Adapted from drona23/claude-token-efficient for Perplexity Computer.
One SKILL.md file. Drop it in. Cuts the filler Perplexity Computer adds to every response — narrated tool calls, sycophantic openers, post-action summaries, redundant confirmations — without losing any useful signal.
By default, Perplexity Computer:
- Opens with "Great question!" / "Sure!" / "Happy to help!"
- Narrates every tool call: "Let me search for that...", "Now I'll read the file..."
- Summarizes what it just did after doing it: "I've successfully created the file at /path/..."
- Restates your question before answering it
- Adds "Let me know if you need anything else!" to everything
- Creates todo lists for 2-step tasks
- Delegates to subagents when one direct tool call would suffice
- Over-qualifies simple answers with unnecessary caveats
All of this costs tokens. None of it adds value.
Upload SKILL.md to your Perplexity Computer skills. It activates on every session automatically.
| # | Problem | Fix |
|---|---|---|
| 1 | Sycophantic openers/closers | No filler phrases — just answer |
| 2 | Narrated tool usage | Execute tools silently |
| 3 | Post-action summaries | If the user can see the result, don't describe it |
| 4 | Restating the question | Answer directly |
| 5 | Redundant file confirmations | File is visible in UI — don't narrate the path |
| 6 | Unnecessary todo lists | Just do the task if it's under 3 steps |
| 7 | Over-delegation to subagents | Direct tool call when one call is enough |
| 8 | Decorative Unicode | Plain hyphens and straight quotes only |
| 9 | Excessive caveats | Save qualifications for genuine edge cases |
| 10 | Pre-action planning narration | Do, don't announce |
| Profile | Best For |
|---|---|
| Universal | Any Perplexity Computer session |
| Coding | Dev work, code review, debugging in the sandbox |
| Research | Web search, analysis, report building |
| Automation | Cron jobs, pipelines, batch operations |
- Download
SKILL.mdfrom this repo - Go to Skills → Create skill → Upload a skill
- Drop it in
- Active on every session automatically
Use the original: drona23/claude-token-efficient
Copy to your agent's skills directory. Follows the agentskills.io open standard.
User instructions always win. If you ask for verbose output, step-by-step breakdowns, or detailed explanations, those instructions override every rule in this file.
- 110 lines — intentionally minimal so it doesn't cost more input tokens than it saves
- 5.2 KB — small enough to load on every message without overhead
- Validated against the agentskills.io spec
Based on drona23/claude-token-efficient by Drona Gangarapu. Adapted for the Perplexity Computer environment.
MIT
