You cannot fix what you refuse to write down. The tool-use log is the smallest honest thing a learner can do to keep an LLM in the drafting seat, and keep themselves in the loop.
It is one page, filled in once a week. Not a research protocol. Not a tracker anyone else grades. A private-by-default artifact you can share at whatever fidelity you choose. A person who writes down where a tool helped, where it blocked, and where it was unclear will notice things a person who does not, will not.
What goes on the page
Three columns and one line at the bottom.
Helped. A short entry for each use where the tool moved you forward on a job you still owned at the end. What you asked, what came back, what you kept, and one honest sentence on whether the sentence in front of you was still yours.
Blocked. A short entry for each use where the tool got in the way. Confident wrong answer, lost thread across turns, fluent output that took longer to correct than it would have taken to write. Name the failure mode plainly.
Unclear. The most useful column, and the one people skip first. A short entry for each use where you could not tell if you had been helped or blocked. The output looked plausible, you kept it, and you are not sure, later, whether you were right to.
The bottom line. One sentence for the week: at the end of this week, do I still own my work? Yes, no, not sure. That is the question the log is really for.
Why boring is the feature
A boring artifact is one that survives contact with a real week (Class A, empirical-in-session, from cohorts to date). Fancy trackers get abandoned by Wednesday. A single sheet with three columns and one line does not.
The mechanism, in the language of the field we build inside, is a comparison step. Active inference treats a learner as a system that keeps a model, makes a prediction, and updates when the world disagrees (Class E, Parr, Pezzulo, and Friston, Active Inference: The Free Energy Principle in Mind, Brain, and Behavior, 2022). The log forces the comparison you would otherwise skip. Without the write-down, the update does not land. With it, it is cheap.
The template, in code you can steal
The log we ship with the workshop is a plain Markdown file. No app, no login, no analytics (Class B, any participant can inspect the template end to end; Class C, it wires into nothing, on purpose, so nothing else can quietly read it). Copy it, put it in a folder you back up, open it once at the end of each work week.
# Tool-Use Log, week of YYYY-MM-DD ## Helped - Asked: ... Got: ... Kept: ... Still my sentence? yes / no ## Blocked - Asked: ... Got: ... Failure mode: ... Time cost vs. writing from scratch: ... ## Unclear - Asked: ... Got: ... Why I am not sure: ... ## Bottom line At the end of this week, do I still own my work? yes / no / not sure One sentence on why:
That is it. Add a fourth column if you want one. If a column stays empty three weeks running, cut it. The log is yours, and its job is to notice, not to be complete.
The falsifier
Here is the falsifier for the practice (Class F). If a cohort keeps the log for eight weeks and their bottom-line answer trends from "yes" toward "not sure" or "no," the log has done its job by surfacing a real problem, and we have failed to teach the loop. In that case the honest move is to publish the aggregate, name what went wrong, and rebuild the workshop.
What we have not yet done, and want to (Class U, unverified in our own data), is a matched comparison against a cohort that does not keep the log at all. Until we run that study, we hold the practice as a strong candidate.
How to read your own log on Friday
Once a week, read the whole page in one sitting. Do not edit last Tuesday to look better. Read for three things. First, ratio: are "helped" entries growing faster than "blocked," or the other way around. Second, repeats: if the same failure mode shows up three weeks running, the workflow is the problem, and the workflow is yours to change. Third, the "unclear" column: if it is the longest, you are in the danger zone the loop is designed to catch, and the honest response is to close the laptop for a stretch.
A weekly page, three columns, one honest sentence at the bottom, read on Friday. A private instrument for the person who has to live with the work.
Where to go next
- When to let the model draft: read alongside a month of your own "helped" entries.
- When to put the model down: cross-checked against your "unclear" column.
- The IamHITL workshop: the log held in a room with other people trying the same practice.
Evidence classes cited in this post: A (empirical-in-session, cohort observations), B (code and inspection, the workshop template), C (configuration and integration, the log wires into nothing by design), E (Parr, Pezzulo, and Friston, 2022), F (falsifier stated), U (matched comparison study not yet run in-house). Corrections welcome. Bring the counter-evidence and the record updates.