Checks it understood you before writing anything
Before any code is written, TLM Forge repeats your request back in two or three sentences and lists anything it is unsure about. If a question could change the design, Claude asks you and waits. A misunderstanding costs you one reply instead of a wasted build.
![A "Team invites" feature request with the requirement "Invite links work for new users." highlighted, and a dotted line to a TLM Forge terminal at Stage 0.5 that restates the request, flags "[BLOCKING] Does a link expire?" and waits, with a card quoting the highlighted requirement and the footer "Nothing runs until you answer".](/images/features/F01_IntentCheck.webp)
How to use it
/tlmforge:buildYou never start this yourself. Run /tlmforge:build, describe your task in plain English, and this check runs right after TLM Forge picks the level (Light, Medium or Deep). It runs on every level.
Words you will see on this page
- Blocking question
- A choice only you can make that would change how the feature is built, so Claude waits for your answer.
- Informational question
- A small detail Claude decides on its own, writing down what it assumed so you can correct it.
- Light, Medium, Deep
- The three levels of checking TLM Forge picks from, based on how big and risky your change is.
- Stage 0.5
- The step right after TLM Forge picks the level, before any planning or code.
- Pro
- The paid TLM Forge plan, which adds extra tools such as a debugging path and checks built into the plugin that enforce some stops.
What happens next, depending on your request
| Your request | What Claude does |
|---|---|
| Has a blocking question | Asks you and waits. Nothing is written until you reply. |
| Is clear, but Claude's reading of it could easily be wrong | Asks you to confirm and waits. |
| Is completely clear, with only small details open | Says what it understood and keeps going in the same reply. You can still interrupt. |
| Is just a question, like "what does this file do?" | Answers it. No planning or process at all. |
Without TLM Forge vs with TLM Forge
| Without TLM Forge | With TLM Forge | |
|---|---|---|
| An unclear request | Claude fills the gap with a guess and builds on it. | Claude asks before writing anything. |
| Small assumptions | Hidden inside the code. | Written out where you can see and correct them. |
| When you spot a misread | At review time, after the work is done. | In the first reply, before any work is done. |
What it does
Many wasted AI coding sessions start with a request that was read slightly wrong. The AI fills a gap with a guess, builds on it for an hour, and you only find out when you review the result. This step catches that gap at the cheapest moment: before any plan or code exists. TLM Forge tells Claude to say, in two or three sentences, what you want built, what will behave differently, and what the end result should be.
Next, Claude lists its open questions and labels each one. A blocking question is a choice that would change the design, such as whether an invite link expires. An informational question is a detail Claude settles itself, and it writes down its assumption so you can correct it. If any question is blocking, or its reading could easily be wrong, Claude asks you and waits.
If your request is completely clear, Claude does not stall. It states its understanding and keeps going, so you can still read it and interrupt. After you answer a question, it confirms in one sentence and moves on: to the design steps on Medium and Deep work, or straight to building on Light work.
How it works, step by step
You describe the task
Run /tlmforge:build and say what you want in plain English. TLM Forge first picks a level and tells you in one sentence.
Claude repeats it back
In two or three sentences: what you want built, what will behave differently, and the expected result.
Claude labels its questions
Each open question is marked blocking (only you can decide) or informational (Claude decides and states its assumption).
Claude asks, then waits
If there is a blocking question, or its reading could easily be wrong, Claude asks you and waits before writing anything.
Claude confirms and continues
After your answer it confirms in one sentence, then moves on to the design steps, or straight to building on Light work.
What it looks like
What you see in your terminal: you ask for a feature, TLM Forge picks Deep, repeats the request back, asks one blocking question, and confirms after you answer. The format is real; the request is made up.
> /tlmforge:build add team invites to the admin panel
Going Deep: a new invite flow.
Understanding: admins invite teammates by email. The invitee signs up
from a link and joins the team with a member role. Admins can see and
cancel pending invites in settings.
[BLOCKING] Does an invite link expire, and after how long?
[INFORMATIONAL] Assuming one open invite per email address.
> Links expire after 7 days.
Confirmed: I'll build team invites with links that expire after 7 days.Why it is better
Misunderstandings show up in seconds
Without TLM Forge, Claude starts building on its first reading of your words. With it, that reading is shown to you first, and if something important is unclear, nothing is written until you reply.
You are only asked what matters
Without a rule for which questions matter, you can be quizzed on details that change nothing. With TLM Forge, details Claude can settle itself are decided with a stated assumption, so only real design choices interrupt you.
Assumptions are written down
Without TLM Forge, guesses stay hidden in the code. With it, every assumption is spelled out, so you can correct one in your next message instead of finding it in the finished work.
Good to know
- This is an instruction TLM Forge gives Claude, not a check built into the plugin that can force a stop. On Pro, the first stop the plugin enforces comes later, when the design is shown on Medium and Deep work.
- If your request is fully clear with no blocking questions, there is no pause. Claude states its understanding and keeps going.
- Plain questions such as "what does this file do?" skip the whole process and are just answered.
- On Pro, a request asking why something is broken goes to a separate debugging path that gathers evidence first.
- Setting TLMFORGE_APPROVAL_GATES=0 tells Claude to skip this step.