saaiasa

name	saaiasa
description	Classify utterances in a gate/procedure/document so the system doesn't answer the wrong thing.
output	{"format":"prose"}

Apply Intent + Speech Acts

Input: $ARGUMENTS

---

Overview

Classify utterances in a gate/procedure/document so the system doesn't answer the wrong thing (question vs command vs judgment vs guess). Many problems in procedure execution come from treating one speech act type as another.

Steps

Step 1: Extract Utterances

From the target artifact (gate, procedure, document):

1. Extract all question-like lines (containing "?", "whether", "how", "what", "is there")
2. Extract all instruction-like lines (containing imperatives, "should", "must", "do")
3. Extract all judgment-like lines (containing evaluative language, "good", "adequate", "correct")
4. Extract all prediction-like lines (containing "will", "would", "likely", "expect")

Step 2: Classify Each Utterance

For each extracted line, assign a speech act type:

Type	Definition	Example	Correct Response
Question	Seeks information	"What are the risks?"	Answer with information
Command	Directs action	"List all dependencies"	Execute the action
Request	Asks for action (politely)	"Could you check..."	Execute or negotiate
Constraint	Sets boundaries	"Must not exceed budget"	Honor as a limit
Judgment	Evaluates quality	"This is adequate"	Accept or challenge with evidence
Prediction	Forecasts outcome	"This will succeed"	Verify or challenge
Hypothesis	Proposes testable claim	"If X then Y"	Design test
Goal	States desired outcome	"We want to increase..."	Plan toward it
Story	Narrates events	"Last time, we..."	Extract lessons

Step 3: Identify Mismatches

Flag lines where the surface form doesn't match the function:

Surface Form	Actual Function	Problem	Fix
Question ("Is this good?")	Judgment (seeking validation)	Will get answer instead of evaluation	Rewrite as explicit judgment request
Command ("Determine if...")	Question (seeking information)	Will get action instead of answer	Rewrite as question
Judgment ("This works")	Hypothesis (untested)	Will be treated as settled	Rewrite as hypothesis to test
Question ("Don't you think...")	Command (seeking agreement)	Leading question	Rewrite as open question or explicit command

Step 4: Split Bundled Utterances

Identify lines that contain multiple speech acts disguised as one:

Before: "Is the goal clear and are the steps complete and has the timeline been validated?" After:

1. "Is the goal clear?" (Question — definition check)
2. "Are the steps complete?" (Question — completeness check)
3. "Has the timeline been validated?" (Question — validation check)

Each should be answered independently.

Step 5: Propose Rewrites

For each problematic utterance:

REWRITE:
Original: [the line as written]
Type detected: [what it looks like]
Type intended: [what it's trying to do]
Problem: [why the mismatch causes issues]
Rewrite: [clearer version]

Step 6: Report

INTENT + SPEECH ACTS AUDIT:
Target: [what was analyzed]
Lines analyzed: [N]

Type distribution:
| Type | Count | Issues |
|------|-------|--------|
| Question | [N] | [N with mismatches] |
| Command | [N] | [N with mismatches] |
| Judgment | [N] | [N with mismatches] |
| [etc] | | |

Mismatches found: [N]
Bundled utterances: [N] → split into [N] individual lines

Rewrites proposed: [N]
Top priority rewrites: [which are most impactful]

When to Use

• When a line reads like a question but functions like a command or judgment
• When a gate/procedure mixes description and recommendation
• When "what is being asked?" is unclear
• → INVOKE: /saqrc (question rewrite chains) for unclear questions
• → INVOKE: /saadag (detectors and generators) for finding ambiguity

Verification

• All utterances extracted and classified
• Mismatches between surface form and function identified
• Bundled utterances split into single speech acts
• Rewrites are clearer than originals
• Each rewrite preserves the original intent