent 'gets dumber toward evening', the window is full — /context shows exactly what by: tools, skills, messages, briefing. Then choose deliberately: /compact WITH instructions on what to preserve (it's an expensive operation — the agent re-reads everything), or the free /clear if the task is done.
Claude Code

We build and operate the systems behind tips like this one — agents, automations and internal tools, shipped into production. See what the studio does →