DESIGN -----> MAP -----> BUILD -----> DIAGNOSE -----> PERSIST -----> COMMUNICATE
What are      Who owns   Does this    When did the    Is this fix    What does it
the actual    each       still fit    symptom first   a rule, a      do, and where
requirements? step?     in a loop?   appear?         reference, or  is the human
                                                     a procedure?   gate?

+=====================================================================+
|                    WORKFLOW DESIGN & DIAGNOSIS                        |
|                                                                      |
|  PART 1: DESIGN          PART 2: MAP           PART 3: BUILD        |
|  +------------------+   +------------------+   +------------------+  |
|  | C1: Step > Session|   | C5: 3 Criteria   |   | C9: Build Loop   |  |
|  | C2: Requirements  |   | C6: Two Maps     |   | C10: Escalation  |  |
|  | C3: Compute, not  |   | C7: 3 Errors     |   +------------------+  |
|  |     write         |   | C8: Map Owner    |                        |
|  | C4: Window Test   |   +------------------+                        |
|  +------------------+                                                |
|                                                                      |
|  PART 4: DIAGNOSE        PART 5: PERSIST       PART 6: OPERATE      |
|  +------------------+   +------------------+   +------------------+  |
|  | C11: 4 Causes    |   | C13: Reaction -->|   | C17: Bad Output  |  |
|  | C12: Diagnostic   |   |      Instruction |   |      Gets Out    |  |
|  |     Sequence      |   | C14: Rule/Ref/   |   | C18: Say What It |  |
|  +------------------+   |      Procedure   |   |      Does + Gate |  |
|                          | C15: Find Friction|   +------------------+  |
|                          | C16: Measure      |                        |
|                          +------------------+                        |
+=====================================================================+

"I use AI"                        "Our workflow uses AI"
+---------------------------+     +---------------------------+
| Personal habit            |     | Repeatable process        |
| Lives in your head        |     | Named steps, fixed owners |
| Varies by mood            |     | Same every time           |
| Disappears on your leave  |     | Works whoever's at desk   |
| You judge it AFTER        |     | You decide BEFORE         |
+---------------------------+     +---------------------------+
       Only this one                    Only this one
       is comfortable                   grows in value
                                        AND can be wrong
                                        in a way that
                                        KEEPS being wrong

A vague request:  "We need better reporting"
                         |
                         v
               THE TASK DEFINITION
        +--------------------------------+
        | 1. WHAT is produced?           |
        | 2. FOR WHOM?                   |
        | 3. HOW OFTEN does it run?      |
        | 4. FROM WHAT DATA?             |
        | 5. IN WHAT FORMAT?             |
        +--------------------------------+
                         |
                         v
Each answer = something you can CHECK
Each one still MISSING = a question you OWE somebody
                         (NOT a blank for the model to fill)

PASS 1: EXTRACT                    PASS 2: ATTACK
"Extract every distinct            "Which of these are ambiguous?
 requirement. For each:             Which could two people read
 - short label                      differently? Which implies a
 - exact source section              requirement stated only
 - ambiguous enough to               indirectly?"
   need clarification?"
                 |                              |
                 v                              v
         What was WRITTEN              What was MEANT
         (cheap to find)               (expensive to miss)

SYNTHESIS                          CALCULATION
(gathering + shaping options)      (numbers that matter)

AI does this WELL                  AI CAN do this
Holding many factors at once       BUT: you can't tell from the
without losing the thread          page whether it DID

                                   A written figure and a
                                   calculated one LOOK THE SAME

BAD:  "Recommend the headcount needed next quarter"
       --> AI WRITES a number into a sentence (you can't verify it)

GOOD: "Using code execution on the attached ticket data,
       CALCULATE quarterly volume growth and average tickets
       resolved per analyst. THEN recommend headcount and
       SHOW THE ASSUMPTIONS."
       --> Separates computation from recommendation
       --> Assumptions are visible = somebody can disagree
          with the recommendation WITHOUT redoing the arithmetic

SYNTHESIS STEP                     JUDGMENT STEP
Runs on information you            Runs on information you
CAN supply                         CANNOT supply (or wouldn't)

--> Hands over well                --> Stays human

Examples:                          Examples:
- Trend analysis                   - Risk appetite
- Option generation                - A hiring freeze
- Data patterns                    - Whether this client will
                                     accept one more delay

+--------------------------------------------------+
|              THE CONTEXT WINDOW                    |
|                                                    |
|  [Your prompt] [Files you supply] [Conversation]  |
|                                                    |
|  Can I put EVERYTHING this step needs in here?     |
|                                                    |
|   YES --> step is ELIGIBLE (not yet owned!)        |
|   NO  --> step is MINE                             |
+--------------------------------------------------+
     |
     |  Things that DON'T fit:
     |  - Decisions made in meetings, not datasets
     |  - Internal announcements not in documents
     |  - Risk appetite
     |  - Relationship context ("a human touch")
     |  - Anything after the model's knowledge cutoff
     v
 Passing the window test makes a step ELIGIBLE.
 It still must survive the 3 criteria (Part 2).

LAYER 1: ELIGIBILITY  ------>  LAYER 2: OWNERSHIP  ------>  LAYER 3: IMPLEMENTATION
"May AI touch this?"            "Who owns the step?"         "What carries it?"

Two gates:                      Three criteria:               Decided LAST:
1. May this data enter          1. Reversibility              - Skill
   this tool?                   2. Stakes                     - Project knowledge
2. Can everything be            3. Accountability             - Code execution
   supplied? (Window test)                                    - Standing instruction
                                                              - Named review gate
A NO at either gate
ends the question.              NEVER let Layer 3
                                influence Layer 2!
                                "Must be computed" belongs
                                in Implementation, NOT
                                in Ownership.

+------------------+  +------------------+  +------------------+
|  AI-APPROPRIATE  |  |  COLLABORATIVE   |  |  HUMAN-RETAINED  |
|                  |  |                  |  |                  |
| AI may own it    |  | AI produces      |  | A person owns    |
|                  |  | Named person     |  | it outright      |
| Can be undone +  |  | JUDGES           |  |                  |
| somebody judges  |  |                  |  |                  |
+------------------+  +------------------+  +------------------+

+-------------------------------------------------------------------+
|                                                                     |
|  1. REVERSIBILITY    Can it be undone if AI gets it wrong?         |
|     - Draft can be rewritten          --> tolerates delegation      |
|     - Sent email can't be unsent      --> does NOT                  |
|     - Cost of error = a REDO, not the error itself                  |
|                                                                     |
|  2. STAKES           What does a mistake cost? (BAD CASE, not avg) |
|     - Misfiled label = 1 minute                                    |
|     - Wrong penalty figure = whatever the penalty IS                |
|                                                                     |
|  3. ACCOUNTABILITY   Who is answerable for this step's outcome?    |
|     - Is this step THE CALL they answer for?                        |
|       --> Human (the decision itself)                               |
|     - Or is it an INPUT someone judges BEFORE it acts?              |
|       --> Can be handed over (accountability sits at the judging)   |
|     - Accountability NEVER moves onto the machine                   |
|       (the machine cannot be asked)                                 |
|                                                                     |
+-------------------------------------------------------------------+

Both workflows, despite being completely unrelated:

MECHANICAL steps  ---------> AI
DRAFT steps       ---------> AI (collaborative if high stakes)
FIGURE-CHECKING   ---------> Human
IRREVERSIBLE step ---------> Human

Same criteria --> same pattern --> in two different fields

ERROR 1: HALO DELEGATION
+--------+    "This went well"    +--------+
| Step A | -------- X ---------> | Step B |
| (draft)| tick carried forward  | (decide)|
+--------+                       +--------+
Drafting competence is evidence about the DRAFT.
It is NOT evidence about the DECISION.

ERROR 2: UNSTAFFED GATE
+--------+    "Reviews become     +--------+
| AI     | --> glances, then  --> |   ?    |
| drafts |     clicks             | nobody |
+--------+                       +--------+
Nothing was decided. Nobody changed the map.
But the step is now AUTOMATED.
The map won't tell you (it records what you DESIGNED,
not what is RUNNING).

ERROR 3: MAPPING THE TOOL INSTEAD OF THE WORK
+--------+         +--------+
| Cool   | ------> | Step   |
| Skill  |         | shaped |
|        |         | like   |
|        |         | Skill  |
+--------+         +--------+
Team maps around features they LIKE.
The workflow uses the tool well.
It's just not the WORK.

+------------------------------------------+
|         THE DELEGATION MAP                |
|                                           |
|  MUST HAVE:                               |
|  1. ONE named owner (a person, not team)  |
|     --> answerable for map = reality      |
|                                           |
|  2. A REVIEW DATE                         |
|     --> quarterly is reasonable default   |
|     --> sooner if volume, staffing, or    |
|        legal position changes             |
+------------------------------------------+

        +----------+
        | IDEATE   |  Produce options
        +----+-----+
             |
             v
        +----------+
        |PROTOTYPE |  Make one concrete enough to react to
        +----+-----+
             |
             v
        +----------+
        | FEEDBACK |  Shows what's wrong (couldn't know in advance)
        +----+-----+
             |
             v
        +----------+
        | REFINE   |  Fix that
        +----+-----+
             |
             +-------> back to IDEATE (next cycle)

DESCRIPTION PROBLEM                FEATURE PROBLEM
"I described it badly"             "This needs a different capability"

Fix: describe it better            Fix: use the right feature
(better prompt)                    (e.g., code execution for numbers)

Cycle 3 was this.                  Cycle 2 was this.

MONTH 1:              MONTH 6:
One small team        Three departments
uses the dashboard    open it every Monday
internally            + one puts numbers
                      in a board report
        |                     |
        v                     v
   "A tool"             "INFRASTRUCTURE"
   (prompt-and-iterate) (needs engineering)

TIMELINE OF ONE SESSION
|                                                              |
|  [First response]  [Middle of session]  [After weeks]        |
|       |                    |                  |              |
|       v                    v                  v              |
|  UNDER-          CONTEXT              STALE                  |
|  SPECIFICATION   OVERLOAD             CONFIGURATION          |
|                                                              |
|  + WRONG FEATURE OR MODEL (shows as one repeatable error)    |
|                                                              |

Same shape as the unstaffed gate (C7):
- Both are controls you set up CORRECTLY
- Both decay while everyone ASSUMES they hold
- Both have ZERO signal at moment of failure
- Both found by a SCHEDULED REVIEW

CHEAP                                              EXPENSIVE
  |                                                    |
  v                                                    v
+--------+   +--------+   +--------+   +--------+   +--------+
| 1.     |   | 2.     |   | 3.     |   | 4.     |   | 5.     |
| PROMPT |-->| SESSION|-->| FEATURE|-->| CONFIG |-->| TASK   |
|        |   | LENGTH |   | & MODEL|   |        |   | FIT    |
| Re-read|   | Over-  |   | Right  |   | Current|   | Can    |
| against|   | loaded?|   | tool?  |   | date?  |   | this   |
| 5 parts|   | Fresh  |   | Right  |   | Check  |   | be     |
|        |   | session|   | tier?  |   | sources|   | done?  |
| seconds|   | fixes? |   |        |   |        |   |        |
+--------+   +--------+   +--------+   +--------+   +--------+
                                                     Most people
                                                     START here
                                                     (wrong!)

REACTION                              INSTRUCTION
(names how output FEELS)              (names what to CHANGE)

"Too generic"          -------->      "Name the audience and the
                                       one action I want them to take"
                                       LEVER: The prompt

"Wrong tone"           -------->      "Add a tone constraint that
                                       applies to every draft"
                                       LEVER: The standing instruction

"Missed the point"     -------->      "State the single question the
                                       output must answer, up front"
                                       LEVER: The prompt

"It's using old data"  -------->      "Replace the source document
                                       in the knowledge base"
                                       LEVER: The knowledge

+-------------------+   +-------------------+   +-------------------+
|      RULE         |   |    REFERENCE      |   |    PROCEDURE      |
|                   |   |                   |   |                   |
| A limit that      |   | Material the work |   | A sequence of     |
| ALWAYS applies    |   | has to READ       |   | steps             |
|                   |   |                   |   |                   |
| HOME:             |   | HOME:             |   | HOME:             |
| Standing          |   | Knowledge base    |   | Skill             |
| instruction       |   | (on the Project)  |   | (at account level)|
|                   |   |                   |   |                   |
| Changes:          |   | Changes:          |   | Changes:          |
| BEHAVIOR inside   |   | What is KNOWN     |   | HOW work is DONE  |
| a Project         |   |                   |   |                   |
+-------------------+   +-------------------+   +-------------------+

THE MARKETER (did it right):
- Noticed: briefs always miss target segment + bury CTA
- Wrote TWO standing instructions into the briefs Project
- Result: every draft arrives correct, for EVERYONE who opens the Project

THE ANALYST (did it wrong):
- Figured out: revenue report must exclude canceled orders
- Typed reminder into chat EVERY MONTH
- It worked every month... then he went on leave
- Colleague ran the report --> numbers went out WITH canceled orders

The fix existed. It was just in a place only one person could find.

1. CONSOLIDATE first   --> reduces NUMBER of steps
   (3 prompts that each need same background = 1 prompt)

2. PROMOTE second      --> reduces COST of remaining steps
   (move repeated patterns into configuration)

Do consolidate FIRST because promoting a step
you're about to merge away is wasted work.

Run the workflow ONCE, unchanged. Record:
1. How long it took
2. How many revision rounds
3. How many manual steps

+------------------------------------------------------------+
| YOU CANNOT REBUILD A BASELINE AFTERWARDS.                   |
| Your estimate of the old way will FLATTER YOU.              |
+------------------------------------------------------------+

OLD WORKFLOW  =====>  compare  <======  NEW WORKFLOW
(for 2-3 cycles)      |
                       v
                 EVIDENCE that new is better on YOUR metric
                 + working FALLBACK if it's not
                 + AGREEMENT of people whose work changed
                 (turns "we're changing your job" into
                  "we ran both, here's what happened")

BAD OUTPUT GETS OUT
         |
         v
+------------------+     +------------------+
| 1. STOP          |     | 2. SCOPE         |
| How is it halted?|     | How many outputs  |
| Who can do it?   |     | were affected?    |
| Need permission? |     | Where did they go?|
+------------------+     +------------------+
         |                        |
         v                        v
+------------------+     +------------------+
| 3. TELL          |     | 4. CHANGE MAP    |
| Who is told?     |     | Which of the 4   |
| By when?         |     | causes was it?    |
| (customer, owner,|     | Step classified   |
|  risk/compliance)|     | wrong, or gate    |
| Late disclosure  |     | not staffed?      |
| costs FAR MORE   |     | Change + date     |
| than early one   |     | = fix outlasts    |
+------------------+     | the week everyone |
                         | was paying        |
                         | attention         |
                         +------------------+

1. Say what the tool DOES
2. Name the human CHECKPOINT
3. (For trust-critical audiences) Say what it REFUSES
   --> proves boundary was designed in advance, not found by incident

Can this data enter this tool? --> NO --> HUMAN (governance gate)
         |
        YES
         |
Can everything it needs be supplied? (Window test) --> NO --> HUMAN
         |
        YES (step is ELIGIBLE)
         |
Can it be undone?
   NO + externally binding --> HUMAN
   YES --> check stakes + accountability
              |
         What does error cost (bad case)?
         High + this IS the decision --> HUMAN
         High + someone judges after --> COLLABORATIVE
         Low + reversible + someone judges --> AI-APPROPRIATE

When did the symptom first appear?
    |
    +-- Wrong from first answer
    |     --> Probably UNDER-SPECIFICATION
    |     --> Test: re-read prompt, is missing thing in it?
    |
    +-- Started fine, then degraded
    |     --> Probably CONTEXT OVERLOAD
    |     --> Test: restate in fresh session
    |
    +-- One repeatable error type
    |     --> Probably WRONG FEATURE/MODEL
    |     --> Test: run once with right feature/stronger tier
    |
    +-- Used to work, now doesn't
    |     --> Probably STALE CONFIGURATION
    |     --> Test: check date on instructions/knowledge
    |
    +-- None of the above fixed it
          --> EXPECTATION MISMATCH
          --> Reshape the task

Will this correction be needed again? --> NO --> leave it
         |
        YES --> PROMOTE it
         |
Is the fix...
    |
    +-- A limit that always applies?
    |     --> RULE --> Standing instruction
    |
    +-- Material the work has to read?
    |     --> REFERENCE --> Knowledge base
    |
    +-- A sequence of steps?
          --> PROCEDURE --> Skill

DELEGATION MAP: [workflow name]
Owner: [one named person]
Last reviewed: [date]
Next review: [date]

ELIGIBILITY (both must pass before any step below is mapped)
  Data gate:  [may this data enter this tool? who confirmed, when?]
  Input gate: [can everything each step needs actually be supplied?]

STEPS
#  Step                  Owner          Criterion                  Carried by
1  ...                   AI             reversible, low stakes     Skill
2  ...                   AI             reversible, checked at #4  code execution
3  ...                   Collaborative  high stakes, human judges  Skill + gate
4  ...                   Human          this step is the decision  named reviewer
5  ...                   Human          irreversible               named signer

GATES (one line per collaborative or human step)
  Step 3: reviewed by [name], at [when], checking [what specifically]
  Step 4: approved by [name], before [what happens next]

BASELINE (measured before any change)
  Time per run: ___   Revision rounds: ___   Manual steps: ___
  Metric this workflow optimises for: [time | consistency | accuracy | variance]

IF IT GOES WRONG
  Stopped by:   [who, and can they act without permission?]
  Scope check:  [how do we find out how many outputs were affected?]
  Told:         [who, within what time]
  Map change:   [recorded here, with a date]