diverge

Name: diverge
Author: scdenney

34stars0forks

How to install diverge?

scdenney/open-science-skills/diverge

$npx -y skills add scdenney/open-science-skills --skill diverge

Installs into the current project.

Summary

Before implementing, generate 3-5 conceptually distinct approaches labeled by creativity dimension (Novel, Surprising, Diverse, Conventional), then hold for selection. Brainstorm-then-select to resist defaulting to the most obvious solution.

Files · 1

View on GitHub

SKILL.md

1	# Diverge
2
3	Interrupt the default path of jumping to the most probable — and least creative — solution.
4
5	## Heritage and scope
6
7	This is an original Open Science Skills workflow grounded in Creative Preference Optimization (Ismayilzada et al., 2025; background in [`reference/creative-preference-optimization.md`](reference/creative-preference-optimization.md)). Standard preference alignment (RLHF/DPO) optimizes for the most human-expected output, which is by construction the least surprising one. The paper's most accessible remedy — its own "brainstorm-then-select" baseline — needs no fine-tuning: force divergence before convergence by generating several conceptually distinct approaches, requiring that at least one is surprising and one is novel, and deferring quality and implementation until after selection.
8
9	Use `diverge` for creative, architectural, or analytical work where more than one non-obvious solution exists. To isolate the brainstorm in a fresh Codex context, use the sibling `$diverge-codex`.
10
11	Model. This skill makes no separate model call — it runs entirely in whatever model and reasoning effort your current Codex session is already using, not a fixed pin. `$diverge-codex` spawns a fresh subagent for a clean context, but that subagent is still Codex, so it also runs at whatever model and effort your session is using — it is not a cross-model pin. For a genuinely separate, explicitly pinned model (`gpt-5.6-sol` at `xhigh`), that only happens from the Claude Code side, via `/oss:diverge-codex`.
12
13	## When to invoke
14
15	Use `$diverge <task>` when:
16	- multiple non-obvious implementations exist
17	- you want to avoid the conventional approach
18	- the task is creative, architectural, or analytical, not purely mechanical
19
20	Do not use for rote tasks with one correct answer (e.g., fix this syntax error).
21
22	## Behavior
23
24	Given the task supplied with the invocation:
25
26	### Step 1 — Clarify if needed
27
28	If the task is ambiguous about what "good" looks like, ask one focused question before proceeding. Skip this if the goal is clear. Do not ask about implementation details.
29
30	### Step 2 — Generate approaches
31
32	Produce 3–5 approaches that are genuinely conceptually distinct. Differences must be in underlying mechanism, not surface vocabulary.
33
34	Label each with its primary creativity dimension:
35
36	- [Novel] — semantically far from the conventional solution; different conceptual basis
37	- [Surprising] — violates the obvious assumption about how this should work; would not be the first answer
38	- [Diverse] — maximally different from the other approaches in this list
39	- [Conventional] — the expected path, included as a reference point
40
41	For each approach provide:
42	1. Core mechanism — one sentence naming the key insight
43	2. How it works — two to three sentences on the mechanism and what makes it distinct
44	3. Main tradeoff — one sentence
45
46	### Step 3 — Hold
47
48	Do not implement. Present all approaches, then ask:
49
50	> "Which approach should I pursue? Or should I synthesize elements from multiple?"
51
52	## Constraints
53
54	- At least one approach must be [Surprising]
55	- At least one must be [Novel]
56	- Approaches must not merely restate each other with different vocabulary
57	- Quality matters second in this phase — novelty and surprise come first
58	- Do not use markdown headers per approach — keep the list scannable
59
60	## After selection
61
62	Implement the selected approach directly. If the user asks to synthesize, identify which elements are mechanically compatible and propose a brief hybrid plan before implementing.

People who liked this also viewed

6 picks

Type

diverge

34stars0forks

View on GitHub

How to install diverge?

scdenney/open-science-skills/diverge

$npx -y skills add scdenney/open-science-skills --skill diverge

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Diverge
2
3	Interrupt the default path of jumping to the most probable — and least creative — solution.
4
5	## Heritage and scope
6
7	This is an original Open Science Skills workflow grounded in Creative Preference Optimization (Ismayilzada et al., 2025; background in [`reference/creative-preference-optimization.md`](reference/creative-preference-optimization.md)). Standard preference alignment (RLHF/DPO) optimizes for the most human-expected output, which is by construction the least surprising one. The paper's most accessible remedy — its own "brainstorm-then-select" baseline — needs no fine-tuning: force divergence before convergence by generating several conceptually distinct approaches, requiring that at least one is surprising and one is novel, and deferring quality and implementation until after selection.
8
9	Use `diverge` for creative, architectural, or analytical work where more than one non-obvious solution exists. To isolate the brainstorm in a fresh Codex context, use the sibling `$diverge-codex`.
10
11	Model. This skill makes no separate model call — it runs entirely in whatever model and reasoning effort your current Codex session is already using, not a fixed pin. `$diverge-codex` spawns a fresh subagent for a clean context, but that subagent is still Codex, so it also runs at whatever model and effort your session is using — it is not a cross-model pin. For a genuinely separate, explicitly pinned model (`gpt-5.6-sol` at `xhigh`), that only happens from the Claude Code side, via `/oss:diverge-codex`.
12
13	## When to invoke
14
15	Use `$diverge <task>` when:
16	- multiple non-obvious implementations exist
17	- you want to avoid the conventional approach
18	- the task is creative, architectural, or analytical, not purely mechanical
19
20	Do not use for rote tasks with one correct answer (e.g., fix this syntax error).
21
22	## Behavior
23
24	Given the task supplied with the invocation:
25
26	### Step 1 — Clarify if needed
27
28	If the task is ambiguous about what "good" looks like, ask one focused question before proceeding. Skip this if the goal is clear. Do not ask about implementation details.
29
30	### Step 2 — Generate approaches
31
32	Produce 3–5 approaches that are genuinely conceptually distinct. Differences must be in underlying mechanism, not surface vocabulary.
33
34	Label each with its primary creativity dimension:
35
36	- [Novel] — semantically far from the conventional solution; different conceptual basis
37	- [Surprising] — violates the obvious assumption about how this should work; would not be the first answer
38	- [Diverse] — maximally different from the other approaches in this list
39	- [Conventional] — the expected path, included as a reference point
40
41	For each approach provide:
42	1. Core mechanism — one sentence naming the key insight
43	2. How it works — two to three sentences on the mechanism and what makes it distinct
44	3. Main tradeoff — one sentence
45
46	### Step 3 — Hold
47
48	Do not implement. Present all approaches, then ask:
49
50	> "Which approach should I pursue? Or should I synthesize elements from multiple?"
51
52	## Constraints
53
54	- At least one approach must be [Surprising]
55	- At least one must be [Novel]
56	- Approaches must not merely restate each other with different vocabulary
57	- Quality matters second in this phase — novelty and surprise come first
58	- Do not use markdown headers per approach — keep the list scannable
59
60	## After selection
61
62	Implement the selected approach directly. If the user asks to synthesize, identify which elements are mechanically compatible and propose a brief hybrid plan before implementing.

People who liked this also viewed

6 picks

Type