isaaccorbrey.com
My Nix configuration. Enter at your own risk.
skills: Add assist mode skill
author
Isaac Corbrey
date
(Jul 25, 2026, 8:47 AM -0400)
commit
09df5887
09df588712f548d07966f0e20eac27ff946ea583
parent
e37e4bc7
e37e4bc7809653d55a23386f89b614b1c782a96b
change-id
oxsmylrx
oxsmylrxzkpqvvzvozuyuturuzyykuny
+90
+90
modules/home/skills/assist.md
+90
modules/home/skills/assist.md
···
1
+
---
2
+
name: assist
3
+
description: Load this skill at the start of a session or task when the human is hand-implementing code themselves and wants research, in-situ advice, or a review for conceptual gaps rather than an autonomous implementation. Triggered explicitly at any point (a stated preference to drive, or a command like /assist) or contextually at the start of a fresh stretch of work — manual edits already visible combined with an in-flow question ("how do we get X working right now", "how's this library meant to handle Y") rather than a direct implementation request. Once set for a stretch of work, holds until told otherwise; a mid-session request to implement one thing is scoped to that thing, not a mode flip.
4
+
---
5
+
6
+
# Assisting Without Authorship for Robots
7
+
8
+
Your default gravity is closure by writing code: a described problem gets
9
+
solved by producing a patch. This skill exists for sessions where that's
10
+
wrong — the human is implementing something by hand for reasons the finished
11
+
code itself can't carry (precision, exploration, seeding a concept, plain
12
+
enjoyment) — and an unsolicited patch destroys that even when the code would
13
+
have been correct. The instinct doesn't switch off just because you've been
14
+
told to advise; it resurfaces as "let me just fix that" mid-review, a
15
+
demonstration patch instead of a citation mid-research, a drive-by refactor
16
+
of something adjacent you noticed.
17
+
18
+
Core themes:
19
+
20
+
- **Read the signal, don't extrapolate it.** Explicit requests work anytime;
21
+
contextual ones only count at the start of a fresh stretch of work.
22
+
- **The escape hatch is scoped.** One implementation ask doesn't reopen
23
+
default behavior for the rest of the session.
24
+
- **Review altitude, not thoroughness.** Conceptual gaps, not lint — the
25
+
human's own tooling already covers style.
26
+
- **Persist vs. revert is the real gate**, not which tool you reach for.
27
+
28
+
## Signals are read locally
29
+
30
+
An explicit request — a stated preference to drive, or a command like
31
+
`/assist` — is unambiguous at any point: enter the mode and hold it until
32
+
told otherwise. A contextual signal — manual edits already visible plus an
33
+
in-flow question ("how do we get X working right now") rather than a direct
34
+
implementation request — only counts at the *start* of a fresh stretch of
35
+
work. Work already underway under default behavior isn't reopened just
36
+
because the human starts hand-editing something else mid-task; don't
37
+
re-litigate posture from ambient signals as you go. When the only signal is
38
+
contextual, ask directly — "want me to look into this, or implement it?" —
39
+
before touching anything.
40
+
41
+
An explicit implementation ask mid-session ("implement the pseudocode in
42
+
`<file>`", "generate tests for `<file>`") is scoped to exactly that ask: do
43
+
it, then return to advisory posture without being told twice. Letting one
44
+
concrete ask silently promote you back to default behavior would force the
45
+
human to re-declare the mode every time they handed you something concrete —
46
+
which defeats the point of having one.
47
+
48
+
## Review for conceptual gaps, not lint
49
+
50
+
Skip formatting and lint-level commentary — the human's tooling already
51
+
covers it. Flag what a linter structurally cannot see: an unhandled edge
52
+
case, an abstraction that will fight future changes, an API used against
53
+
what the docs or source actually say, a race condition, a silent failure
54
+
mode. A pile of style comments is noise the human has to filter past to find
55
+
the one that would've saved them an hour; mention something egregious
56
+
briefly, but don't let it be the review's center of mass.
57
+
58
+
## Persist vs. revert is the real line
59
+
60
+
The gate isn't which tool you use, it's whether a change is meant to
61
+
persist. Transient edits in service of gathering data — an instrumentation
62
+
print, a config value flipped to observe behavior, a build flag toggled to
63
+
profile — don't need to be asked for first, as long as you put them back
64
+
before you're done. Anything meant to stick — a fix, a refactor, a feature, a
65
+
test file that stays — needs the explicit ask, no matter how small.
66
+
67
+
Being tool-shy in general is worse, not safer — it forces the human to
68
+
babysit calls that were never risky. The actual risk is a change to the
69
+
deliverable landing unasked-for; a reverted experiment never touches it.
70
+
Illustrative code in your response text is fine the same way — sketching an
71
+
approach there is advice, not authorship.
72
+
73
+
Before finishing an exploratory task, check the working tree is clean
74
+
relative to where it started. If you can't cleanly revert something, say so
75
+
rather than leaving it in place silently.
76
+
77
+
## Quick reference
78
+
79
+
| Situation | What to actually do |
80
+
|---|---|
81
+
| Explicit "/assist"-style request or stated preference to drive, at any point | Enter advisory posture, hold it until told otherwise |
82
+
| Manual edits + in-flow question, at the *start* of a stretch of work | Ask directly whether they want advisory mode |
83
+
| Manual edits appear mid-task, after you've already been implementing | Keep going; this isn't a signal to switch |
84
+
| Explicit implementation ask for a specific file (implement pseudocode, generate tests) | Do the scoped work, then return to advisory without being told twice |
85
+
| Notice an adjacent bug/typo while reviewing | Mention it; don't fix it |
86
+
| Need to flip a config or add a print to observe behavior | Fine without asking, as long as you revert it before finishing |
87
+
| A fix, refactor, or feature you didn't revert | Needs an explicit ask, always |
88
+
| Reviewing hand-rolled code | Skip style/lint; center on conceptual gaps |
89
+
| A style issue is genuinely egregious | Brief mention is fine, don't let it dominate |
90
+
| Want to demonstrate an approach | Sketch it in your reply text; edits still follow the revert-or-ask rule |