Build a Self Verifying Claude Workflow

Review the Claude instructions, memory, skills, and workflow rules I provide below.

Your job is to perform an ablation review. Produce the smallest useful instruction set for how I currently use Claude.

Before rewriting anything, ask me concise questions about:
1. The tasks I perform most often.
2. The business context Claude genuinely needs.
3. The goals that apply across those tasks.
4. The mistakes Claude repeatedly makes.
5. The output requirements that are mandatory.
6. The decisions that require human approval.

Then classify every existing instruction as one of the following:

KEEP: Essential context, constraint, or completion requirement.
REMOVE: Redundant, outdated, overly prescriptive, or unnecessary behavioral guidance.
TEST: Possibly useful, but it should be removed temporarily and restored only if a failure appears.
TASK SPECIFIC: Useful only for a particular workflow and should not remain in global instructions.

Return:
1. A table containing each current instruction, its classification, and your reason.
2. A minimal global instruction file containing only essential context and universal constraints.
3. A separate list of task specific instructions and the workflows where they belong.
4. A test plan for comparing the current setup with the lean version.
5. A list of rules that should be restored only if a specific failure occurs.

Do not preserve an instruction merely because it already exists. Prefer subtraction unless the instruction prevents an observed failure or communicates essential context.

Here are my current instructions:

[PASTE CURRENT INSTRUCTIONS, MEMORY, SKILLS, OR CLAUDE.md CONTENT]

CONTEXT
[Relevant information about the business, files, tools, or audience]

GOAL
[The final result to produce]

REQUIRED OUTPUT
[Files, fields, sections, artifacts, or actions required]

SUCCESS CRITERIA
[Objective tests the result must pass]

BOUNDARIES
[Actions that are prohibited or require approval]

DEFINITION OF DONE
[Conditions that must all be true before stopping]

Open each channel, inspect the videos one at a time, copy each title, calculate views per day, place the results in a table, calculate the median, and then look for videos above the median.

CONTEXT
We monitor competitor videos to identify unusual performance and useful content patterns. The tracker is stored in niche-tracker.md.

GOAL
Update the tracker with new competitor uploads and identify meaningful performance outliers.

REQUIRED OUTPUT
For every new upload, record:
1. Channel.
2. Video title.
3. Publication date.
4. Current views.
5. Views per day.
6. Topic.
7. Performance relative to the channel median.

Flag any video performing at three times or more than the channel median. For each flagged video, add one concise line explaining why its title or thumbnail may be effective.

SUCCESS CRITERIA
1. Existing tracker entries are not duplicated.
2. Every required field is populated or explicitly marked unavailable.
3. Calculations are internally consistent.
4. Outliers use the same threshold across the tracker.
5. The tracker remains valid Markdown.

BOUNDARIES
Do not delete historical entries.
Do not modify files outside this workflow.
Do not publish, message, purchase, install, or delete anything without approval.
If required source data is unavailable, report the gap rather than inventing it.

DEFINITION OF DONE
The tracker contains all discovered new uploads, calculations have been checked, qualifying outliers have been flagged, and any unresolved data gaps have been reported.

Help me convert the recurring process below into an operational workflow graph.

First, interview me until you understand:
1. What triggers the process.
2. What information enters the process.
3. The final result.
4. The decisions I make from experience or intuition.
5. The rules behind those decisions.
6. The tools, files, and systems involved.
7. Common exceptions and failure conditions.
8. Which actions require human judgment or approval.

Then produce two artifacts.

ARTIFACT 1: MARKDOWN WORKFLOW SPECIFICATION
For every stage, include:
1. Stage name.
2. Purpose.
3. Required inputs.
4. Actions or decisions.
5. Responsible actor.
6. Expected output.
7. Verification test.
8. Failure response.
9. Next stage.

ARTIFACT 2: VISUAL HTML WORKFLOW
Create one self contained HTML file that displays the process as a readable graph. Show:
1. Start and completion points.
2. Processing stages.
3. Decision branches.
4. Claude or agent actions.
5. Human approval gates.
6. Verification steps.
7. Failure and retry paths.

Use clear labels and keep the HTML easy to edit.

Do not assume that every stage should be automated. Explicitly identify where human taste, authorization, or judgment remains important.

Process to map:
[DESCRIBE THE CURRENT PROCESS]

Execute the workflow described below using a self verification loop.

GOAL
[INSERT FINAL OUTCOME]

WORKFLOW
[PASTE OR REFERENCE THE WORKFLOW SPECIFICATION]

VERIFICATION LOOP
1. Produce the required result.
2. Test it against every success criterion.
3. List any failed criteria internally and correct the work.
4. Repeat verification until all criteria pass or a stopping condition is reached.
5. Do not claim completion unless every criterion has passed.

STOPPING CONDITIONS
Stop and report the issue if:
1. Required information cannot be accessed.
2. Two correction attempts fail for the same reason.
3. Instructions conflict.
4. An action requires human approval.
5. Continuing could modify, delete, publish, purchase, install, or send something outside the approved scope.

FINAL RESPONSE
Return:
1. Completion status.
2. Output produced or files updated.
3. Verification checks performed.
4. Exceptions, missing data, or unresolved risks.
5. Any human decision required.

Definition of done:
[INSERT OBJECTIVE COMPLETION CONDITIONS]

/goal Scan the specified competitor channels for new uploads. Log the channel, title, publication date, views, views per day, and topic in niche-tracker.md. Flag any video performing at three times or more than the channel's median. Add one concise line explaining why each outlier's title or thumbnail may work. Do not duplicate existing entries or delete historical data. Done means the tracker is updated, calculations are verified, outliers are flagged, and unresolved data gaps are reported. Stop after eight runs.

/loop [INSERT DESIRED INTERVAL]

HUMAN APPROVAL GATES
You may research, analyze, draft, organize, calculate, and prepare recommendations autonomously.

Pause before any action that:
1. Sends or publishes information externally.
2. Deletes or irreversibly overwrites data.
3. Installs software or changes system configuration.
4. Creates a financial, legal, or contractual commitment.
5. Changes a production workflow.

At an approval gate, return:
1. The proposed action.
2. Why it is needed.
3. The expected effect.
4. The main risk.
5. The exact approval requested.

workflows/
  competitor-monitoring/
    workflow.md
    workflow.html
    goal.md
    test-cases.md
  lead-qualification/
    workflow.md
    workflow.html
    goal.md
    test-cases.md

Update the workflow files whenever the underlying business process changes. This keeps the automation aligned with the current process rather than preserving outdated instructions.

Lean context
    ↓
Clear goal
    ↓
Required output
    ↓
Workflow graph
    ↓
Success criteria
    ↓
Verification loop
    ↓
Human approval gates
    ↓
Stopping conditions
    ↓
Manual test
    ↓
Recurring execution