HJ / WORKSPACE
← Open the whiteboard
YOUR EVERYDAY GUIDE

One way of working.
For everything you do.

A school assignment, a new business, research or client work. Tell me the result you want. I find the right files, apply the relevant lessons and check the work.

No special command or skill name needed.
?Start a project for my new business idea. Here is what I know. Help me work out who would buy it and what we should test first.?
01

Read our baseline

AGENTS.md is short and general. It holds working rules and links, with no offers, named clients or old task instructions.

02

Understand the job

Identify the result, format and constraints. Pick the right method. Ask only about missing details that affect the work.

03

Read this project's files

Use its goals, requirements, sources and saved feedback. School work does not inherit client marketing rules.

04

Choose the right lessons

Read the relevant saved guides. Open original sources where needed. Briefly tell you what I will use, make and check.

05

Apply them to the work

Make the requested result. Keep a short record of which lesson shaped which part, with its source and the check performed.

06

Pass the quality gate

Check every output against its full checklist. Fix every failure and check again. Hand over the finished set only when all required checks pass.

Read left to right, then continue below. The same preparation and care apply across subjects; the selected method changes with the job.

What loads automatically? Codex reads AGENTS.md when starting a session inside this workspace. The baseline tells it to select and read the right skill. You do not need to name that skill. An ordinary ChatGPT chat outside the workspace does not automatically have these files.

What happens when...

Choose a situation. See what you provide and what I take care of.

You say

I will

    Where your information goes.

    Give me a file or link and tell me what it is for. I save it and tell you where it went.

    Rules for all work

    Our baseline: use plain English, read the right inputs, apply relevant learning, check work and correct confirmed method errors.

    AGENTS.md

    A school, business or other project

    Its goal, requirements, decisions and links to source files. Clients keep their existing profile and research folders.

    projects/{project}/PROJECT.yaml
    clients/{client}/PROFILE.yaml

    Preferences for one project

    Lasting corrections, the tasks they affect and how to check them. A one-time edit stays with that task.

    The project's or client's FEEDBACK.yaml

    Original learning material

    Videos, transcripts, articles and posts with their source links. Downloaded material is not automatically an adopted method.

    research/learnings/{topic}/

    The method we actually use

    Reviewed lessons and task steps. Correct the owning skill or its existing guide when a confirmed reusable error comes from it.

    skills/agent/{task}/SKILL.md
    Its linked references / skills/knowledge/

    Results and saved progress

    Results stay in the project's usual output folder. Save decisions, next steps and where lessons shaped the result.

    Project or client output folder
    workspace/tasks/{run}/

    Unsure where a file goes? Say ?save this for my history project? or ?study this as shared learning.? To drop files yourself, use research/inbox/ or the project/client inbox/. Tell me what you added; these folders are not automatically watched.

    What I need from you.

    The result you want, which project it belongs to, and anything new that matters. For school: the assignment. For a business: the idea and constraints. For a client: the brief and any new feedback.

    You do not need to repeat saved details or choose a skill. I ask only when a missing detail changes the work.

    What changes for you.

    Before substantial work, you get a short update: ?I will use these inputs and this method, make this result, and check these things.? This is not another approval step.

    At delivery, you get the actual work, its location and a short note on what was used and checked. The detailed source record stays with the task.

    The bouncer before delivery.

    Before making the work, I build its checklist from your request, the method, relevant lessons and saved rules. After production, I inspect each actual output and record the evidence.

    Every required check passes

    You receive the complete checked set. Approval applies to those exact files. A later change, crop or export needs another review.

    One check fails or is missing

    The output goes back for correction. I fix it, repeat the affected checks and run the whole gate again. A high average score cannot cancel a failure.

    A static ad, for example

    Check the customer insight, format method, exact copy, genuine quotes, supported claims, real assets, layout, phone-size readability and final export. Every requested ad gets its own checks.

    Something prevents completion

    I explain the missing evidence or decision and what is needed next. I keep rejected work out of the finished handoff. I do not invent a pass.

    How mistakes improve the next job.

    The method caused the error

    I correct the owning skill or its existing reference, add the relevant check there, and verify the case that failed. The lasting fix does not live only in chat.

    A step was skipped, or your preference changed

    If the instructions were correct, I fix the work and use the existing check. A preference for one project stays in its own feedback. It does not become a rule for every subject.

    Consistent standards, with honest limits. The gate rejects incomplete checklists, missing outputs and changed files. I still have to perform the actual review and judge whether the lessons were applied well. This is a required review process and checking command, not an installed T3 stop hook or a guarantee of perfect judgment.

    The shared baseline and project routes are installed

    Built from research, checked against real files.

    20 videos downloaded, 20 timestamped transcripts saved, and screen samples inspected from every video. The practical ideas were compared with current official documentation and the existing workspace.

    Review covered full transcripts or relevant sections and sampled screens, rather than continuous viewing of all footage. Some linked templates were unavailable without signup or purchase.

    Rollout status: 22 task routes are installed, including general research, business planning, other project work and workspace maintenance. 86 client folders have readable profiles and a place for feedback. Two client source maps have been reviewed in detail. The remaining maps are marked for review; missing client facts are left unknown. This is not a claim that every historical file or client deliverable has been checked.

    See the 20 research videos

      Technical reference: OpenAI's project instructions, OpenAI's skills documentation, and T3's source. The detailed internal evidence report is saved in the workspace.