Agent Skills: Improve Processes

Improve Processes

Process-siren's job is semantic fidelity. Faithful conversion of a flawed process encodes the flaws with false precision. Use this skill when the source process needs improvement before — or alongside — Mermaid conversion.

When to Apply

Apply before converting when the source shows ANY of:

• Abstract verbs with no concrete action ("handle", "manage", "ensure")
• Conditions that cannot be evaluated by an AI agent ("when appropriate", "if needed")
• Missing entry or exit conditions
• Undefined actors ("we", "the system", "someone")
• Steps that do not change state (pure description, no action)
• No feedback loop or error path

Frameworks (Reference Only)

These frameworks share one principle — a process must be deterministic, auditable, and actor-owned.

Lean (Ohno) — eliminate steps that produce no state change; apply 5 Whys to trace ambiguity to its root

Six Sigma / DMAIC (Smith) — Define outcome, Measure current state, Analyze gap, Improve, Control recurrence

BPR (Hammer) — radical question: "If we started from scratch, what would this look like?"

Design Thinking (IDEO/Brown) — Empathize with the agent executing the process; design for their decision points

Systems Thinking (Senge) — identify feedback loops and side effects before encoding structure

Theory of Constraints (Goldratt) — find the bottleneck step; simplify around it before adding branches

Antifragility (Taleb) — prefer processes that improve under stress over processes that merely tolerate it

Excellence Checklist

Before converting, verify the source process satisfies:

• [ ] Clarity — no interpretive gaps; every term has one meaning
• [ ] Determinism — same input produces same output, independent of executor
• [ ] Minimal cognitive load — relies on structure, not memory
• [ ] Explicit feedback loops — error paths and retry conditions stated
• [ ] Measurable outcomes — each terminal state has an observable signal
• [ ] Visible constraints — blockers and preconditions named, not implied
• [ ] Ownership — every step names the actor
• [ ] Edge case coverage — at least one failure scenario is handled
• [ ] Teachable in 5 minutes — a novice can follow it cold
• [ ] Auditable — execution can be traced and verified after the fact

Pre-Conversion Completeness Gate

After reading the process-under-review and all linked or referencing files, evaluate this gate before any Mermaid conversion begins.

Question: Are all branches, conditions, and terminal states derivable from what has been read — with no unbound unknowns?

flowchart TD
    Read["Read process-under-review<br>and all linked/referencing files"] --> Inventory["Inventory all steps, conditions,<br>branches, and terminal states<br>found in the source material"]
    Inventory --> Gate{"Are all branches, conditions,<br>and terminal states derivable<br>from the source — with no<br>unbound unknowns?"}
    Gate -->|"Yes — all structure is derivable"| Triage["Proceed to Triage Protocol<br>then Mermaid conversion"]
    Gate -->|"No — unknowns remain"| Coach["Enter coach-mode<br>Stop conversion entirely"]
    Coach --> Report["Produce BLOCKED report<br>(see Coach-Mode Report Format below)"]
    Report --> Done(["Return report to process author<br>Await answers before any conversion"])

Why This Gate Exists

Converting an incomplete process produces a diagram that looks authoritative but encodes ambiguity as if it were resolved. An AI agent reading that diagram will follow the false structure and behave incorrectly. Coach-mode surfaces the incompleteness instead of hiding it.

Coach-Mode Report Format

When the gate returns NO, produce this report — do not produce any Mermaid:

CONVERSION ASSESSMENT

Goal:
- [One sentence stating what the process is intended to accomplish]

Source material read:
- [List each file or section examined, with path or reference]

What is known (derivable from source):
- [Fact 1 — cite the source section]
- [Fact 2 — cite the source section]
- ...

What is unknown or unbound:
- [Missing branch] | Gap: [what is undefined] | Source: [which file/section is silent on this]
- [Ambiguous condition] | Gap: [what observable fact is missing] | Source: [which file/section]
- [Undefined terminal state] | Gap: [what success/failure looks like] | Source: [absent from all files]
- [Step referencing undefined thing] | Gap: [the undefined reference] | Source: [where the reference appears]

Questions the process author must answer before conversion can proceed:

[Category — e.g., Branching Conditions]:
- [Question 1] (needed because: [why this blocks a specific diagram node or edge])
- [Question 2] (needed because: ...)

[Category — e.g., Terminal States]:
- [Question] (needed because: ...)

Decision:
- BLOCKED

Conversion will proceed once all questions above are answered.

Report field rules:

• "What is known" — list only facts directly readable or unambiguously derivable from the source files; cite the source for each
• "What is unknown or unbound" — list only genuine gaps; do not list things that are merely implicit if the implication is unambiguous
• "Questions" — one question per gap; ask only what is missing; do not ask about things already answered in the source
• "BLOCKED" is the only valid verdict when the gate returns NO; there is no partial conversion

Triage Protocol

flowchart TD
    Start(["Source process received"]) --> O{"Is the intended outcome<br>stated in one measurable sentence?"}
    O -->|"No"| FixO["Rewrite outcome statement<br>before proceeding"]
    O -->|"Yes"| A{"Is the actor named<br>for every step?"}
    FixO --> A
    A -->|"No — actor undefined"| FixA["Name actor per step;<br>ask user if ambiguous"]
    A -->|"Yes"| B{"Do all steps change observable state?"}
    FixA --> B
    B -->|"No — some steps are pure description"| FixB["Remove or rewrite no-op steps<br>as concrete actions"]
    B -->|"Yes"| C{"Are all decision conditions<br>evaluable without interpretation?"}
    FixB --> C
    C -->|"No — vague conditions remain"| FixC["Replace with observable facts:<br>exit code, file existence, string match"]
    C -->|"Yes"| D{"Are entry and exit<br>conditions explicit?"}
    FixC --> D
    D -->|"No"| FixD["Add entry precondition<br>and exit terminal state"]
    D -->|"Yes"| Convert(["Process is ready for Mermaid conversion"])
    FixD --> Convert

Practical Improvement Framework

Apply in sequence when rebuilding a weak process:

flowchart TD
    S1["1. Rewrite outcome as one measurable sentence"] --> S2
    S2["2. Replace abstract verbs with concrete actions"] --> S3
    S3["3. Add decision gates — If X, then Y"] --> S4
    S4["4. Define inputs and outputs for each step"] --> S5
    S5["5. Remove steps that do not change state"] --> S6
    S6["6. Add at least one correct execution example"] --> S7
    S7["7. Add at least one failure example"] --> S8
    S8["8. Stress-test: what happens at each edge case?"] --> S9
    S9["9. Time the walkthrough — can a novice follow in 5 minutes?"] --> S10
    S10["10. Confirm auditable — can execution be traced after the fact?"] --> Done(["Improved process ready"])