Caravel, AI Led SDLC by Midships

Same request. Same specification. Verifiable output.

Most teams use AI by prompting it. What comes back depends on who asked, how they phrased it and what the model guessed to fill the gaps. Caravel works from versioned context instead, so delivery is predictable and the knowledge stays with your organisation, not in someone's chat history.

Prompt driven
Ticket, individual prompt, model guesses, review, rework
Context driven with Caravel
Declared intent, locked specification, personas with versioned knowledge, golden tests, human arbitration
1. One request, two paths

Where the variation comes from

The same change request run through both approaches. Prompt driven delivery inherits the inconsistency of the people prompting. Context driven delivery removes it before any code is written.

One request

Prompt driven Each developer prompts in their own words Context driven Caravel works from a locked specification Change request Add step up authentication to the new payee journey Developer A "add MFA when apayee is added" Developer B "OTP step up onrisky payees" Developer C "reuse the loginflow for payees" Model fills the gaps Three differentimplementations PASS FAIL NO AUDIT EVENT Review compares against personal judgement, so the work goes round again Specification v1, locked Acceptance criteria and golden tests Knowledgebase v2.4 Templatestep up journey AnAnalyst ArArchitect DvDeveloper TsTester Run 1, run 2, run 3: verified against the same tests PASS PASS PASS Human arbitrates against the spec

Prompt driven: context per call

Each call resends the conversation so far

Context driven: context per call

Each gate passes a bounded package, not the history

Bars show context sent on each model call; the line shows the running total. Illustrative, same scale on both sides.

2. One delivery, step by step

How Caravel takes the request to release

Humans declare intent and arbitrate outcomes. Specialised personas do the work in between, and every step is recorded against the specification it serves.

Change request: step up authentication on the new payee journey

    Illustrative scenario. Version numbers and identifiers are examples.

    3. What changes

    A governed delivery system, not a faster keyboard

    Prompt driven tools make individuals quicker. They also make each individual's habits and gaps part of the output. Caravel moves the quality decision out of the prompt and into context the organisation controls.

    Prompt driven
    Context driven with Caravel
    Starting point
    A prompt written by whoever picks up the ticket
    An intent declared by a human, turned into a machine readable specification that cannot change mid build
    Requirements
    Gaps are filled by the model's best guess
    Personas prompt the humans for what is missing before any code is written
    Where knowledge lives
    In individual heads and chat histories
    In versioned knowledge bases your organisation owns
    How work is shaped
    One general purpose assistant for everything
    Specialised personas and task specific templates
    What review checks
    Whether it looks right to the reviewer
    Conformance to the specification and the golden test set
    Running it again
    A different answer
    Checked against the same specification and tests
    Documentation
    Written afterwards, if at all
    Derived from the code, which is canonical, so it stays current
    Token use
    Compounds: each call resends the conversation so far, and every rework loop adds more
    Bounded at each gate: the next persona receives the specification and relevant knowledge, not the history

    Output quality is set by the quality of the context, not the intelligence of the model.

    4. How it works

    Six settled principles

    Humans declare and arbitrate

    People state intent and judge outcomes. They do not hand craft prompts.

    Personas, not one assistant

    Specialised personas for analysis, architecture, build and test, each with a defined job.

    Versioned knowledge

    Personas work from knowledge bases that are versioned, reviewed and owned by you.

    Task specific templates

    Each kind of task has its own template. There is no universal one.

    Immutable specifications

    Machine readable specifications and golden test sets define done before work starts.

    Code is canonical

    Documentation is generated from the code, so the two stay aligned.

    5. Working with us

    Start with one workload domain

    Caravel's first workload domain is Keycloak journeys, where Midships has deep delivery experience. We start with a bounded set of journeys, build the knowledge base and templates with your team, and prove predictable delivery before widening scope.

    Caravel takes individual prompting habits out of the delivery path, rather than asking people to be more careful.

    Assess one delivery workload
    ImplementationOne off, to establish the framework in your environment.
    Annual framework licenceIncludes the first delivery pack.
    Delivery creditsConsumed only on verified successful delivery.
    InferenceYou carry your own inference costs.