BitShovel.One real problem at a time
Open the current workflowCurrent Workflow中文

One real problem at a time

中文

HISTORICAL DIG · JUL—AUG 2026

Should this task start in an editor, terminal, or office agent?

These products can all look like “agents,” but they work in different places: editor-based coding, terminal repository work, multi-task workspaces, and office-file actions need different context and permissions.

First recorded · Last public revision · Revision 3
CodexClaude CodeCursorTRAETencent WorkBuddy
Pinned open-source Promptfoo interfaceDATED SOURCE SNAPSHOT
Promptfoo evaluation setup with provider selection before prompts, test cases, and run options

This empty setup screen shows one comparison order: define candidates and test conditions before reading results. It does not establish local data handling, controlled permissions, or a completed run.

About this image and its source

Promptfoo repository evaluation setup, © Promptfoo 2025, MIT; visible icons are separately covered by the Lucide ISC/MIT terms. BitShovel only resized it proportionally and converted it to WebP. This empty setup screen shows that candidates, prompts, and test cases are defined before comparison; it does not establish local data handling, controlled permissions, or a completed run. Promptfoo does not participate in or endorse BitShovel.

Open source page ↗

WORKFLOW 001 · NODE POSITION

Current node 017 nodes total

Set the deliverable before the tool

  1. 00Theme contract · organized
  2. 01Entry · historical record
  3. 02Control · historical record
  4. 03Run boundary · historical record
  5. 04Delivery check · historical record
  6. 05Recovery · historical record
  7. 06Editorial focus · product experiment

HOW IT RELATES TO WORKFLOW 001

Relation to the current subject

This was an earlier cut into the problem.

This record narrowed “which agent is better?” to “what result do I need, and where will the work happen?” Workflow 001 continues from there: once tools, sessions, and plugins multiply, how can a project still be finished and resumed days later? Each record keeps its own date and evidence state instead of being merged into a falsely continuous conclusion.

Record state at the time

  • Coding & agents
  • AI agent collaboration
  • In-depth guide
  • Developing
  • Sources cross-checked · no same-task run

THE GUIDANCE AT THE TIME

Judgment at the time

Choose the result and working environment before the product name.

Open full rationale and return rules
Define the result and working environment first. For changes you want to see inside an editor, compare Cursor or TRAE; for repository work from a terminal, compare Codex or Claude Code. For office files and browser tasks, check Tencent WorkBuddy separately for file, model, and external-service permissions.
Who it was for
People with access to AI who feel buried by tools, plans, and jargon and want one concrete result first
Why it mattered then
Open note

Starting from a product name often means taking on learning, permission, and subscription costs before knowing whether it will help deliver the result.

Where the judgment stops
Open note

Official pages show what each product currently offers and claims, not how well, reliably, or cheaply it handles the same task. The added open-source Promptfoo screens show only a method for defining tests before comparing results; they are not real benchmark results for these agent tools. Features, plans, and data handling change, so check again before a trial.

CHECK BEFORE READING ON

Fit check

Check whether it fits the task you are trying to finish.

This record may help

  • You have access to AI but feel buried by tools, plans, and jargon, and want one concrete result first.
  • You can define the result and working environment before comparing two categories that could actually finish the task.

Do not use it yet when

  • You still cannot say what result should exist at the end.
  • The first trial already requires consequential permissions, long-term payment, or a large data migration.
  • Your decision requires a BitShovel same-task test or an independent-user completion rate—this record does not have either.

THE 20-MINUTE TRIAL IT LEFT BEHIND

Historical trial

Compare only two categories and finish with a result you can open.

Time needed
About 20 minutes
Likely cost
Use a free tier or local sample first; do not prepay annually
Permission boundary
No real accounts and no write access to production
  1. 01

    First step

    Open this step

    Write down the required result, inputs, and deadline, then compare only two tool categories that could finish it.

  2. 02

    What should exist

    Open this step

    One category choice and one minimal test capped at 20 minutes.

  3. 03

    How to tell it worked

    Open this step

    The test ends with a result you can open, not merely a chat transcript or demo.

  4. 04

    Stop when

    Open this step

    You cannot define the result, or the trial demands consequential permissions, long-term payment, or a large data migration first.

  5. 05

    How to back out

    Open this step

    Export the result, revoke access, cancel the trial, and record why it did not finish.

COMPARISON METHOD, NOT A BENCHMARK

Method and interfaces

Define test conditions first, then place results from the same inputs side by side.

Open method boundary

The two Promptfoo documentation screens show only the order of comparison. They are not same-task tests of Codex, Claude Code, Cursor, TRAE, or WorkBuddy.

Pinned Promptfoo result interfaceDATED SOURCE SNAPSHOT
Promptfoo result table comparing three prompt variants on the same yes/no inputs with pass rates, assertions, failure reasons, and latency

All three columns are prompt variants of Model A—not different agent tools or a BitShovel or market benchmark. It only demonstrates how results from the same inputs can be compared side by side.

About this image and its source

Promptfoo repository Web Viewer documentation image, © Promptfoo 2025, MIT; visible icons are separately covered by Lucide ISC/MIT, MUI MIT, and Google Material Icons Apache-2.0. BitShovel only cropped the top metadata area containing the project author email and evaluation ID, then converted the image to WebP. The three columns are prompt variants of Model A, not different agent tools and not a BitShovel or market benchmark. Promptfoo, MUI, Lucide, and Google do not participate in or endorse BitShovel.

Open source page ↗

SOURCES AND EVIDENCE

Sources and evidence

As of August 8, 2026, these sources established how the products described their entry points and capabilities—not how they performed on the same task.

Open evidence boundary

Official pages for five tools were cross-checked with community discussions and release information. The same task was not run across these tools, and there are no independent-user results.

  1. 01
    Community report · Checked then

    User signal: many subscriptions, unclear daily return

    Open source boundary

    The discussion brings selection back to whether a tool solves a core problem and creates visible impact.

    Open original source ↗
  2. 02
    Community report · Checked then

    User signal: too many tools, no clear foundation

    Open source boundary

    Several replies favor starting with one small project and one workflow that can actually be completed.

    Open original source ↗
  3. 03
    Method source · Checked then

    How Product Hunt launches and ranking work

    Open source boundary

    Used to treat launch attention as discovery, not a user outcome.

    Open original source ↗
  4. 04
    First-party source · Checked then

    Official OpenAI Codex app introduction

    Open source boundary

    Checked maker descriptions of multiple agents, workspaces, review, and automation.

    Open original source ↗
  5. 05
    First-party source · Checked then

    Official Anthropic Claude Code overview

    Open source boundary

    Checked maker descriptions of terminal use, repository understanding, edits, and command execution.

    Open original source ↗
  6. 06
    First-party source · Checked then

    Official Cursor Agent modes documentation

    Open source boundary

    Checked permission differences among Agent, Ask, and manual modes.

    Open original source ↗
  7. 07
    First-party source · Checked then

    Official TRAE product entry

    Open source boundary

    Checked the current editor and agent positioning; no same-task run was performed.

    Open original source ↗
  8. 08
    First-party source · Checked then

    Tencent WorkBuddy terms and risk boundaries

    Open source boundary

    Disambiguates Tencent's desktop office agent and checks its third-party model, tool-call, and human-safeguard terms.

    Open original source ↗

REVISION RECORD

Revision record

Every change keeps its date and reason.

Open 2 revisions
  1. Revision 3

    Added Codex, Claude Code, Cursor, TRAE, and Tencent WorkBuddy, then reorganized the choices around editors, terminals, and office tasks.

  2. Revision 2

    Changed a broad tool comparison into choosing the simplest workable option for one concrete result, cost limit, and permission level.

WHAT THIS RECORD LEAVES

What this record leaves

Keep the old judgment—and the distance between it and today.

The current Workflow continues with delivery, context recovery, complexity, and permissions. This historical record keeps its date and evidence state so the narrowing of the problem remains visible.

Back to the Dig index