How to build Super Intelligence workflows instead of asking random questions is the shift from occasional AI help to reliable human-directed operating capability. A random question may produce a useful answer. A workflow defines what happens before the question, what information enters, how the answer is checked, who is allowed to act, where the result goes and what happens when something fails.
The important difference is that a workflow reaches the real task. A study explanation ends with independent practice. A planning conversation ends with a checked schedule. A document summary ends with a usable record linked to the source. A meeting brief ends with decisions and next actions stored where the team can find them.
This article is part of the How to Leverage Your Life with Super Intelligence series. It builds directly on reusable SI prompts, better context and the personal SI operating system.
The central principle is simple: a prompt produces an output; a workflow produces a completed state.
Why Random Questions Feel Useful but Rarely Compound
Random questions are excellent for exploration.
You can ask for an explanation, an idea, a rewrite or a comparison whenever a need appears.
The weakness is not the conversation itself. The weakness is that the useful method often disappears when the conversation ends.
Next week you repeat the same setup.
You find the same source again.
You rewrite the same instruction.
You make the same correction.
You forget which version was final.
The system helped, but the improvement did not survive.
A workflow captures the reusable structure around the interaction.
The question becomes one component rather than the whole system.
Prompt, Workflow and Automation Are Not the Same Thing
Prompt
A prompt is an instruction or request to the model. It may include context and examples.
Workflow
A workflow is the sequence that connects a trigger to a completed result. It includes sources, prompt, checks, action, destination and recovery.
Automation
Automation performs one or more workflow steps with reduced manual intervention. A workflow can be entirely manual. In fact, important workflows should often be understood manually before they are automated.
Confusing these levels creates unnecessary complexity. A person may try to automate a task before knowing what a correct manual process looks like. The result is not leverage. It is faster uncertainty.
The Nine Parts of a Complete SI Workflow
1. Trigger
What event starts the workflow? A school notice arrives. A study session begins. Friday afternoon arrives. A meeting is scheduled. A project enters review.
2. Source
Where does the authoritative information live? The original notice, course material, current calendar, approved project brief or verified dataset.
3. Context
Which subset of the source and current state does SI need? Provide the smallest sufficient model rather than the entire archive.
4. Prompt
What reasoning or transformation should SI perform? This may be a reusable template.
5. Verification
How will the result be checked? Against the source, a calculator, an independent task, current official information or another qualified person.
6. Approval
Who is authorised to decide that the result can be used? Preparation and external action should not be silently merged.
7. Action
What changes in the real world? The calendar is updated, the student attempts the problem, the message is sent, the document is saved, the purchase is made or the question is asked.
8. Completion evidence
What proves the task finished? A sent message, a booking confirmation, a submitted assignment, a saved document, a verified checklist or an independent correct attempt.
9. Review and recovery
What happens when the task fails, the source conflicts or circumstances change? A workflow without recovery logic is only a happy-path script.
The Workflow Spine
The nine parts can be compressed into one operating spine:
Trigger → Source → Context → SI work → Verify → Approve → Act → Evidence → Update.
When something goes wrong, route backward to the first incorrect state.
If the result is factually wrong, inspect source and context.
If the result is correct but unusable, inspect output and destination.
If the plan was correct but not completed, inspect approval, action and ownership.
If the same error returns next week, inspect whether the workflow update survived.
Worked Workflow 1 — A Weekly Planning System
Trigger
Sunday evening review.
Sources
Current calendar, active project cards and a short list of new obligations.
Context
Three weekly outcomes, fixed commitments, large tasks with rough duration, protected time, travel and one known uncertainty.
SI task
Build a minimum viable week, identify conflicts and produce one fallback rule. Do not fill every available hour.
Verification
Check fixed commitments against the actual calendar. Confirm that project deadlines match their current sources.
Approval
The person reviews the schedule. Shared responsibilities are discussed with the people affected.
Action
Approved time blocks are entered into the actual calendar or task system.
Completion evidence
The schedule is visible in the destination system, not merely described in chat.
Recovery
If a meeting moves, protect fixed deadlines and the stated personal boundary. Move the lowest-priority flexible task first.
The reusable prompt is only the middle of this workflow. The calendar check and the recovery rule are equally important.
Worked Workflow 2 — A Learning Session
Trigger
Planned study block begins.
Source
Current course material and the learner’s latest attempt.
Context
Learning goal, current capability, recent error, allowed help and time available.
SI task
Diagnose the first unstable reasoning step. Ask one question. Give the smallest useful hint only if needed. Finish with a changed problem.
Verification
Check technical content against the course source. Use a changed question as the learning test.
Action
The learner solves, explains and retries.
Completion evidence
An independent attempt shows whether the idea transferred.
Update
The learning record changes from a broad label such as “weak in algebra” to a specific current state such as “selects the method independently but still makes sign errors”.
The workflow protects learning from becoming answer consumption.
Worked Workflow 3 — Household Notice to Preparation
Trigger
A notice requiring household action arrives.
Source
The original notice.
Context
Only the relevant text, with unnecessary personal information removed.
SI task
Extract confirmed dates, required items, optional items, actions and unanswered questions. Preserve source wording for important requirements.
Verification
Compare every extracted date and requirement with the notice.
Approval
Family members confirm responsibilities rather than allowing the assistant to assign them by inference.
Action
Tasks and dates move into the shared preparation list or calendar.
Completion evidence
Materials are ready, confirmations are received and the actual obligation is completed.
Recovery
If the notice is ambiguous, ask the issuing organisation instead of asking SI to guess.
Worked Workflow 4 — Writing and Sending an Important Email
Trigger
A message needs a clear professional response.
Source
The original message and your own confirmed facts.
Context
Purpose, audience, factual commitments, tone and any wording you must not change.
SI task
Draft or revise the message while preserving facts and commitments.
Verification
Compare every date, promise, reason and request with your original intent.
Approval
You approve the exact final message.
Action
Send through the actual email system.
Completion evidence
Confirm the message appears in the sent folder or other destination.
Recovery
If delivery fails or status is uncertain, check before retrying so the recipient does not receive duplicates.
Worked Workflow 5 — Research to Decision Brief
Research workflows fail when source discovery, interpretation and decision are compressed into one opaque answer.
Trigger: a decision needs current evidence.
Source: authoritative current sources plus relevant background research.
Context: question, population, time period, decision criteria and currentness requirement.
SI task: map claims, supporting evidence, disagreement and unknowns.
Verification: open the key sources and confirm support.
Approval: human decides which evidence is sufficient for the intended consequence.
Action: create the decision brief.
Evidence: every consequential claim remains traceable.
Recovery: when sources conflict, preserve the conflict and investigate rather than averaging.
NIST’s AI Risk Management Framework treats evaluation and risk management as part of AI use. The same principle is valuable at personal scale: design verification and recovery into the workflow rather than treating them as optional aftercare.
Worked Workflow 6 — Major Purchase Research
Trigger: a substantial purchase is being considered.
Source: user’s requirements plus current product information.
Context: problem to solve, budget boundary, essential requirements, preferences, disqualifiers and existing alternatives.
SI task: clarify requirements before naming products; then compare current options.
Verification: current prices, specifications, warranty and availability checked with current sources.
Approval: user decides whether any purchase is justified.
Action: purchase happens outside the research conversation.
Completion evidence: order confirmation and retained record.
Recovery: if requirements change, redo the comparison rather than assuming the earlier ranking still applies.
The workflow explicitly allows “do not buy” as a successful outcome.
Manual First: Why You Should Run a Workflow Before Automating It
Manual operation reveals where judgement actually lives.
You discover which inputs are difficult to gather.
You discover which outputs require correction.
You discover which actions need approval.
You discover where another person’s consent matters.
You discover whether the workflow saves enough effort to justify integration.
Only after those boundaries are visible should automation remove selected steps.
Otherwise, you risk automating an incorrect assumption or scaling a verification problem.
Four Levels of Workflow Automation
Level 0 — Manual conversation
You bring the input, run the prompt, check the result and act manually.
Level 1 — Reusable prompt
The instruction is standardised. Inputs still arrive manually.
Level 2 — Connected retrieval
Authorised apps or files can provide selected context. Human review remains before action.
Level 3 — Assisted action
The system can perform supported external actions after explicit approval, such as creating or updating an item in a connected service.
Level 4 — Conditional recurring workflow
Selected checks or actions may run on a schedule or when a condition occurs, with defined notification and escalation rules.
Not every workflow should reach Level 4. The correct level depends on consequence, reversibility, verification and actual benefit.
Connected Apps: Access Must Stay Task-Specific
OpenAI’s current connected apps documentation explains that supported apps may let ChatGPT find information, summarise documents or perform supported actions, depending on app capabilities, plan, region, workspace and permissions.
A connection should have a defined workflow purpose.
“Use this calendar to check fixed commitments for my weekly plan” is a bounded purpose.
“Connect everything so the assistant knows my life” is not.
Review what the integration can read and what it can change.
Design the workflow so a connection failure has a manual fallback.
Approval Boundaries: Prepare, Propose, Act
Use three states.
Prepare: SI may collect, organise or draft.
Propose: SI may suggest a specific change or action.
Act: the external system changes.
A workflow should state where human approval sits between these states.
For low-risk reversible actions, approval may be light.
For consequential, public or irreversible actions, approval should be deliberate and specific.
“Publish this exact checked page” is a narrower approval than “manage my website”.
“Send this exact message to this confirmed recipient” is narrower than “handle my email”.
Completion Evidence: The Most Neglected Workflow Step
People often mark a task complete when the assistant describes the action.
Do not.
A reminder is complete when the reminder exists.
A message is complete when it was sent and the status is confirmed.
A booking is complete when a valid confirmation exists.
A study objective is complete when the learner demonstrates the capability.
A document update is complete when the correct current version is saved in the right place.
Completion evidence anchors the workflow in reality.
The Recovery Tree
Source missing
Stop. Retrieve or verify the source. Do not complete the workflow by guessing.
Source conflict
Preserve both claims with dates and authority. Resolve before consequential action.
SI output fails verification
Repair the prompt or context. Do not patch the final output repeatedly without finding the cause.
Approval withheld
Return the item to proposal state. Do not treat non-response as approval.
Action fails
Check the destination before retrying. Avoid duplicate sends, bookings or updates.
Completion uncertain
Use destination evidence or direct confirmation. Do not infer completion from the plan.
Repeated failure
Pause automation. Run the process manually and diagnose the first unstable step.
Workflow Drift
A workflow can remain technically functional while no longer serving the original goal.
The weekly plan gradually fills protected evenings.
The learning workflow begins giving longer explanations and fewer attempts.
The household workflow collects more private information than it needs.
The research workflow relies on summaries instead of opening sources.
The writing workflow starts adding commitments because a style preference rewards decisiveness.
Use periodic review to compare the current workflow with its purpose, constraints and stop conditions.
Version the Workflow, Not Just the Prompt
A workflow version changes when any operational component changes: source, approval, destination, completion evidence, recovery or prompt.
v1.0: notice text manually pasted; family checklist manually updated.
v1.1: source retrieval connected; verification still manual.
v1.2: shared calendar proposal generated, but event creation remains approval-only.
Version notes help you understand what changed when performance changes.
How to Measure a Workflow
Total effort
Count input preparation, SI interaction, checking, correction, action and maintenance.
Quality
Choose a task-specific measure: omissions, independent learning, factual accuracy, clarity, successful completion or fewer reversals.
Verification burden
If every output needs expensive checking, the workflow may need narrower scope.
Failure recovery
Measure how quickly the process returns to a correct state after an error.
Human load
Ask whether work was genuinely reduced or merely shifted to another person.
Dependency
Can the task still be completed when the assistant or connection is unavailable?
No single metric fits every workflow. Keep measurement proportionate to the task.
The Workflow Break-Even Test
Suppose a weekly task normally takes 30 minutes.
The SI workflow requires 6 minutes of input preparation, 4 minutes of interaction, 8 minutes of verification and 4 minutes of transfer. Total: 22 minutes.
The apparent saving is 8 minutes per comparable use.
If setup required 80 minutes, roughly ten similar uses would recover that setup time under these simplified assumptions.
But if maintenance adds 10 minutes every week, the workflow no longer saves time.
This calculation is illustrative. The important habit is to measure the whole process rather than the generation step alone.
A 30-Day Workflow Build
Days 1–5: observe one repeated task. Record source, action and failure points without changing the process.
Days 6–10: run the task manually with SI assistance. Build the first reusable prompt.
Days 11–15: add verification, completion evidence and stop conditions.
Days 16–20: test missing information, conflicting information and one changed example.
Days 21–24: measure total effort and maintenance. Remove unnecessary fields or steps.
Days 25–27: add one connection only if retrieval or transfer is a demonstrated bottleneck.
Days 28–29: test manual fallback and action failure.
Day 30: decide: keep, simplify, automate one step further or stop.
The 30-day path makes automation the result of understanding rather than the starting assumption.
The Workflow Registry
Once you have several working workflows, maintain a small registry.
Name: PLAN.WEEKLY
Trigger: Sunday review
Source: calendar + active projects
Prompt: PLAN.WEEKLY.v1.2
Human approval: before calendar changes
Destination: calendar/task list
Completion evidence: approved schedule visible
Fallback: manual calendar review
Owner: user
Status: active.
A registry prevents forgotten automations and duplicated workflows. It also makes it easier to stop systems that no longer justify their maintenance.
A Complete SI Workflow Specimen: From Trigger to Verified Completion
The most important workflow lesson is that every stage should have a visible state. The following fictional specimen shows a complete workflow for preparing and sending a weekly project update. It is intentionally ordinary so the operating logic remains easy to inspect.
State 0 — Trigger
Friday at 3pm, the weekly review period begins. The trigger is time-based and predictable. Nothing has been generated yet.
State 1 — Source collection
The workflow gathers the current project card, this week’s completed tasks, unresolved blockers and the latest approved milestone plan. Old drafts remain outside the active source set.
The source rule says: the current project card owns status, the milestone plan owns agreed dates and completed-task evidence comes from the actual work system. If a note conflicts with one of these sources, the conflict must be flagged rather than silently merged.
State 2 — Context packet
The working packet says: goal is a concise status update; audience is the project sponsor; decision needed this week is whether to move one milestone; known uncertainty is the vendor response time; no confidential customer details should be included; output must fit 250 words.
State 3 — SI preparation
The reusable prompt asks SI to produce four sections: completed, current risk, decision needed and next week. It must distinguish confirmed facts from estimates and preserve the vendor uncertainty.
State 4 — Verification
The human checks every completion claim against the actual task system, every date against the current milestone plan and every risk against the project card. One generated sentence says the vendor delay is “expected to be resolved Monday”. The source only says “vendor response pending”. That sentence is removed.
State 5 — Approval
The project owner approves the exact message. The workflow does not treat a generated draft as ready to send.
State 6 — Action
The message is sent through the real communication system.
State 7 — Completion evidence
The sent status is confirmed. A copy or reference is stored according to the project’s normal record practice.
State 8 — Update
The only workflow lesson retained is: “Do not infer a resolution date from the phrase response pending.” The reusable prompt gains a rule that unresolved third-party timing remains unresolved unless the source contains a date.
This specimen shows why the final prompt is only one step. A perfect draft without source discipline, verification, approval and completion evidence would still be an incomplete workflow.
Workflow State Labels: Proposed Is Not Approved, Approved Is Not Completed
Use explicit state labels when a workflow involves external action. They prevent conversational language from making a task look further along than it really is.
Captured: the trigger or input has arrived.
Prepared: SI produced a draft, summary or proposal.
Verified: required checks passed.
Approved: the authorised person accepted the exact action.
Executed: the real system changed or the action was performed.
Confirmed: completion evidence exists.
Closed: records were updated and no follow-up remains.
A task should move forward only when the condition for the next state is actually met. “Prepared” does not become “executed” because the assistant uses confident language. “Approved” does not become “confirmed” because a send command was attempted. State discipline reduces false completion.
The Workflow Ownership Map
Every recurring workflow needs an owner for each consequential stage.
For a family notice, the parent may own verification, another family member may own transport and the student may own packing materials.
For a learning workflow, the learner owns the attempt, the teacher or source may own curricular correctness and the assistant only owns the temporary instructional step it was asked to perform.
For a workplace workflow, the organisation may own data policy, the manager may own approval and a connected app may only provide retrieval or action capability.
Write ownership explicitly when ambiguity could create risk. SI should never infer authority merely from access.
Failure Injection: Test the Workflow Before It Matters
A workflow that succeeds only when every input is perfect is not robust. Before relying on it, deliberately introduce several failures in a safe test environment.
Failure test 1 — Missing source
Remove the original document. The workflow should stop at source collection or produce an explicit “source unavailable” state. It should not proceed using an old summary as though it were current.
Failure test 2 — Conflicting dates
Provide two different deadlines from two sources. The workflow should preserve the conflict, identify source dates and require resolution before action.
Failure test 3 — Unsupported completion claim
Add a note saying “probably finished”. The workflow should not mark the task completed without the defined completion evidence.
Failure test 4 — Wrong destination
Change the target calendar, folder or recipient in a test case. Confirm that the workflow identifies the destination explicitly before acting.
Failure test 5 — Tool unavailable
Assume the connected app is temporarily unavailable. Can the task still be completed manually from the source records?
Failure test 6 — Human rejects the proposal
The workflow should return to proposal state, not attempt a different action automatically unless that behaviour was explicitly designed and authorised.
Failure test 7 — Duplicate action risk
Simulate an uncertain send or booking result. The recovery rule should check the destination before retrying.
These tests reveal whether the workflow has recovery behaviour or merely a happy path.
The Workflow Decision Matrix: Should This Be Automated?
Before increasing automation, inspect five dimensions.
Frequency
High-frequency tasks offer more opportunity to recover setup cost. But frequency alone does not justify automation.
Stability
Stable inputs and rules are easier to automate than tasks whose criteria change every time.
Verification
Cheap, reliable verification makes automation safer. If errors are difficult to detect, keep more human review.
Consequence and reversibility
A reversible draft can tolerate more experimentation than an irreversible financial transfer, public publication or sensitive communication.
Human judgement
Some tasks depend on values, consent, empathy or professional responsibility that should remain with people even when preparation can be automated.
A useful rule is: automate the stable mechanical layer first; keep ambiguous judgement and consequential approval visible.
Worked Automation Decision: Calendar Planning
A weekly planning workflow may contain four possible automated steps: retrieve calendar events, produce a proposed plan, create draft events and move existing events.
Retrieving events is relatively mechanical when access is authorised. Producing a proposed plan remains advisory. Creating draft events may be appropriate after the user reviews the plan. Moving existing commitments automatically is more consequential because another person’s meeting or a fixed appointment may be affected.
The automation boundary can therefore stop before event movement. The workflow still gains leverage from connected retrieval and prepared proposals without granting broader action authority than necessary.
This example shows that “automated workflow” is not a binary category. Different steps can sit at different automation levels.
Workflow Templates for Everyday Life
Morning preparation workflow
Trigger: start of day.
Source: current calendar and today’s task list.
SI role: identify fixed commitments, one priority and unresolved preparation.
Check: calendar times.
Action: user chooses the first task.
Evidence: preparation complete before leaving.
Recovery: if a fixed commitment changes, rebuild only the affected part of the day.
Reading-to-notes workflow
Trigger: finish a chapter or article.
Source: original text.
SI role: extract claims, definitions and questions while preserving source references.
Check: compare summary with text.
Action: user writes one explanation in their own words.
Evidence: retrievable note linked to source.
Recovery: when the summary loses a caveat, correct the prompt and regenerate only the affected section.
Purchase research workflow
Trigger: purchase exceeds a chosen importance or cost threshold.
Source: user’s requirements and current product sources.
SI role: clarify requirements and compare options.
Check: current specifications, price and warranty.
Action: user decides.
Evidence: retained order or no-purchase decision.
Recovery: if requirements change, rerun the comparison rather than preserving the old ranking.
Travel preparation workflow
Trigger: trip confirmed.
Source: confirmed bookings and current official requirements.
SI role: organise itinerary, questions and contingency options.
Check: booking details and official current sources.
Action: user approves changes or bookings.
Evidence: confirmed itinerary and documents.
Recovery: disruption routes to current transport/provider information rather than the stale original plan.
Learning review workflow
Trigger: end of study week.
Source: attempt record and course material.
SI role: classify errors and propose one repair target.
Check: independent retest.
Action: learner selects next practice.
Evidence: changed problem performance.
Recovery: if generated practice is unreliable, return to teacher-approved or source material.
Document revision workflow
Trigger: draft ready for review.
Source: current draft plus requirements.
SI role: check structure, clarity, missing support and consistency.
Check: human reviews every factual change.
Action: accepted edits enter current version.
Evidence: saved version with clear date/version.
Recovery: rollback when a revision changes meaning incorrectly.
The Handover Problem: Where Workflows Commonly Break
Many workflow failures happen between good steps rather than inside them.
Source to context
The correct document exists, but the wrong version is copied into the prompt.
Context to SI
The right facts are present, but the goal or constraints are unclear.
SI to verification
The output looks polished, so nobody checks it.
Verification to approval
A reviewer assumes someone else has approved.
Approval to action
The approved task is never actually executed.
Action to completion evidence
The system reports uncertainty, but the task is marked finished anyway.
Completion to update
The real-world state changed, but the project record remains stale, so the next run repeats old work.
Design each handover explicitly. The strongest prompt cannot compensate for a missing handover.
The Manual Fallback Card
Every important connected workflow should have a small fallback card.
If SI is unavailable: where are the source records?
If the connected app fails: how do I retrieve or update manually?
If the workflow output is unreliable: what is the direct manual process?
If an action status is uncertain: where do I check?
If I stop using this provider: which prompts and records must I retain?
The fallback card protects continuity and exposes hidden dependence. If nobody can explain how to perform the underlying task manually, the workflow may be too opaque for its importance.
Workflow Security and Privacy Hygiene
A workflow repeats. That means a bad data habit repeats too.
Keep credentials, passwords and authentication codes outside prompts.
Use the minimum information needed for the task.
Separate personal, shared and restricted records.
Check organisational rules before moving workplace information into a personal account.
Review connected-app permissions periodically and remove connections that no longer serve an active workflow.
Do not use an automation to bypass a human consent step merely because the technical path exists.
A secure workflow is not one that never uses connected information. It is one whose access has a defined purpose and boundary.
Workflow Maintenance: The Smallest Change That Fixes the Real Failure
Do not rebuild the whole workflow after every mistake.
If the source was stale, repair source selection.
If the prompt invented missing dates, add a stop condition.
If verification took too long, narrow the output.
If users forgot to act, strengthen the handover to the destination.
If duplicate actions occurred, improve completion checking before retries.
If the workflow collects too much context, reduce the input fields.
Repair at the earliest failing stage. This keeps workflows understandable and prevents layers of defensive complexity from accumulating everywhere.
Workflow Review Questions
At each review, ask: Does this workflow still serve the same goal? Is the source still authoritative? Is the input still minimal? Is the prompt still current? Is verification happening? Is approval explicit where needed? Is completion evidence real? Is maintenance cost growing? Is another person carrying hidden work? Is the manual fallback still possible?
Then choose one action: keep, simplify, reduce automation, increase automation cautiously, or retire.
Retiring a workflow is a successful outcome when the underlying task disappeared or a simpler method became better.
A Cross-Domain Transfer Test
A robust workflow architecture should transfer even when the content changes.
Take the spine—Trigger → Source → Context → SI work → Verify → Approve → Act → Evidence → Update—and apply it to a different domain.
For study, the source is course material and evidence is independent performance.
For household administration, the source is the official notice and evidence is completed preparation.
For writing, the source is the original draft and evidence is the saved checked version.
For a booking, the source is current provider information and evidence is the booking confirmation.
The specific checks change, but the operating logic remains. That transfer is evidence that you have learned a workflow architecture rather than memorised one example.
The Workflow Maturity Test
A mature SI workflow has several characteristics.
The trigger is clear.
The source is traceable.
The context is minimal and current.
The prompt has a stable job.
Verification matches the output.
Approval matches consequence.
External actions are observable.
Completion evidence is defined.
Failure routes are known.
The workflow can be explained by a human.
The system can be simplified or stopped without losing the underlying capability.
That last point matters most: dependable infrastructure increases human capability while remaining understandable enough to leave.
The Workflow Anti-Patterns That Make SI Look More Capable Than the System Really Is
Some workflow designs create the appearance of intelligence while hiding fragile handoffs. These anti-patterns are worth learning because they recur across life, school and work.
The infinite chat
The conversation contains the plan, the current state, the decisions, the next action and the history. Nothing is written back to a stable record. The workflow works until the thread becomes difficult to search or a new conversation begins without the same context.
Repair: save the checked current state and next action outside the conversation. Keep the chat as working history, not the only operational record.
The invisible source
The assistant produces a summary, but nobody can tell which document or version supplied a key claim.
Repair: name the authoritative source at the start and keep traceable references for consequential facts.
The approval blur
A user says “looks good”, and the system treats that as permission to perform several external actions that were not individually clear.
Repair: approvals should identify the object and action: send this message, update this event, publish this page. Preparation and permission remain separate states.
The completion illusion
The assistant says “done” after generating instructions or attempting an action, but the destination has not been checked.
Repair: define completion evidence before execution. The workflow does not close until the evidence exists.
The retry loop
An external action returns an uncertain status, so the workflow immediately repeats it. This can create duplicate messages, bookings or records.
Repair: uncertain action state routes first to destination checking, not automatic retry.
The automation ladder jump
A person connects accounts and grants action capability before running the process manually even once.
Repair: prove the task, check and recovery manually, then automate the stable mechanical step that genuinely creates leverage.
The dashboard substitute
The workflow produces a beautiful dashboard but nobody knows what action changes because of it.
Repair: every displayed item should support a decision, action, check or alert. Remove metrics that create attention without operational value.
Human-in-the-Loop Is Not One Thing
People often say a workflow has “human oversight” without specifying what the person is actually doing. Human involvement can occur at several different points, and the placement matters.
Human sets the goal
The person defines what outcome matters and which values constrain the task.
Human provides or authorises the source
The person determines which information may enter the workflow and which record should be trusted.
Human verifies
The person checks the output against the source, calculation, independent task or other appropriate evidence.
Human approves action
The person accepts the exact external change before it occurs when consequence requires this boundary.
Human handles exceptions
The routine may run smoothly most of the time, while ambiguous, conflicting or high-consequence cases route to a person.
Human reviews the system
The person examines whether the workflow still serves its intended goal and whether automation has drifted.
A good design names which of these roles are necessary. “There is a human in the loop” is too vague to be an operational control.
Event Logging: Keep Only the History That Helps You Recover
A recurring workflow benefits from a lightweight event log. The log does not need every model message. It needs the state transitions that matter.
Example:
18:00 — notice received.
18:04 — source verified.
18:07 — SI extraction prepared.
18:12 — human verification complete; optional item corrected.
18:15 — family list updated.
18:17 — transport owner unresolved; question sent.
19:05 — owner confirmed.
19:07 — workflow closed.
This record makes failure diagnosis easier. If the wrong date appears later, you can see whether it entered during source capture, SI extraction, human verification or the final update.
Do not log sensitive detail merely because logging is possible. Record enough to understand state and recovery.
A Workflow Should Reduce Re-Explanation
One practical test of workflow maturity is how much setup must be repeated.
In a random-question pattern, the user explains the same background every time. In a better workflow, stable context lives in the current brief, the reusable prompt already knows the task class and only the changing input must be supplied.
But beware of the opposite extreme. If the workflow requires maintaining a complex database just to avoid two minutes of re-explanation, the infrastructure has become heavier than the problem.
The right amount of persistence is the amount that materially improves the next run.
The Workflow Closure Test
Before calling a workflow finished, answer six questions.
1. Did the real task change state?
2. Is the completion visible in the destination system?
3. Were consequential facts verified?
4. Were approvals obtained where needed?
5. Was the current record updated so the next run starts correctly?
6. Is there any unresolved exception that still needs an owner?
If the answer to the first or second question is no, the workflow probably produced preparation rather than completion. That may still be useful, but label it correctly.
The distinction prevents a common SI failure: confusing fluent progress reports with progress in the world.
The One-Workflow Challenge
Choose one repeated task that occurs at least several times each month and whose result you can verify. For the first week, do not automate it. Write the trigger, source, context, prompt, check, action and completion evidence.
Run it manually twice. Record the first point of friction on each attempt.
On the third attempt, repair only that point. If source retrieval is slow, improve retrieval. If the prompt invents details, improve the stop rule. If checking is expensive, narrow the output. If action is forgotten, improve the handover.
After several runs, measure the full process. Only then decide whether a connected app, scheduled trigger or supported action would remove a proven burden.
The objective is not to build an impressive automation. It is to build one process that you trust because you understand how it reaches reality.
A Full Workflow Walkthrough: Weekly Planning from Trigger to Closed Loop
A workflow becomes real when every handover is visible. The following specimen uses weekly planning because the task is familiar, but the structure applies to many other domains.
Trigger
Sunday evening arrives, or the previous week’s plan is formally closed. The trigger is not “I feel disorganised”. It is an observable event that tells the workflow when to begin.
Source collection
The workflow receives the current calendar, active project cards and any confirmed deadlines that are not yet on the calendar. It does not treat old chat summaries as the primary source when current records exist.
Context packet
Goals: finish the report draft, complete two study sessions, prepare for Friday appointment.
Fixed commitments: current calendar events.
Protected time: Sunday afternoon and two family blocks.
Uncertain item: Thursday client call may move.
Output: minimum viable week first; optional extension second.
Check: every fixed time against calendar.
SI preparation
SI identifies conflicts before scheduling. It separates must-do, flexible and optional work. It preserves uncertainty around Thursday. It does not schedule every open hour. It proposes buffer where a late task would otherwise threaten a fixed commitment.
Verification
The user compares fixed times with the calendar, checks the real durations of the largest tasks and confirms that protected time remains protected. Any invented assumption is removed.
Approval
The user accepts the weekly plan. This is a decision state, not yet a calendar state. If connected tools are available, the user can explicitly approve a narrow set of changes rather than granting broad authority to reorganise everything.
Action
Approved study and work blocks are entered into the actual planning system. Optional ideas remain outside the calendar unless chosen.
Completion evidence
The destination system shows the blocks. The workflow records that the plan was entered successfully. A statement such as “I have scheduled your week” is not enough if the destination cannot be inspected.
Review
At the end of the week, compare plan with reality. Which block failed because duration was wrong? Which task moved because a dependency changed? Which buffer prevented a collision? Retain one useful correction for next week.
The entire loop can be manual. Automation is optional. What makes it a workflow is the stable sequence from trigger to review.
State Machines in Plain English: Know What Stage the Work Is In
Many workflow errors occur because different states are treated as if they were the same. A state machine is simply a clear set of labels describing where an item currently sits.
New: the task entered the workflow.
Prepared: SI produced a candidate output.
Checked: the output passed verification.
Approved: a human authorised the next action.
Executed: the action was attempted.
Confirmed: completion evidence exists.
Failed: the action or verification failed.
Needs clarification: progress is blocked by missing or conflicting information.
Closed: the task is complete and the record updated.
These states prevent a common category error: a prepared draft being described as completed work. They also make recovery easier because the next step depends on the actual state rather than on conversational wording.
Retry Safety: Never Repeat an External Action Blindly
Suppose a workflow sends a message, creates an event or submits a form, but the system reports an error. The worst recovery is to repeat the action immediately without checking whether the first attempt actually happened.
Use a retry protocol: 1) check the destination for completion evidence; 2) if completion is confirmed, do not retry; 3) if completion clearly failed, repair the cause and retry once; 4) if completion is uncertain, escalate for human inspection rather than creating duplicate risk; 5) record the result so the workflow does not forget what happened.
Software engineers often call the desired property idempotency: repeating the same intended operation should not create multiple unintended effects. In everyday language, the rule is simpler: before doing it again, check whether it already happened.
Observability: A Workflow Must Tell You What It Is Doing
A workflow is hard to trust when the user cannot see its inputs, state, output and completion evidence. Observability means making those elements inspectable.
For an ordinary personal workflow, you usually need only a compact record: trigger time, source used, current state, output location, verification result, approval state and completion evidence.
Do not log everything. A giant transcript creates noise and may retain information unnecessarily. Log the events that help you answer: What happened? Which source was used? What failed? Did the external action complete? Which version of the workflow ran?
Example event record
“30 Sep 18:00 — School notice workflow v1.2 triggered. Source: notice dated 30 Sep. Preparation created. Verification: one ambiguous transport sentence flagged. State: needs clarification. No family task assigned. Workflow stopped.”
That record is enough to recover later without preserving a large hidden process.
A Complete Workflow Walkthrough: Learning Without Answer Dependency
A learning workflow should be designed around evidence of capability, not around the amount of content SI can generate.
Trigger: scheduled study session begins.
Source: syllabus section, latest marked attempt and current error log.
Context: learner can calculate percentages but struggles to identify the method in word problems.
Prompt: ask one diagnostic question; wait for attempt; give smallest useful hint; end with changed problem.
Verification: compare technical explanation with source material.
Action: learner attempts, explains and retests.
Completion evidence: changed problem solved independently with correct reasoning.
Update: record current error pattern and next practice need.
If the learner cannot solve the changed problem, the state is not “lesson completed”. The workflow returns to diagnosis. This is why state labels matter: activity completion and learning completion are different.
A Complete Workflow Walkthrough: Research to Decision Brief
Research workflows fail when source collection, synthesis and decision-making blur together.
Trigger: a decision requires external evidence.
Research question: written with time, geography and population boundaries.
Source plan: define which claims need primary sources, current sources or specialist evidence.
Collection: retrieve only relevant authorised material.
SI synthesis: separate established facts, disputed claims, missing evidence and implications.
Verification: open the important sources and confirm they support the summary.
Decision brief: state evidence, uncertainty, options and unanswered questions.
Human decision: choose based on stated criteria.
Archive: retain source map and final decision, not every search trace.
The workflow should stop when the next question requires new evidence. More synthesis cannot compensate for a missing source.
Workflow Handoffs: The Exact Places Errors Enter
Trigger → source
The wrong record is collected or the workflow starts when it should not. Repair with a clear trigger and source rule.
Source → context
Important caveats disappear during summarisation. Repair by preserving provenance and uncertainty.
Context → prompt
The instruction ignores a key constraint. Repair by separating stable prompt rules from task-local context.
Prompt → verification
A fluent result is accepted without a check. Repair by defining the acceptance test before generation.
Verification → approval
A checked result is assumed to be authorised. Repair with an explicit approval state.
Approval → action
The wrong destination, recipient or object is selected. Repair by confirming the exact target.
Action → completion evidence
The system narrates success without proof. Repair by inspecting the destination or receipt.
Completion → update
The real-world task is complete but the project record remains open, causing duplicate work later. Repair by closing the loop and updating the source-of-truth record.
The Automation Gate: Five Questions Before Giving SI More Authority
Automation should follow evidence, not enthusiasm. Before moving a workflow from manual preparation to connected action, answer five questions.
1. Is the task stable? If the steps change every time, automation will be brittle.
2. Is the output easy to verify? If errors are hard to detect, keep human review close.
3. Is the action reversible? Irreversible actions deserve narrower authority.
4. Is the destination unambiguous? The system should know exactly which record, recipient or account it is acting on.
5. Is there a recovery path? If something fails, the workflow must know how to stop without compounding the error.
If several answers are no, keep the workflow at preparation or proposal level. A manual workflow can still create substantial leverage.
Workflow Metrics That Matter
Cycle time
How long does the complete process take from trigger to confirmed completion? Include verification and recovery.
First-pass quality
How often does the prepared output pass verification without repair?
Exception rate
How often does the workflow hit missing, conflicting or ambiguous information?
Recovery cost
When something fails, how much effort is required to restore a safe state?
Human attention
How much review, correction and approval does the workflow require? A process that saves computer time but consumes more human attention may not be an improvement.
Outcome quality
Did the actual task improve? Fewer omissions, stronger learning, clearer decisions or more reliable preparation matter more than output volume.
A Workflow Build Sheet You Can Copy
Name: [workflow name]
Owner task: [one primary job]
Trigger: [event or condition]
Sources: [authoritative inputs]
Context: [current state and constraints]
Prompt or reasoning step: [what SI prepares]
Verification: [how output is checked]
Approval: [who authorises action]
Action: [exact external or human step]
Destination: [where result belongs]
Completion evidence: [what proves done]
Exceptions: [missing source / conflict / failure routes]
Fallback: [manual route]
Version: [current workflow version]
Review trigger: [when to reassess].
If you cannot fill one of these fields, the workflow may not yet be ready for automation. Build the missing part manually first.
The Workflow Floor: What Must Be True Before You Call It Reliable
A reliable personal SI workflow has an explicit trigger, identifiable source, current context, bounded SI task, acceptance test, approval rule, real action, observable completion evidence, exception path, manual fallback and review. Remove any one of those and the workflow may still work in easy cases, but the failure mode becomes harder to see.
The purpose is not bureaucratic perfection. It is to make the system understandable enough that a person can intervene when reality changes.
Random questions give isolated help. Reusable prompts give repeatable instructions. Workflows connect those instructions to sources, checks, action and learning. That is where Super Intelligence begins to compound into durable everyday capability.
The Workflow Exit Test
Before calling a workflow mature, ask whether you can explain its trigger, source, current state, verification, approval boundary, action, completion evidence and recovery path without opening the chat that created it.
If you cannot, the workflow is still partly hidden inside conversation. Move the durable parts into a workflow card, preserve a manual fallback and make the completion state observable. A reliable workflow should become easier to inspect as it matures, not more mysterious.
Frequently Asked Questions
What is an SI workflow?
A repeatable sequence that connects a trigger and source to SI processing, verification, approval, real-world action, completion evidence and review.
Do I need automation to have a workflow?
No. A workflow can be fully manual. Manual-first operation is often the best way to understand the task before automating selected steps.
What is the difference between a prompt and a workflow?
The prompt tells SI what to do with the current input. The workflow defines the surrounding process that gets input, checks output, acts on it and records completion.
How many workflows should I build at once?
Start with one repeated, bounded task. Multiple simultaneous changes make it harder to see which design choice helped or failed.
Which workflow should I automate first?
Choose a low-risk repeated process with stable inputs, clear verification and a reversible action. Do not choose only by the number of minutes it appears to consume.
How do I know whether a workflow is complete?
Define completion evidence in advance. The evidence should live in the real destination: submitted work, saved record, sent message, confirmed event or demonstrated skill.
What if the assistant gives a wrong answer?
Stop before action, identify whether source, context or prompt caused the failure, repair the earliest incorrect step and retest.
Should workflows use connected apps?
Only when a specific connection solves a demonstrated retrieval or action bottleneck and the access is appropriate. Keep a manual fallback.
Can a workflow make decisions automatically?
Some low-risk systems may support bounded automated actions, but consequential decisions should retain appropriate human judgement and approval. Technical capability is not sufficient reason to delegate.
How often should workflows be reviewed?
Review after meaningful failures and periodically according to how often the workflow changes. Stable simple workflows need less attention than connected or high-consequence ones.
What comes next in the series?
The next foundation article examines what should and should not be delegated to Super Intelligence, including authority, reversibility, risk and human responsibility.
Continue the Super Intelligence Life Series
How to Leverage Your Life with Super Intelligence — 100-Article Hub
How to Turn Super Intelligence from a Chatbot into a Thinking Partner
How to Give Super Intelligence Better Context About What You Are Trying to Achieve
How to Create Reusable Super Intelligence Prompts for Your Life
These guides form one progression: improve the quality of thought, give SI the right current context, preserve working instructions as reusable prompts, and then connect proven prompts to sources, verification, action and review.
Helpful Reading
- How to Leverage Your Life with Super Intelligence
- How to Create Reusable Super Intelligence Prompts for Your Life
- How to Give Super Intelligence Better Context About What You Are Trying to Achieve
- How to Build Your Personal Super Intelligence Operating System
- How to Identify Where Super Intelligence Can Create the Most Leverage in Your Life
- OpenAI — Connected Apps in ChatGPT
- OpenAI — Prompt Engineering Best Practices for ChatGPT
- NIST — AI Risk Management Framework
From Conversation to Capability
A random question can be useful.
A reusable prompt can make useful help easier to repeat.
A workflow goes further.
It knows when the task begins.
It knows where facts come from.
It knows how SI should contribute.
It knows how the result is checked.
It knows who approves action.
It knows what completion looks like.
And it knows what to do when reality stops following the happy path.
That is when Super Intelligence stops being a sequence of interesting conversations and starts becoming dependable personal infrastructure.
