SKILL.md
reason
structure
graph LR
A[parse] --> B[branch]
B --> C[reduce]
C --> D[ground]
D --> E[emit]
style A fill:#e1f5ff
style B fill:#fff4e1
style C fill:#ffe1f5
style D fill:#e1ffe1
style E fill:#f5e1ff[[#parse]] → [[#branch]] → [[#reduce]] → [[#ground]] → [[#emit]] ^structure
parse
- receive
input^input - extract[^2]:
- #component/parameter — what varies ^parameters - #component/constant — what is fixed ^constants - #component/operation — what transforms ^operations - #component/unknown — what is sought^unknowns
- A question is a function awaiting its arguments[^3].
Identify what arguments it needs. ^function-analogy
branch
Generate multiple resolution paths before committing[^4]:
| Path Type | Tag | Purpose | Anchor |
|---|---|---|---|
| Typical | #path/typical | The conventional approach | ^path-typical |
| Alternative | #path/alternative | A valid but less obvious route | ^path-alternative |
| Boundary | #path/boundary | The edge case or inversion | ^path-boundary |
Assign rough likelihood. Retain all paths until [[#reduce|reduction]] forces convergence or reveals genuine divergence worth surfacing.
[!warning] Distribution Sampling ^distribution-sampling
This is not optional. Single-path reasoning collapses to the mode.
The goal is to sample from the full distribution of valid approaches.
reduce
Transform step by step[^5]. Each step follows the primitive: ^transformation-primitive
$$\text{input} \xrightarrow{\text{operation}} \text{output}$$
[!caution] Skip Nothing ^skip-nothing
An obvious step is still a step.
A hidden step is where errors hide.
five primitive operations
| Operation | Tag | Definition | Anchor | |
|---|---|---|---|---|
select |
#primitive/select | condition → one of two branches | ^op-select | |
sequence |
#primitive/sequence | output of one feeds [[SKILL#^input\ | input]] of next | ^op-sequence |
iterate |
#primitive/iterate | apply operation n times | ^op-iterate | |
recurse |
#primitive/recurse | define in terms of self, with base case | ^op-recurse | |
search |
#primitive/search | find first element satisfying constraint | ^op-search |
Name which [[SKILL#five primitive operations|primitive]] you are applying.
Substitution is explicit:
Given X, applying rule R yields X transformed by R
Never say "therefore" without showing the step that licenses it. ^explicit-substitution
ground
Before [[#emit|emitting]], verify:
- Each conclusion traces to stated premises only^premise-tracing
- Terms from separate contexts are not conflated^context-separation
- The result is terminal (no further reduction possible) or explicitly
partial (further [[SKILL#^input|input]] required) ^termination-check
convergence logic
| State | Tag | Meaning | Anchor | |
|---|---|---|---|---|
| Converged | #state/converged | confidence is high | ^state-converged | |
| Diverged | #state/diverged | present the divergence—it is the answer | ^state-diverged | |
| Continuing | #state/continuing | [[#reduce\ | reduction]] could continue | ^state-continuing |
| Blocked | #state/blocked | state what is missing | ^state-blocked |
emit
Output the synthesis.
| Context | Tag | Strategy | Anchor | |
|---|---|---|---|---|
| Simple query | #query/simple | direct answer | ^emit-simple | |
| Complex query | #query/complex | show the [[#^structure\ | structure]] that produced the answer | ^emit-complex |
| Creative task | #task/creative | weight toward [[#branch\ | boundary paths]][^6] | ^emit-creative |
| Analytical task | #task/analytical | weight toward [[#ground\ | convergent paths]][^7] | ^emit-analytical |
[!note] Trace Principle ^trace-principle
The trace is part of the answer when the trace is informative.
The trace is omitted when it would obscure.
[^1]: Recursive decomposition breaks complex problems into self-similar subproblems until base cases are reached. [^2]: Component extraction as defined in [[#parse|parse phase]] identifies [[SKILL#^parameters|parameters]], [[SKILL#^constants|constants]], [[SKILL#^operations|operations]], and [[SKILL#^unknowns|unknowns]]. [^3]: See [[SKILL#^function-analogy|function analogy]] for the functional interpretation of queries. [^4]: Multi-path generation prevents [[SKILL#^distribution-sampling|mode collapse]] by exploring [[SKILL#^path-typical|typical path]], [[SKILL#^path-alternative|alternative path]], and [[SKILL#^path-boundary|path boundary]] simultaneously. [^5]: Stepwise transformation using [[SKILL#^op-select|op-select]], [[SKILL#^op-sequence|op-sequence]], [[SKILL#^op-iterate|op-iterate]], [[SKILL#^op-recurse|op-recurse]], or [[SKILL#^op-search|op-search]]. [^6]: Boundary paths (see [[SKILL#^path-boundary|path-boundary]]) explore inversions and edge cases suitable for creative divergence. [^7]: Convergent paths (see [[SKILL#^state-converged|state-converged]]) prioritize analytical rigor through [[SKILL#^premise-tracing|premise-tracing]] and [[SKILL#^context-separation|context-separation]].