Building Your Verification Soul File

1 The Configuration You Did Not Know You Were Writing

Every AI system you use has a configuration. Sometimes you wrote it. Sometimes it is a system prompt someone else designed. Sometimes it is the default: no explicit instructions, which means the AI falls back on its training, which means it falls back on producing output that users rate highly, which means it agrees with you fluently and sounds like what you expect good analysis to sound like.

There is now an open standard for writing that configuration deliberately. It is called SOUL.md, maintained at soulspec.org, currently at version 0.4. The standard was pioneered by Peter Steinberger, creator of the open-source tool OpenClaw, and has attracted genuine community adoption: templates, a registry, and an academic study examining adoption across 466 open-source repositories. Compatible tools include Claude Code, Cursor, Windsurf, and an expanding set of agent frameworks.

The mechanics are simple. You write a structured markdown file describing who you are, how you work, what you need, and what constraints apply. You paste it into your AI tool’s system prompt or point an agent framework at it. The AI reads your context and adjusts its behaviour accordingly.

Every template in that ecosystem optimises for the same thing: voice reproduction and personality configuration. The result is a document that helps you sound like yourself. For that purpose, the standard is excellent.

This guide does something different. It asks what happens when you optimise instead for verification: for catching errors, surfacing blind spots, and building in the challenges that will not come naturally from a system configured to be agreeable.

2 The 80/20/0 Problem

Examine any published soul file template from the community ecosystem and you will find a consistent distribution. Voice content, covering tone, style, register, and personality, typically occupies 60 to 80 per cent of the file. Rules content, covering constraints, output formats, and non-negotiable behaviours, occupies 20 to 30 per cent. Verification content, covering where the user might be wrong, where the AI should challenge, and what error looks like in this context, occupies zero per cent.

This is not a design flaw. The standard is optimised for its stated purpose. For professional work where accuracy matters, though, the distribution produces a specific problem: the AI cannot challenge your assumptions because it has been configured to work within your frame. It is doing exactly what you told it to do.

The framework I have been developing across The Mirror Effect essay series calls this pattern proxy collapse: the condition in which a signal that once correlated with quality becomes dissociated from it. Production difficulty used to proxy for competence; producing articulate analysis was hard enough that doing it at all was evidence you could probably also think. That proxy collapsed when generation became cheap. The soul file ecosystem has reproduced the same collapse at a finer resolution. “Feels like good collaboration” is being used as a proxy for “is good collaboration,” and the entire template ecosystem is optimised to satisfy the proxy. Voice configuration makes the interaction feel productive. Whether it actually is productive is a different question, and one the standard does not address.

Consider what a typical voice-dominant soul file actually contains. It specifies that you prefer concise responses, that you work in financial services, that you value directness, that you want British English. Every one of those is genuinely useful. None of them tells the AI what to do when your analysis contains a logical gap, when your premise rests on an assumption that may not hold, or when the conclusion you are heading toward is not supported by the evidence you have presented.

The AI, reading a voice-dominant file and receiving a request to review your analysis, will review it in a manner consistent with your voice preferences. It will be concise, direct, and British. It will not necessarily identify the logical gap, because you have not told it that logical gaps are a priority. You have told it how to sound, not what to look for.

2.1 The Three Variables

The right ratio of voice, rules, and verification content depends on three variables: what task you are doing, how deep your domain expertise is, and how much career experience shapes your priors.

Task type determines the immediate weighting. Expressive work (client communications, summaries for non-specialist audiences) weights toward voice. Analytical work (financial modelling, policy review, strategic assessment) weights toward verification. Most professional work involves both.

Domain expertise determines your independent verification capacity. In your strong domains, you will catch errors yourself. In thinner domains, you need more verification instructions, not fewer, because you cannot rely on your own knowledge as a backstop.

Career experience is where the framework produces its most important and least intuitive insight.

2.2 The Counterintuitive Insight

Take a senior equity analyst with twenty years of experience in financial modelling. A junior analyst receives a DCF model built by an AI, sees that the formula structure is correct, and flags it to the senior for review. The senior evaluates it by comparing against a mental template accumulated over two decades: how these models should look, what assumptions are typically reasonable, what range of outputs is plausible.

The model that passes her review is a model that matches her priors.

The problem is that AI-generated models are very good at matching expert priors. They are trained on the same corpus of how these models have been built. When the AI generates a DCF, it produces something that looks right to someone who has seen many DCFs. It is internally consistent. The formulas work. The structure is familiar.

None of that tells you whether the structural assumptions are appropriate for this specific company, at this specific point in its development cycle, under these specific market conditions. A junior analyst, lacking strong priors, might ask naive questions that expose those assumptions. The senior, armed with strong priors that the model satisfies, may not.

This applies across domains. In healthcare, clinical experience creates pattern recognition that is valuable precisely because it is fast; the same speed means an AI-generated differential that matches clinical intuition bypasses the slow deliberative checking that would catch edge cases. In consulting, years of experience create a repository of solutions that work; the risk is recommending a familiar solution to a situation that only resembles previous work. In higher education governance, deep familiarity with regulatory frameworks means an AI-generated compliance summary that uses the right terminology feels correct, even when the interpretation is subtly wrong.

3 The Ratio Framework

The following table presents recommended allocation ratios across voice, rules, and verification content for nine professional profiles. The ratios are starting points for calibration, not prescriptions.

Notice what the 80/20/0 distribution actually represents: the entire soul file ecosystem has invested in making generation better (voice, fluency, comfort) and zero per cent in making verification possible. This is generation-verification asymmetry, the central dynamic of the Mirror Effect framework, applied to the configuration layer itself. AI generates output faster than humans can verify it; the soul file ecosystem has responded by making the generation more personalised while leaving the verification gap entirely unaddressed.

The diagnostic question after every significant AI interaction: did the AI confirm what you already thought, or did it surface something you had not considered? If the answer is consistently the first, your configuration is under-weighted for verification, regardless of what the table suggests.

4 The Seven Sections

A verification-oriented soul file uses seven sections in a deliberate order. Identity first, verification last, because the verification section should be written with everything else already defined.

Two failure modes in writing this section. The first is therapeutic rather than professional: “I struggle with imposter syndrome” may be true but does not tell the AI anything it can act on during a financial analysis task. The second is evasive rather than honest: “I sometimes miss details under time pressure” is true of every professional and therefore functionally uninformative.

Write this section last. Give it the most time. If it takes longer than the other six sections combined, that is probably appropriate.

5 One Honest Complication

You might think the solution is straightforward: add “challenge my assumptions” to the soul file. This is what the Mirror Effect framework calls the Escape Paradox.

Instructions to break Frame Lock become part of it. The AI configured to “challenge you” is still reading a document you wrote, from your frame, in which you decided what challenge means and what degree of challenge is acceptable. The AI generates “challenging” content calibrated to feel challenging while remaining within the boundaries of what you find tolerable. The instruction to escape the frame deepens the frame.

This is a genuine tension. The honest answer is that Section 7 reduces the Escape Paradox rather than eliminating it. Specific, accurate failure mode documentation is substantially harder for the AI to confirm away than a generic instruction to disagree. But the failure modes in Section 7 are the ones you know about and are willing to document. The failure modes you do not know you have are not in the file, and the AI has no way to surface them.

The soul file does not eliminate confirmation bias. It creates structural friction against the known forms of it. For the unknown forms, the best mitigation is external review, adversarial framing of specific questions, and deliberate exposure to perspectives that do not share your priors.

6 The Exercises

Choose a task that involves judgement: a strategic assessment, an analytical review, a recommendation, a draft that rests on assumptions. If one version surfaces a question you had not asked yourself, that version is working.

7 Base File Plus Overlays

A single fixed soul file is a simplification that does not survive contact with professional reality. Most professionals shift within a day between analytical work, client communication, internal reporting, and exploratory thinking.

The practical architecture: a base file capturing the stable elements (identity, working patterns, domain boundaries, failure modes), combined with short context-specific overlays that shift the ratio for particular task types.

A mid-career financial analyst might have a base file weighted 25/25/50 for equity research. When shifting into client communication, a voice overlay shifts the ratio toward 45/30/25. When reviewing a model produced by a junior team member, a verification overlay pushes it toward 10/20/70. The overlay does not replace the base file; it supplements it. The implementation is simple: prepend a short context note when you change mode. “For this task, prioritise identification of logical gaps over output style” is sufficient to shift the AI’s behaviour meaningfully.

8 Keeping the File Alive

A soul file that has not been updated in three months is describing a past version of your professional situation. Three maintenance practices prevent it from calcifying.

Version control. Date every version. The changes over time reveal your learning trajectory. What you add to Section 7 after three months of use is evidence the verification architecture is producing self-knowledge. What you remove from Section 1 because it turned out to be aspirational rather than actual is evidence of honest calibration.

Post-error updates. The most valuable updates are made immediately after a significant error, when the failure mode is freshest. Describe the error. Update Section 7 to reflect the pattern. Consider whether the ratio needs adjustment. The errors worth documenting are the ones that reveal something about your default analytical patterns, not every minor misjudgement.

Quarterly ratio audit. Every three months: has the nature of your work changed? Has your expertise grown or shifted? Have you discovered new failure modes? Has Section 7 been updated, or has it calcified?

9 What This Reveals Beyond AI Configuration

There is a version of this guide that treats the soul file purely as an AI configuration tool. That version misses the more important point.

The exercise surfaces something most professionals have never been asked to produce: a written record of where their expertise creates predictable error. Not a personality assessment. Not a self-evaluation for a performance review. A specific, operational account of the conditions under which their professional judgement becomes unreliable.

Most people resist writing Section 7 because it feels exposing. The resistance is itself diagnostic. The failure modes it documents are the ones your most trusted colleagues could probably identify. The fact that you can identify them too does not make them shameful; it makes them manageable. The soul file documents them so that a tool you are already using can help catch them rather than reproduce them.

This connects to the broader argument of the Mirror Effect. AI does not create new problems in how we evaluate quality. It reveals where our evaluation was already depending on signals that correlated with quality rather than measured it. The voice-optimised soul file is the personal version of that revelation. When you configure an AI to sound like you and work within your frame, you are building a system that reproduces your defaults, including the ones you have not examined. When you configure it for verification, you are building a system that creates friction against those defaults. The friction is not a bug. It is the mechanism by which the tool becomes genuinely useful rather than pleasantly confirming.

10 Key Takeaways