Every file you paste, every task you stack, every "and also…" lands on the same desk. Past a point the agent stops working and starts guessing what you probably meant. This module shows you exactly where that point is, and the three habits that keep you on the right side of it.
THE BRUTAL TRUTHZido built a feature and reviewed it in the same session. The context went 8–9× heavier and the review approved every assumption the build had made. That is not a Zido story. It is physics. The same desk cannot check itself, and you are doing this every week.
Zido's lesson, replayed
REAL THG LESSON · SAME SESSION · BUILD + REVIEW
8–9×
That is how much heavier the context got when the review ran in the same session as the build. The reviewer needed the entire build history to make sense of the code, so it dragged every earlier turn, every tool output and every dead end along. The agent spent the desk re‑reading its own reasoning, then approved its own assumptions. Worthless and expensive at the same time. The fix costs nothing: new session or subagent, file + spec in, report out.
Press both. Same code, same bug, same reviewer model. The only difference is which desk the review runs on.
thg‑dev · agent session · build → review
press ▶ to replay
The three rules
These are not tips. They are the difference between an agent that ships and an agent that burns your afternoon. Every context failure at THG is a violation of one of them.
Never build AND review in one session
The builder's desk is full of the builder's reasons. A reviewer needs none of them.
Review needs the result, the spec, and a report format. Nothing else. Everything else on the desk is noise that makes the reviewer agree with the builder. Spawn a subagent (Module 05) or open a new session. Never type "now double‑check your work" into the session that did the work.
AT THG THIS ISBuild the Unibit lobby change in session A. Review it from a fresh session B with git diff and the ticket pasted in. Twenty lines of input instead of two hours of history.
One small step at a time
One task. Verified. Then the next. Not "do all seven tickets".
Every step you stack adds its files, its test output and its dead ends to the desk. Seven steps in one prompt means step seven is executed by an agent that can barely see step one. Do one, check it, commit it, then say what's next.
AT THG THIS IS"Implement the RTP display. Run the tests. Stop and report." Then: "Now the bonus cap. Stop and report." Not: "Implement RTP display, bonus cap, new endpoint, migrate the old one, update docs and write the release notes."
Plan, then execute
The plan lives in a file. The session lives and dies. Guess which one is your memory.
First session: produce PLAN.md: steps, decisions, done/next. Each execution session reads the plan and does one step. When a session bloats, you kill it without losing anything, because nothing important lived only on the desk.
AT THG THIS ISThe "keep the old endpoints" decision is written in the plan. Session four, at 15% context, still knows it. Without the plan, session one at 70% context had already forgotten it.
Context is not memory. The plan file is memory. The session is a disposable desk you should be happy to throw away. If throwing it away scares you, something important is only on the desk. Write it down now.
How to see the bloat before it costs you
You won't get a warning light. You get behaviour. Learn the behaviour, and act at the first sign, not the fourth.
Answers get vaguer and longer. It restates the task before doing it.
early
Finish the current step. Write state to the plan. New session for the next step.
"As we agreed…" followed by something you never agreed. Or a decision quietly reversed.
early
Stop. The decision was compressed. Re‑state it from the plan. If there is no plan, write one now, then restart.
The tool says it is compacting or summarising the conversation.
late
Whatever wasn't in a file is now lossy. Assume every earlier decision is at risk. Dump state to the plan immediately.
It re‑introduces a bug it fixed earlier, or edits a file it already finished.
late
The fix scrolled off the desk. Do not argue with it. Kill the session. Fresh session, plan in, one step.
"Can you remind me what we were working on?"
dead
Session is gone. If you have a plan file, you lost five minutes. If you don't, you lost the afternoon and you don't yet know which decisions.
The state handoff that makes fresh sessions free
This is what you paste at the top of a new session. Six lines. If you can't fill them in, you weren't keeping a plan and that is the real problem.
STATE HANDOFF (fresh session · paste first · then one step)
Goal: migrate Unibit lobby to the new API, keep old endpoints working.
Done: 1) endpoint map 2) adapters for /v2/lobby 3) tests green 41/41
Next: step 4 only — wire RTP display to /v2. Run tests. Stop and report.
Decisions that must survive: keep /v1/lobby live · bonus cap 2500× · no schema changes
Do not: re-plan, refactor outside step 4, or review your own work.
Simulator: watch the session bloat
You have six pieces of work to ship. Each can be done the pile‑it‑in way or the hygienic way. Both finish the task. Only one keeps the desk usable. The meter is your context load. Red is where the agent starts guessing. Ship all six and see what your habits cost.
Win condition: all 6 done, meter never past 60%, review done by fresh eyes, zero decisions lost. Anything else is how most afternoons actually go.
One session. Six tasks. Your call.
Done 0/6Peak 0%Sessions 1Decisions lost 0
context load 0%1.0× a clean session
freshthinningforgettingguessingdead
Agent quality: sharp. Every decision is on the desk and legible.
thg‑dev · session 1 · what the agent is doing
session 1 open · desk empty · PLAN.md: none
Quiz: prove it
Five questions. You need 4 right (80%) to pass. Fail and you retake it. Only passes are recorded.
Module 6 quiz
THE ONE THING TO REMEMBERBuild in one session. Review in another. Write the plan to a file so the next session starts smart. One small step, verified, then the next. Your context is not memory. The plan file is.