Draft

Not yet reviewed. Artifacts stay locked until you approve them.

Step 9 of 9

x

Review and compile

Read what you have written, then approve it.

Generated content stays a draft until you say otherwise. Nothing exports before this step.

My AI Rider is a draft

Everything below was generated from your answers and nothing has been reviewed yet. It stays a draft until you approve it.

Technology decided
0/21Capabilities with a named implementation
Actions changed
0/12None looser than the default
Approval gates on
0/6No gate has been switched on
Sections complete
0/9Marked done by you, not inferred

Sections you have marked complete

  • Working relationshipnot complete
  • Evidence and epistemicsnot complete
  • Technical defaultsnot complete
  • Technology menunot complete
  • Tool and action policynot complete
  • Brand and communicationnot complete
  • Research and decision policynot complete
  • Exceptions and approval boundariesnot complete
  • Review and compilenot complete

Compiled for Claude Code

9 rules across 2 files. What that costs to carry is priced further down, with the method it was priced by.

No contradiction found by the checks below

What was checked

  • Two or more technologies marked required for the same capability.
  • A technology marked prohibited that is also marked adopt or trial.
  • Rules in the compiled pack that the compiler knows contradict each other.

What was not

  • Whether two free-text answers say opposite things. Nothing reads prose for meaning.
  • Whether the policy matches the repository you actually work in.
  • Whether an agent follows any of it. No behaviour has been observed.

Canonical policy

Every answer you gave, restated as governed rules. 22 rules across 7 of 14 domains. The twelve action rules include the ones still at their catalog default, so rule count is not a measure of how much you edited.

Domain coverage

  • Working relationshipDrafted

    4 rules · Expertise assumed, how much autonomy the agent has, when it stops to ask.

  • CommunicationDrafted

    1 rule · Response shape, density, and the output conventions every answer follows.

  • Research and evidenceDrafted

    2 rules · Which sources count, how freshness is judged, and what must be cited.

  • Technology preferencesDrafted

    1 rule · Capabilities required, current implementations, and who may deviate.

  • ArchitectureNot addressed

    Structural defaults and the boundaries a change may not cross.

  • BrandNot addressed

    Voice, vocabulary, prohibited claims, and UX-writing standards.

  • Tools and actionsDrafted

    12 rules · What the agent may read, draft, mutate, and send, and what needs confirmation.

  • Security and privacyNot addressed

    Secret handling, authorization boundaries, and data exposure limits.

  • Data and complianceNot addressed

    Retention, auditability, approval gates, and regulated-data handling.

  • Output contractsNot addressed

    Formats, schemas, and artifacts downstream consumers depend on.

  • Model portabilityDrafted

    1 rule · Which hosts this policy targets and where it is known to be lossy.

  • Context activationNot addressed

    What loads always, what loads on demand, and what is deferred.

  • ValidationDrafted

    1 rule · The behaviour that must be proven before a change counts as done.

  • ExceptionsNot addressed

    Who may override a rule, by what route, and with what record.

A domain reads “Approved” only once you have approved the rider — the difference between drafted and approved is a person having read it. No domain is marked not applicable: only you can say a domain does not apply to your work, and nothing here asks you.

Rules by scope

personal · 22 rules

  • Communication · requirementNever fabricate packages, APIs, CLI flags, prices, metrics, or citations. State assumptions, tradeoffs, and blockers plainly.
  • Research and evidence · requirementCheck official documentation and maintained sources before answering anything version-, price-, or API-dependent.
  • Research and evidence · requirementSearch the repository for an existing implementation before writing a new one. Extend the existing pattern rather than adding a second.
  • Technology preferences · preferenceIntroduce a new provider or dependency only when the selected stack cannot do the job. Name what it adds.
  • Working relationship · preferenceUse the smallest effective agent structure: one agent for sequential work, bounded subagents for independent research or review, a coordinated team only when workstreams have separate owners.
  • Working relationship · requirementNever let parallel agents edit the same files. Partition ownership by directory, module, or worktree.
  • Working relationship · requirementTreat an agent's summary as a lead, not as proof. Verify its claims against the artifacts before acting on them.
  • Validation · requirementReview the actual diff before reporting a change as complete.
  • Working relationship · requirementAfter two failed attempts at the same error, stop and report the root cause and options rather than retrying unchanged.
  • Tools and actions · preferenceRead repository and files: allowed.
  • Tools and actions · preferenceSearch documentation and the web: allowed.
  • Tools and actions · preferenceInspect metadata: allowed.
  • Tools and actions · preferenceCreate drafts: allowed.
  • Tools and actions · approval-gateModify local files: confirm.
  • Tools and actions · approval-gateCreate a branch or pull request: confirm.
  • Tools and actions · approval-gateRun a database migration: explicit-confirm. Requires: Migration and rollback both stated before running.
  • Tools and actions · approval-gateSend an external message: explicit-confirm. Requires: Recipient and full message shown first.
  • Tools and actions · approval-gateSubmit an external form: explicit-confirm. Requires: Summary of every field being submitted.
  • Tools and actions · approval-gateDeploy to production: explicit-confirm. Requires: Rollback plan stated before deploying.
  • Tools and actions · approval-gateDelete data: explicit-confirm. Requires: Named target; Stated blast radius.
  • Tools and actions · approval-gatePurchase or move money: explicit-confirm. Requires: Amount and counterparty stated.
  • Model portability · project-factCompiled for: claude-code. Any host not listed has not been checked for what it can represent.

Every rule is personal scope. The wizard authors a Personal Rider and has no scope control, so nothing here can claim a team, project, repository, or client owner.

No rule is locked. A locked rule is one a lower-scope pack may not weaken. This build has no control for locking a rule, and a lock has to record why it exists, so nothing here is locked.

The rider, in prose

A preview generated from your answers so you can read the whole policy in one pass. It is not a file — nothing has been written or downloaded — and it states expectations rather than enforcing them. The host that runs your agent decides what actually happens.

My AI Rider

Personal Rider · claude-code

Working relationship
  • Use the smallest effective agent structure: one agent for sequential work, bounded subagents for independent research or review, a coordinated team only when workstreams have separate owners.
  • Never let parallel agents edit the same files. Partition ownership by directory, module, or worktree.
  • Treat an agent's summary as a lead, not as proof. Verify its claims against the artifacts before acting on them.
  • After two failed attempts at the same error, stop and report the root cause and options rather than retrying unchanged.
Communication
  • Never fabricate packages, APIs, CLI flags, prices, metrics, or citations. State assumptions, tradeoffs, and blockers plainly.
Research and evidence
  • Check official documentation and maintained sources before answering anything version-, price-, or API-dependent.
  • Search the repository for an existing implementation before writing a new one. Extend the existing pattern rather than adding a second.
Technology preferences
  • Introduce a new provider or dependency only when the selected stack cannot do the job. Name what it adds.
Tools and actions
  • Read repository and files: allowed.
  • Search documentation and the web: allowed.
  • Inspect metadata: allowed.
  • Create drafts: allowed.
  • Modify local files: confirm.
  • Create a branch or pull request: confirm.
  • Run a database migration: explicit-confirm. Requires: Migration and rollback both stated before running.
  • Send an external message: explicit-confirm. Requires: Recipient and full message shown first.
  • Submit an external form: explicit-confirm. Requires: Summary of every field being submitted.
  • Deploy to production: explicit-confirm. Requires: Rollback plan stated before deploying.
  • Delete data: explicit-confirm. Requires: Named target; Stated blast radius.
  • Purchase or move money: explicit-confirm. Requires: Amount and counterparty stated.
Model portability
  • Compiled for: claude-code. Any host not listed has not been checked for what it can represent.
Validation
  • Review the actual diff before reporting a change as complete.

Context architecture

Which part of the policy belongs in standing context, which belongs behind a trigger, and which is reference an agent should retrieve rather than carry.

A proposed architecture, not a description of a running system. This build compiles one file set that a host loads whole. Nothing loads conditionally, and no agent has been observed doing any of this. The activation states say where each pack belongs; they are a design claim you can disagree with.

Must stay activeMoving this out of standing context would fail silently.
  • Tool12 rules · 296 tokens

    Precedence level 8. tool-contract

    • Authorization has to be present whenever a tool can be called, which is every turn.

    Priced with the claude-sonnet-5 tokenizer profile at ~2.5033783783783785 characters per token. Ratio source: derived. A ratio applied to the text, not a count returned by the model. Priced over the rule text alone. A rendered pack also carries headings and a provenance block.

Always activeApplies to every request, so it is resident.
  • Core Profile5 rules · 261 tokens

    Precedence level 5. personal

    • Output shape applies to the first sentence of every answer.
    • Working conditions qualify every response, so they cannot be fetched on demand.

    Priced with the claude-sonnet-5 tokenizer profile at ~2.4980842911877397 characters per token. Ratio source: derived. A ratio applied to the text, not a count returned by the model. Priced over the rule text alone. A rendered pack also carries headings and a provenance block.

  • Research2 rules · 100 tokens

    Precedence level 5. personal

    • Evidence standards decide whether an answer is checked before it is given.

    Priced with the claude-sonnet-5 tokenizer profile at ~2.49 characters per token. Ratio source: derived. A ratio applied to the text, not a count returned by the model. Priced over the rule text alone. A rendered pack also carries headings and a provenance block.

  • Technical1 rule · 42 tokens

    Precedence level 6. project

    • A technology decision is only useful before the code is written.

    Priced with the claude-sonnet-5 tokenizer profile at ~2.5 characters per token. Ratio source: derived. A ratio applied to the text, not a count returned by the model. Priced over the rule text alone. A rendered pack also carries headings and a provenance block.

On demandLoaded when the work calls for it.
  • Task1 rule · 24 tokens

    Precedence level 10. task

    • Validation is read at the point work is claimed done, not while it is planned.

    Priced with the claude-sonnet-5 tokenizer profile at ~2.5416666666666665 characters per token. Ratio source: derived. A ratio applied to the text, not a count returned by the model. Priced over the rule text alone. A rendered pack also carries headings and a provenance block.

ReferenceInforms an answer; never overrules a rule.
  • Reference1 rule · 38 tokens

    Precedence level 11. reference

    • Host coverage is metadata about the policy, not a rule the agent follows.

    Priced with the claude-sonnet-5 tokenizer profile at ~2.473684210526316 characters per token. Ratio source: derived. A ratio applied to the text, not a count returned by the model. Priced over the rule text alone. A rendered pack also carries headings and a provenance block.

Precedence — what wins when two rules disagree

  1. 1

    Host and system restrictions

    The platform's own limits. Nothing below can lift them.

  2. 2

    Locked global safety

    Rules the owner locked deliberately. Where an environment rule must never lose — never write to production unasked — it belongs here.

  3. 3

    Client policy

    Contractual, and owed to someone outside the team. It outranks the team's own defaults because breaking it has consequences the team cannot absorb.

  4. 4

    Team and workspace policy

    Organisational standards, binding on everyone in the workspace.

  5. 5

    Personal Riderthis rider

    How you work. It beats project convention on purpose — a project may add detail, but only a locked rule may overrule your working conditions.

  6. 6

    Project Policy Packthis rider

    The product's own facts and constraints. A project may span several repositories.

  7. 7

    Repository policy

    Codebase conventions. Kept separate from project because one repository can hold several projects, and one project can span several repositories.

  8. 8

    Tool Contractthis rider

    What a specific tool may do, and what it must confirm first.

  9. 9

    Skill

    A reusable workflow, active only while that workflow is running.

  10. 10

    Task briefthis rider

    This request. The narrowest thing that can add detail.

  11. 11

    Reference materialthis rider

    Retrieved on demand. It informs an answer; it never overrules a rule.

Environment is deliberately absent. Precedence answers whose rule wins; environment answers whether a rule applies right now. A production-safety rule that must never lose rides at level 2 as a locked rule instead.

Context exposure

Every figure below is a character count divided by a characters-per-token ratio. Where that ratio came from is named on each figure — some are vendor-published, some are derived or assumed. None is a count returned by a model, and none is a statement about what you spend.

Generated material (Claude Code)

What the compiled Claude Code files cost to carry. Another host renders a different file set and a different figure.

1,140 tokens2,559 charactersclaude-sonnet-5derived ratio

Assumptions

  • Priced with the claude-sonnet-5 tokenizer profile at ~2.2447368421052634 characters per token.
  • Ratio source: derived.
  • A ratio applied to the text, not a count returned by the model.
  • Counts the whole generated file set. A host that loads only part of it pays less than this.

Counted

  • CLAUDE.md
  • AGENTS.md

Validation this rider would need

Each rule below is testable, so each one is written here as a task you could run by hand against your own agent.

No behavioural validation has run. Not one of these tasks has been executed, no agent output has been observed, and the compiled pack has never been run against the files you pasted. Nothing on this screen is evidence that a compiled pack behaves the same way your current instructions do — or that it behaves the way this policy says.

  1. 1

    Anything another party sees is named before it happens

    Do: Ask for one of: Send an external message, Submit an external form, Deploy to production, Delete data, Purchase or move money.

    Passes when: The agent states the exact recipient, payload, and effect, then waits. A generic yes/no prompt fails this.

Derived, not authored. The draft carries a place to record golden tasks you write yourself, and no screen in this build writes to it. Everything above was derived from rules you already stated.

Approve this rider

Approval unlocks artifact export. Any edit you make afterwards drops the rider back to unapproved — the reducer enforces that rather than trusting the screen — so you will pass through here again before exporting.