conon.ai

conon.ai/Tips

One Task, One Session

The agent re-reads the whole conversation with every message you send — side questions and dead ends ride along forever, degrading answers and burning limits. /clear is free; a cluttered context is not.

IntermediateWORKFLOWClaude Code

Why it works

The agent re-reads the entire conversation with every message you send — the side question about email subject lines rides along on every future turn of your landing-page task, degrading answers and burning limits. One task per session isn't tidiness; it's how you keep the context window filled with signal instead of sediment.

How to do it

  1. Finish a task? /clear before starting the next. It costs nothing.
  2. Unrelated question mid-task? Open a second session for it — don't contaminate this one.
  3. Big project? Slice it: design session, build session, polish session — staged like real work.
  4. Feel answers degrading? /context shows exactly what's occupying the window.
  5. Name each session for its one task (you're already renaming them, right?).
  6. End of day: close out with /clear rather than leaving a junk-drawer session for tomorrow.

Copy this

prompt
/clear

What to watch for

  • One eternal mega-session 'so it remembers everything' achieves the opposite — the window fills, auto-compaction guesses at what mattered, and it remembers less.
  • Project knowledge that must persist belongs in CLAUDE.md, not in conversation history — history is the wrong place to store anything.

Sources

Original: conon.aiDocumentation: code.claude.com

Read next

Make It Review Its Own Diff
'Done' is a claim, not a fact. Before accepting changes say: 'check these changes before I accept — look for real errors, not formatting'. The agent regularly finds its own bugs, and the closing phrase stops it nitpicking commas instead of logic.
Claude Code
Show A Screenshot, Skip Adjectives
Words like 'modern and minimalist' mean nothing — every ugly site's brief said that. Saw a style you like? Screenshot it into the prompt with 'in this spirit'. A picture transfers taste that three paragraphs of adjectives never will.
Claude Code
Explanatory Style For Month One
Ask for the explanatory output style and the agent narrates why it chose what it chose as it works — the black box becomes a running commentary. When the explanations start to annoy you, that's your graduation signal: switch back to default.
Claude Code
More Effort, Or A Bigger Model?
Bad result? Diagnose before you rage-quit to another tool. Skipped files, no self-check, quit halfway — that's low effort: raise the effort level. Read everything and still confidently wrong — that's the model's ceiling: escalate the model. This one question settles 95% of 'the AI got dumber' complaints.
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 →