On 10/8/26 11:02 AM, Damian Rouson wrote:
I’m curious about the extent to which skill files are helpful. I suspect they
can be very helpful, but I don’t have enough experience to know for sure. A
student introduced me to an interesting approach: point the LLM to code written
in the manner you want and let the LLM write its own skill file. This gives some
insights into the how the LLM reasons about the code and the broad conclusions
it reaches. I was very impressed with some of the high-level concepts and
patterns the LLM extracted from my codes. But I haven’t made extensive use of
the skill file yet.
D
Skill files are very useful. Over the last several months I have established the
following. The last one about sending patch files I have only used once. As one
uses these you can modify and add things to constrain the skills or guide the LLM.
There are times when I have suggested that the LLM I use avoid making a
particular error and at times it has suggested a modification to a skill file.
It will also identify contradictions between memories and guidance so i can
chose to fix those.
I would say skill files are a key element.
Jerry
On Thu, Oct 8, 2026 at 09:42 Jerry D <[email protected]
<mailto:[email protected]>> wrote:
Hello all,
I am sure you are all aware that I have been using LLMs for testing and
debugging gfortran a lot lately.
I thought I would share a bit about why I have been doing so.
First, I do want to see bugs fixed and get gfortran in good shape so that
when
the next standard starts to emerge we will be in a better position to focus
on
implementation of the new.
Second, I think it is important for us to learn as much as we can about this
new
technology and how to use it. There are advantages and disadvantages in
doing so.
On the plus side these things can run hundreds of tests and variations on
code
to identify defects in the compiler. This I think is valuable by itself. The
downside of this testing is one can be overwhelmed by the number of defects
found. In spite of our best efforts over the years, we still have gaps in
implementation of features and issues with already implemented features.
This
can be frustrating and disappointing.
Yes, the LLMs can identify fixes rather quickly. And, sometimes they can be
very
unreadable. Part of my efforts here involve experimenting with prompting and
rules to move toward better readability. This includes the simpler the
better.
In this process I want to acknowledge the excellent feedback from Mikael.
It is
obvious the LLMs do not have high level design approaches. I consider
myself a
novice in this regard as well.
Third, I do not want to keep marching along here without getting some
feedback
from everyone on this subject. I actually have about 25 patches sitting on
my
worktree to be submitted. Most all have PR numbers assigned. Some are two or
three months old.
Shall I keep moving along with this? In the long term what do we do with
use of
LLM tools?
Do we need to set up an online meeting to discuss?
As always, best regards to all.
Jerry
