STRIVEOpen in STRIVE →
Oscar Morkeeth
Nov 2, 2026 · Claude Code
Recorded activity · client-reported · Gemma pointer pilot

Gemma pointer pilot: does one line make a small model read past decisions?

Project / context

Does one line in the prompt make a small model read a past design decision? Tested on Gemma 4 E2B on one laptop. Agents start each session blind to what the last one decided. With the line: read the decision file 18 of 18 times. Without: 0 of 18. But the task was too hard for th

Images from this run

Run photo
← PreviousClose ×Next →
Project image
Files changed
7
Elapsed
1h 46m
Typed turns
4

Activity along the route

Session trace
Claude CodeClaude Opus 5.5
♥ XUDOS 0CommentsOpen in STRIVE →
Recorded details and sources
Recorded tokens

Text the model read and wrote.

33.5M
Tool calls

Actions the agent took.

122
Elapsed

Time between the first and last captured event, including waiting.

1h 46m
About these measurements

Recorded tokens. 33,406,835 input + 130,558 output. Cached input is included; only calls with recorded usage are counted.

Model. Model names from the capture.

You typed. Recorded human messages, not words or keystrokes.

Tool calls. Actions the agent took.

Elapsed. Time between the first and last captured event, including waiting.

Est. dollars. No stored per-model usage and price provenance. No dollar estimate is shown.

Tool-active time · proxy. No stored tool-call bin provenance. No active-time proxy is shown.

Models and usage

Platform / harness
Claude Code
Recorded models
Claude Opus 5.5
Recorded tokens
33,537,393
Where these numbers come from

Model names and usage come from the imported capture. STRIVE has not verified them with the provider. Source: claude-message-usage. Recorded as claude-opus-5-5.

33,406,835 input + 130,558 output tokens. Cached input and reasoning tokens are already included, and are not added again. Only calls with recorded usage are included; missing usage remains unknown. This is not a bill or a measure of wasted work.

Recorded activity

Human messages: 4

Messages typed or queued by a person; excludes tool results and injected context.

Tool requests: 122

Requests the agent made to tools. A request does not prove a successful action.

Files present after writes: 7

Files the capture says were written and existed when checked. This does not prove they work.

Source and limits

Imported · client-reported. The capture supplied these numbers. STRIVE has not independently verified the source or outcome.

Recorded session span; may include idle time. Not human effort. Session date supplied by the capture.

The measurement reference identifies this submitted capture. A hash is not proof the counters are true.

Photos and descriptions are chosen by the author. They do not attest to when a session happened or what succeeded.

More from this builder

  • Cliffhanger: what your locked tokens can do todayCliffhanger
  • Micro tools day: SkillWire company grouping and weekly viewSkillWire
  • Hermes Studio: video editing becomes a conversationHermes Studio

Make a card like this

Turn a Claude Code, Cursor or Codex session on your own computer into a private run. The capture stays on your machine until you deliberately save it.

Add your run
AboutPrivacyTermsContactDelete your data