I Spent a Weekend Trying to Break My ChatGPT Context System — It Scored 78-94/100

sam03069 · reddit · 2026-09-14

The author spent a weekend stress-testing a custom ChatGPT operating system, scoring 78/100 to 94/100 across models and reasoning tiers.

How it started: a custom writing skill for natural AI-assisted prose exposed a bigger problem — skills, files, past conversations and project instructions existed, but fresh chats couldn't reliably know what mattered or where to look. The writing problem became a context problem.

The routing system: recover relevant context → identify the right project → retrieve prior decisions → find the source that owns the fact → use the right tool/skill → verify. He then tried to break it: personal facts, prior conversations, project state, automations, conflicting info, same-named files, wrong-person risks, missing evidence, live-vs-historical state.

Scores: GPT-5.5 Medium 78, GPT-5.5 High 82, GPT-5.6 Medium 92, GPT-5.6 High 94 (READY); an independent GPT-5.6 Sol High audit scored 91 (24 passes, 3 partials, 0 failures).

Key failures: cross-chat retrieval is still inconsistent; a CURRENT-STATE file helps resume work but can't replace checking the live Sheet, inbox, automation or database. Weak routing causes classic mistakes: registration ≠ attendance, matching name ≠ same person, failed search ≠ never existed, historical context shouldn't override live systems.

The lesson: split "context" into chat history (what happened), learning files (lessons), CURRENT-STATE (resuming work), AGENTS (system rules), skills (task handling), and live systems (current facts). Not a bigger prompt — better routing, clear source ownership, identity binding, verification, and rules for missing evidence.

Original post →

I Spent a Weekend Trying to Break My ChatGPT Context System — It Scored 78-94/100

More from coding & agent