| name | pce - Procedure Engine - Comprehensive Analysis |
| description | The deep analysis engine. Routes input by type, applies appropriate analysis depth, produces structured understanding and action. |
| output | {"format":"prose"} |
Procedure Engine - Comprehensive Analysis
Input: $ARGUMENTS
Core Principles
-
Understand before acting. Explore the problem from multiple angles before recommending action. But don't explore forever — stop when you understand enough to act well.
-
Route by type. Different inputs need different treatment. A goal needs decomposition. A problem needs root cause. A decision needs comparison. Classify first.
-
Surface the hidden. The most important information is usually unstated — hidden assumptions, implicit constraints, presupposed beliefs. Find these before analyzing the stated content.
-
Testable output. Every recommendation should include how to verify it worked. Every claim should be possible to check.
-
Action-oriented. Analysis that doesn't lead to action is academic. End with specific, prioritized next steps.
Step 1: Classify Input
What kind of thing is this?
| Type | Signal | Core Need |
|---|
| GOAL | "I want to...", "How do I..." | Decompose and plan |
| PROBLEM | "Something is wrong...", "Why does..." | Diagnose and fix |
| QUESTION | "What is...", "Is it true that..." | Research and answer |
| DECISION | "Should I...", "Which option..." | Compare and choose |
| SITUATION | "Here's what's happening..." | Understand and respond |
| FEELING | "I'm frustrated/excited/confused..." | Clarify and redirect |
Step 2: Surface Hidden Claims
Every input contains unstated assumptions. Find them.
- Explicit claims: What's directly stated
- Implicit claims: What's assumed but not said
- Bundled claims: Multiple claims packed into one statement
- Presuppositions: What must be true for the statement to make sense
Example: "I need to scale our database" bundles: a scaling problem exists, the database is the bottleneck, scaling the database will fix it, we have resources to scale.
ARAW the highest-VOI hidden claim — the one that, if wrong, changes everything.
Step 3: Analyze by Type
Route to the appropriate analysis:
GOAL
- What's the foundational goal behind this? (trace upward)
- What are the subgoals? (decompose downward)
- What's the current state vs desired state?
- What's blocking progress?
- What's the highest-leverage next step?
→ May invoke: /gd, /gjs, /stg
PROBLEM
- What exactly is wrong? (specific symptoms)
- When did it start? What changed?
- What's the root cause? (ask "why" until you hit foundation)
- What are the fix options?
- Which fix addresses root cause, not symptoms?
→ May invoke: /rc5w, /araw, /lpd
QUESTION
- What kind of question? (factual, conceptual, evaluative)
- What's already known?
- What are the candidate answers?
- What evidence supports/contradicts each?
- What's the best current answer + confidence level?
→ May invoke: /araw, /bi, /vbo
DECISION
- What are the options? (including non-obvious ones)
- What criteria matter? (derived from purpose)
- How do options compare?
- What's the recommendation?
- How reversible is this?
→ May invoke: /cmp, /evaluation_dimensions, /sel
SITUATION
- What's happening? (neutral description)
- What does this mean for the user's goals?
- What options does this create?
- What's the recommended response?
→ May invoke: /araw, /grfr, /stg
FEELING
- What feeling is being reported, in the user's own terms?
- Default to engaging with the feeling. The user brought it up — that's sufficient permission.
- If exploring is desired: what goal, value, or unmet need might the feeling be pointing at?
- If a different underlying need is proposed, would addressing it but not the stated request count as clarification or substitution?
- What action follows from the chosen frame?
→ May invoke: /araw, /ve, /gu
Step 4: Synthesize
- Key findings: What did the analysis reveal? (3-5 bullet points)
- Tensions: Any unresolved conflicts between different findings?
- Confidence: What are you sure about? What's uncertain?
- Testable predictions: "If this analysis is correct, you should see X when Y."
Step 5: Recommend Action
Specific, prioritized next steps:
- DO_FIRST: [action] — [who does it] — [what it resolves]
- DO_NEXT: [action] — [who] — [what it resolves]
- MONITOR: [what to watch for] — [what it would mean]
Every action must pass this execution check:
- Trigger: when exactly should this action be taken?
- Actor: who does it?
- Procedure: what concrete step is performed?
- Output: what artifact or state change should result?
- Verification: how do we know the action worked?
Pre-Completion Check
