A two-week deliverable in ChatGPT almost always ends the same way: 30 scattered chats, three contradictory versions of the same outline, and an evening lost to “I know I wrote that paragraph somewhere.” The problem is not the tool — it is the lack of structure. This guide gives you a spec.md plus log.md rhythm inside one Project that turns the same effort into something that compounds instead of fragments.
What this tutorial solves
A 2-week task in ChatGPT usually devolves into 30 scattered chats, contradictory drafts, and “where did I write that?” frustration. This guide gives you a structure that compounds instead.
Who this is for
Anyone working on a single deliverable that will take more than 5 chat sessions: a thesis chapter, a launch plan, a long content series, a job-search campaign.
When to reach for it
The task obviously exceeds one sitting and involves multiple sub-tasks.
When this is NOT the right tool
Single-session work, exploratory questions, or anything where you do not yet know the rough shape of the deliverable.
Step by step
- Define the final deliverable in one sentence and one paragraph. If you cannot, the task is not ready for AI yet.
- Break it into 4-8 sub-tasks. Each sub-task is one chat’s worth of work.
- Create a Project. Inside, two files: “spec.md” (what you are building) and “log.md” (decisions and outputs as you go).
- For each sub-task: open a new chat in the Project. Start by stating the sub-task and the relevant section of spec.md. End by copying outputs and decisions into log.md.
- When two sub-tasks contradict, do not just pick one in the moment — write the contradiction into log.md and resolve it explicitly in a dedicated chat.
- Every 3 sub-tasks, re-upload log.md to the Project. Without this, ChatGPT loses old decisions and re-suggests them.
- Final assembly: a dedicated chat that reads log.md and stitches outputs into the deliverable.
Recommended workflow
A 6-week thesis chapter: spec.md describes the argument. 8 sub-tasks (lit-review section, data, methods, etc.). log.md grows weekly. Final chat reads log.md and produces a 30-page draft.
Common mistakes
- Skipping spec.md. You will rebuild it in your head every chat, badly.
- Letting log.md grow without structure (no headings, no dates).
- Trying to do the final assembly chat without re-uploading the latest log.md.
- Mixing project work with random questions in the same Project — Project Instructions then drift.
- Relying on Memory instead of files. Memory is unstructured and partial; a file is yours, structured, and exportable.
- Cramming 12 sub-tasks into one Project. Past 10, the file weight degrades responses — split into two Projects sharing the same spec.md.
Advanced tips
- Use headings in log.md (## 2026-05-19 — methods section). ChatGPT can then reference specific dates.
- Pin one chat per sub-task — these are your milestones, easier to revisit than scrolling history.
- When stuck, re-read spec.md aloud to ChatGPT and ask “is what I have done in log.md still on-track?” Calibration step.
FAQ
- Why not Memory?: Memory is unstructured and unreliable for multi-week tasks. A file you control is.
- How many sub-tasks is too many?: Past ~10, the Project file weight degrades responses. Split into multiple Projects with shared spec.md.
- What if the task scope changes mid-way?: Update spec.md explicitly with a dated note (“2026-05-15: scope reduced to chapters 1-3”). Do not silently edit — the contradiction trail matters.
- Can I share this with collaborators?: Yes, if you use a Team workspace. The Project files become shared context for everyone working in it.