Marius Mutu
f4880a2a18
State files updated by dashboard/heartbeat/cron jobs, plus new KB captures (samsung firmware todo, scout song reel, weekly youtube notes). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3.8 KiB
Your Claude Limit Burns In 90 Minutes Because Of One ChatGPT Habit
- URL: https://youtu.be/5ztI_dbj6ek
- Durata: 26:35
- Data: 2026-05-02
- Tags: @work @growth #token-management #claude #ai-efficiency #agents
TL;DR
Videoclipul e despre cum obiceiurile proaste de folosire a AI-ului (ChatGPT, Claude, Gemini) ard tokens inutil — și cum le poți reduce de 8-10x fără să pierzi calitate. Autorul (Nate) a construit un "Stupid Button" care auditează pattern-urile de utilizare și identifică risipa. Modelele nu sunt scumpe — obiceiurile tale sunt.
Puncte cheie
1. Formatele de fișiere ucid bugetul de tokens
- PDF raw: 100k+ tokens pentru 4.500 cuvinte de conținut (overhead header/footer/metadata)
- Markdown: 4-6k tokens pentru același conținut — economie 20x
- Convertire gratuita: orice serviciu online sau direct Claude
- Screenshots = groaznic pentru tokens. Copy-paste text direct.
2. Nu sprawla conversatiile
- La fiecare turn, modelul reciteste INTREAGA conversatie
- 20-30 turns = context window umplut, model "drift", scadere calitate
- Regula: sesiune noua la 10-15 turns
- Doua moduri separate: gathering info vs. executie. Nu le amesteca.
3. Model selection inteligenta
- Opus: rationament, decizii complexe, arhitectura
- Sonnet: executie, coding, debugging
- Haiku: formatare, polish, taskuri simple
- Nu folosi Ferrari la cumparaturi. Opus nu pentru orice.
4. Auditeaza plugin-urile si conectorii
- Fiecare plugin = tokens incarcati la FIECARE sesiune
- Un utilizator a raportat 50k tokens consumati inainte de primul cuvant scris
- Daca nu il folosesti activ → sterge-l. "Barnacle on a ship"
5. Prompt caching (pentru API / agenti)
- Cache hits Opus: $0.50/M vs $5/M standard = 90% discount
- Cache-uieste: system prompt, tool definitions, reference docs
- Daca nu faci asta, arunci banii pe fereastra
6. Web search eficient
- Perplexity MCP vs Claude native search: 10-15k tokens mai putin per search
- De 5x mai rapid + citations structurate
- MCP = magic pentru token management la search
7. Cei 5 Comandamenti pentru Agenti
- Index references — nu dump documente intregi in context
- Pre-proceseaza — documentul trebuie sa ajunga ready-to-use, nu ready-to-read
- Cache stable context — system prompts, tool defs, persona, reference docs
- Scope minim — un planning agent nu are nevoie de intreaga codebase
- Masoara — daca nu stii cost per call, optimizezi orb
Estimare economii reale
| Workflow | Input tokens | Output tokens | Cost estimat |
|---|---|---|---|
| Sloppy (PDF raw, 30 turns, Opus tot) | 800k-1M | 150-200k | $8-10 |
| Clean (markdown, 10-15 turns/session, model mix) | 100-150k | 50-80k | ~$1 |
Economie: 8-10x pentru acelasi rezultat.
Quote-uri
"Frontier AI can be absurdly cheap when you know what you're doing. The models are not expensive, it's your habits that cost a lot."
"Use Opus for reasoning and Sonnet for execution and Haiku for polish."
"Cache hits on Opus cost 50 cents per million versus $5 per million standard."
"Models perform worse when they're drowning in irrelevant context."
"You cannot improve what you do not measure."
"Your mistakes scale with the price of intelligence."
"As models get more intelligent, we can lean out the context window initially because we can trust the model to retrieve better."
Actiuni practice (pentru Marius)
- Audit plugins/skills active in Claude Code —
/contextcommand verifica ce se incarca - Sesiuni noi mai des in Claude Code (10-15 turns maxim)
- Pentru Ralph agents: verifica daca system prompt-ul e caches
- Web search via Perplexity MCP in loc de Claude native search
- Convertire PDF la markdown inainte de a baga in context