WSIB - What Skill Is Best

Input: $ARGUMENTS

Core Principles

Best means best NOW, not best in general. The right skill depends on current state — what’s already been done, what’s blocked, what the user knows. A skill that’s perfect for step 3 is wrong if you’re at step 1.
Intent over keywords. “I’m stuck” might need /diagnose, /sbfow, /how, or /emotion depending on what kind of stuck. Match the underlying need, not the surface words.
One skill, not a workflow. WSIB picks ONE skill. If the user needs a sequence, route to /fonss. The temptation to recommend three skills “just in case” dilutes the recommendation.
Confidence must be calibrated. HIGH means “this is clearly the right skill and alternatives are substantially worse.” MEDIUM means “this fits but a close competitor exists.” LOW means “best guess but the intent is ambiguous.” Don’t default to HIGH.
The runner-up is diagnostic. If the runner-up is from a completely different category than the best, the intent is probably ambiguous and you should say so. If runner-up is in the same category, your pick is probably right.
Fallback is not second-best. The fallback is what to run IF the best skill fails or produces weak output. It’s a recovery route, not a preference ordering.

Phase 1: Intent Extraction

Parse the prompt into structured intent:

[W1] SURFACE_REQUEST: [what the user literally said]
[W2] INTENT_TYPE: [goal | problem | question | decision | feeling | task | content | claim]
[W3] SPECIFICITY: [vague | moderate | specific]
[W4] URGENCY: [exploring | planning | ready-to-act | stuck | frustrated]
[W5] PRIOR_STATE: [fresh start | mid-process | iteration | recovery]
[W6] IMPLICIT_NEED: [what they probably also need but didn't say]

Intent Classification Table

If intent looks like…	Primary category	Check also
”I want X” / “I need X”	Goal → `/want`	`/gu` if goal is vague
”How do I X?”	Method → `/how`	`/foht` if specific
”Should I X or Y?”	Decision → `/decide`	`/dcp` if binary
”What’s wrong with X?”	Diagnostic → `/diagnose`	`/sbfow` if prior attempt failed
”I think X”	Claim → `/claim`	`/it` to decompose first
”Help me write/create X”	Content → `/create`	`/w` or `/wre` for writing
”I feel X” / “I’m frustrated”	Emotion → `/emotion`	`/cfr` for conflict
”Do X” / “Execute X”	Action → `/action`	`/handle` if vague
”Explore X” / “What about X?”	Search → `/search`	`/poa` for options
”Is X true?” / “Check X”	Validation → `/claim`	`/ver` for verification
”What should I do next?”	Sequence → `/next`	`/fonss` for multi-step
”Make this better”	Iteration → `/iterate`	`/evaluate` first

Phase 2: Candidate Generation

Generate top 3-5 candidate skills. For each:

[W7] CANDIDATE: /skill-id — FIT: [why it matches the intent] — WEAKNESS: [what it doesn't cover]
[W8] CANDIDATE: /skill-id — FIT: [why] — WEAKNESS: [what]
[W9] CANDIDATE: /skill-id — FIT: [why] — WEAKNESS: [what]

Selection Criteria

Score each candidate on:

Criterion	Question	Weight
Intent match	Does this skill address the core intent?	3x
State fit	Is this the right skill for WHERE the user is?	2x
Completeness	Will this skill resolve the need, or just start it?	1x
Specificity match	Does the skill’s scope match the request’s specificity?	1x

Phase 3: Selection

[W10] BEST_NOW: /skill-id
[W11] CONFIDENCE: HIGH | MEDIUM | LOW
[W12] WHY_BEST: [one sentence — what makes this better than alternatives]
[W13] RUNNER_UP: /skill-id
[W14] WHY_NOT_RUNNER_UP: [one sentence — what specifically makes it second]
[W15] RUNNER_UP_DISTANCE: [CLOSE | MODERATE | FAR]
[W16] FALLBACK_IF_FAILS: /skill-id
[W17] FALLBACK_TRIGGER: [what would indicate the best skill failed]

Confidence Calibration

Confidence	Condition
HIGH	Intent is clear, best skill is purpose-built for this, runner-up is FAR
MEDIUM	Intent is clear but 2+ skills fit well, or intent is slightly ambiguous
LOW	Intent is ambiguous, OR the best skill is a general router not a specific tool

If CONFIDENCE is LOW, add:

[W18] AMBIGUITY: [what's ambiguous about the intent]
[W19] CLARIFYING_QUESTION: [what would resolve the ambiguity]

Phase 4: Output

BEST_NOW: /skill-id
CONFIDENCE: HIGH | MEDIUM | LOW
WHY_BEST: [one sentence]
WHY_NOT_RUNNER_UP: [one sentence]
RUNNER_UP: /skill-id
RUNNER_UP_DISTANCE: CLOSE | MODERATE | FAR
FALLBACK_IF_FAILS: /skill-id
FALLBACK_TRIGGER: [when to switch]

→ INVOKE: /skill-id $ARGUMENTS

If CONFIDENCE is LOW, output the clarifying question INSTEAD of invoking.

Failure Modes

Failure	Signal	Fix
Router recursion	BEST_NOW is another router (`/meta`, `/handle`)	Routers route to routers = lost. Pick the concrete skill the router would pick.
Keyword matching	”decision” → `/decide` without checking if it’s actually a decision	Test against intent classification table, not just keywords
Default to category	Always picks `/want` or `/how` because they’re broad	Category skills are for ambiguous inputs. If intent is clear, pick the specific skill.
Overconfident HIGH	Everything marked HIGH confidence	If runner-up distance is CLOSE, confidence can’t be HIGH
Missing state context	Recommends `/gu` when user already understands their goal	Check PRIOR_STATE — mid-process users need different skills than fresh-start users
Workflow in disguise	Returns one skill but the user clearly needs a chain	If the need is sequential, say so and route to `/fonss`

Depth Scaling

Depth	Candidates Generated	Selection Criteria Scored	Disambiguation
1x	3	Intent match only	None
2x	5	All 4 criteria	Clarifying question if LOW
4x	7	All 4 + interaction check	Full ambiguity analysis
8x	10	All 4 + prior-state deep check	Multi-path recommendation

Default: 2x. These are floors.

Pre-Completion Checklist

Intent extracted (not just keywords matched)
Prior state considered (fresh start vs mid-process vs recovery)
At least 3 candidates generated with fit AND weakness noted
BEST_NOW selected with explicit reason it beats runner-up
Confidence calibrated (not defaulting to HIGH)
Fallback is a recovery route, not just second-best
If LOW confidence: clarifying question provided instead of blind invocation
Result is ONE skill, not a workflow (if workflow needed, route to /fonss)

Integration

Use from: /meta, /handle, /next
Differs from /extract: extract finds ALL relevant skills; wsib picks the ONE best
Differs from /fonss: fonss sequences multiple skills; wsib picks one
Differs from /given: given ranks by ROI for a goal; wsib matches by intent fit
Differs from /fnd: fnd is a registry lookup; wsib does intent analysis first
Routes to: whatever skill is selected

wsib - What Skill Is Best