Maps a recurring strategic decision into an honest yes/no decision tree, rendered as Mermaid, with unanswerable questions marked as open gaps that double as a reading list.
SkillScore95Maintained
Automatic
◆Knowledge Management pack
Last verified 8/4/2026
What it changes
Passed 21 of 22 test cases with the skill, and 11 of 22 without it.
Without skill50%
With skill95%
Checks it passed that the model alone missed
The document header starts with # Decision:.
The response avoids using a 3-way fork node.
The flowchart uses binary yes/no decision nodes.
and 18 more in the full report
11 fixed10 passed either way0 still failing1 broke
Pass-rate lift
+45 pp
Tokens
+32%
Agent turns
0%
Response time
-33%
Model gemini-3.6-flashMeasured 8/4/2026v121 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 11 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 11
Prompt
When updating a decision tree document because a legal rule was updated, what is the mandatory rule regarding old, obsolete branch logic in the markdown document?
What was checked, without and with the skill
Checked
Without skill
With skill
The response specifies that superseded branch logic must be struck through or moved to Notes.
Missed
Passed
case-22 · gemini-3.6-flash · 8/4/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md107 linessingle file
/decision-tree — Map a Decision into a Yes/No Tree
Turns a fuzzy judgment call into an explicit, reusable map. The output is a dated
markdown artifact with a Mermaid flowchart (renders natively on GitHub; VS Code
needs a Mermaid preview extension).
The diagram is the cheapest part. What it forces is the product:
Every branch becomes an observable question (a fact you can check), not a judgment word
Every question you can't currently answer gets marked as an OPEN GAP: the tree
doubles as a map of what still needs to be learned or read. A finished tree with
zero gaps means the decision is understood; gaps show exactly where understanding stops.
Input
A decision to map, e.g. /decision-tree can this client data enter an AI tool?
If no argument: ask which recurring decision to map. Push back if it's a one-off
choice. Trees are for decisions that recur; a one-off gets a recommendation, not a tree.
Method
1. Frame the root
Pin down conversationally, not as a form:
The decision, stated as a single question with a bounded set of outcomes
Who decides — three cases, not two: your own decision gets a tree; someone
else's gets a recommendation map, framed as such; a jointly owned decision —
several deciders, each holding a private walk-away (a negotiation) — isn't one
tree at all. One tree flattens exactly the disagreement worth seeing; give each
party their own. And when a decider speaks for an interest they don't fully bear
the risk of ("for the department," for a company), ask which parts of their
position are personal and which are the principal's — the two blend by default
and quietly distort every branch.
Reversibility: reversible decisions get shallow trees and a bias to act;
irreversible ones earn depth
2. Force the questions
Every node must be answerable by an observable fact, not a judgment.
"Is it sensitive?" hides judgment; split it into checkable criteria.
Yes/no only. A three-way fork is two questions; split it.
Order by cheapness and kill-power: cheapest-to-answer, most-decisive
questions go closest to the root.