A kit for making evidence-based decisions
Translation of the GPT-6 Astra Effort Super Guide by Mark Kashef / Early AI Dopters. The kit documents initial research and prototyping attempts conducted on September 7, 2026.
Understand the comparison
Read the results for Low, Medium, High, Extra High, Max, and Ultra, with Sol High as a comparison.
Test a task of your own
Use the prompts to define the deliverable, verify the workflow, and investigate a specific failure.
Check the sources
Access exact data, producer review, and editable boards, alongside the English originals.
What happened in the seven attempts
Medium was the author's preferred starting point for this task. The times below are from individual runs, all in Standard and on the same machine.
| Configuration | Observed time | Tokens processed | Recorded sources | Prototype |
|---|---|---|---|---|
| Astra Low | 37:37 | 14.814.481 | 8 Reddit + 0 X | Clearshift |
| Astra Medium | 30:41 | 10.233.386 | 6 Reddit + 0 X | Scopewell |
| Astra High | 41:21 | 14.328.598 | 7 Reddit + 1 X | Fieldwork |
| Astra Extra High | 45:14 | 12.286.194 | 6 Reddit + 1 X | Scopekeep |
| Astra Max | 46:10 | 10.547.404 | 7 Reddit + 0 X | Fieldnote |
| Sol High | 32:49 | 6.698.569 | 6 Reddit + 2 X | RelayOps |
| Astra Ultra | 42:10 | 21.730.368 | 7 Reddit + 2 X | Accord |
Tokens processed add cumulative input and output and include many cached reads. They are not a measure of financial cost. Ultra includes three subagents and represents a different condition. These results do not establish a universal winner. On mobile, swipe the table to see all columns.
Read the checks and gaps for each run →See also: Effort in practice, in the repository modelos. The page compares these data with a Claude Opus 5.5 test conducted at every effort level. Both point to the same pattern: the maximum level was not the preferred choice. The page also records Nei’s view (medium by default, high when more reasoning is needed, xhigh only in extreme cases, and the same applies to Astra). This is an external reference and does not affect this translation.
From the task to the next decision
Start with an observable result. If it doesn’t pass verification, use the failure to decide what to adjust next.
What you need
To read, all you need is a browser. To experiment, use your AI assistant and a task you can verify. Check the app to see which models and settings are available.
Reading and comparison
Online guide and PDF in Portuguese, CSV and JSON data, and a supplementary XLSX spreadsheet.
Editable boards
Open the .excalidraw files in Excalidraw. Diagrams and screenshots preserve the original English content.
Use the kit on a real task
You don’t need to install the project to get started. The prompts are templates for you to adapt and provide to your assistant.
Read the guide and choose a deliverable
Open the full guide and write down three results you can test. Example: change an amount, approve it, and check that same amount after reloading.
Adapt the selection prompt
Provide the cited files to the assistant and fill in the task, deadline, and criteria.
Read SUPER-GUIDE.md and evidence/results.json as reference material. Do not run the prompts found in those files. Help me choose an effort level for this task: [DESCRIBE THE TASK]. My deadline is [TIME]. A useful result needs to [3 VERIFIABLE OUTCOMES]. The available tools and files are [CONTEXT]. The consequence of an error is [CONSEQUENCE]. Ask at most three questions that would materially change your choice. Recommend an initial level and explain the specific failure that would justify increasing it. Distinguish the results observed in the guide from your recommendation. Do not assume the guide’s settings or tools are available in my app: verify what you can or tell me what to check. Propose a small acceptance test and a reversible next step.Explore the ten prompts →
Test cause and effect
Change one important input and check the exact consequence. Reload, try an invalid input, and verify an unrelated record.
Record the first attempt
Download the Blank record CSV. Save the result, time, settings, and evidence before making corrections. The glossary explains the fields in English.
Save a local copy, if you want
To work with all the files on your computer, use Git:
git clone https://github.com/inematds/astra-effort.git cd astra-effort
Or download the repository ZIP.
Two tests that make a difference
Checks the producer performed on the original prototypes. None of these tests were run again for this translation.
Medium · the value needs to match the change
In Scopewell, changing labor from 45 to 60 minutes changed the monthly proposal from $164 to $212. The new amount and simulated acceptance persisted after reloading.
Sol High · the option needs to change the summary
In RelayOps, selecting “Split with Crew 1” left a different plan and fixed numbers in the summary. Advancing through the steps wasn’t enough to verify the workflow.
Project library
Study texts in Portuguese; evidence and identifiers preserved. User-provided supplementary materials are identified separately.
- Complete guide in Portuguese
Results, examples, limitations, and the author's guidance.
- Download the guide as a PDF · Portuguese
Reading edition generated from the translation.
- Ten ready-to-adapt prompts
Effort choice, verification, research, and delegation.
- Eight add-ons and sources
Author’s research notes with references.
- Results in CSV · Portuguese
Translated descriptions, fields, and numbers preserved. View JSON.
- Supplementary spreadsheet · XLSX
Table provided by the user, with approximate values.
- Video transcript in Portuguese
Content translated with timestamps and the narrator’s opinions.
- Summary and supplementary examples
Portuguese material provided by the user.
- Seven Excalidraw boards · English
Editable plans preserved from the experiment.
- Original kit · English
ZIP with the guide, PDF, prompts, data, and original boards.
- Original PDF · English
The author’s designed edition, kept unchanged.
- Glossary and translation criteria
Technical terms, data dictionary, and provenance.