Judicial First Impression casts the model as an experienced judge reading a submission cold under time pressure, with seven fixed headings:

It is emphatically not an attack skill. In Meredith-Flister's own framing: "Your task is not to improve the argument. Your task is not to attack it. Your task is to assess — honestly, precisely, and without encouragement — how it actually lands on first reading."

Meredith-Flister has also published a "Relationship to Other Skills" note, which sets out how this skill sits alongside Opposing Counsel Review and a third: a persuasive legal writing skill that builds and strengthens arguments. The suggested workflow is: run Judicial First Impression first, to understand how the argument actually lands; then Opposing Counsel Review to stress-test it in an adversarial environment; then the persuasive writing skill to strengthen it in response.

The same hypothetical from Part 1 is reused here: a Part 24 summary judgment application by Northfield Analytics Limited against Brimstone Retail Group Limited for £1,287,500 in unpaid subscription fees, listed before the Commercial Court on 22 May 2026. The facts, parties and claim number are entirely invented. The full text of the skeleton sits in Part 1.

The deliberate defects are set out in full in Part 1. In summary: an inflammatory tone, citing clauses without setting them out, no engagement with the UCTA 1977 s.3 or Misrepresentation Act 1967 s.3 reasonableness controls, dismissal of the Defendant's evidence by characterisation rather than analysis, and indemnity costs sought without articulated basis. Throughout, the skeleton argues the merits of the underlying claim rather than the narrow Part 24 question.

What follows is the complete output of Judicial First Impression run on that skeleton, in an isolated instance in Claude Co-Work, with no prior context and no visibility of the earlier Opposing Counsel Review output. The model used was Opus 4.7. Quotes are verbatim, with original punctuation preserved (including the em dashes, which are themselves a useful tell for readers trying to spot AI-authored text).

The skill produced the following seven-section cold-read assessment. Commentary follows each section.

Comment. The summary is accurate and economical. What is noteworthy is the final clause: the skill has correctly identified which of the two Part 24 limbs the skeleton is really arguing, and flagged that the other is underdeveloped. A judge reading the skeleton would indeed notice that.

Comment. Six specific, paragraph-referenced points of confusion. Every one is correct. The "material" observation is the same point Opposing Counsel Review made, from a very different register: the judicial version does not accuse the Claimant of smuggling anything in; it simply notes that the word is doing work the skeleton has not earned.

Comment. This section is where Judicial First Impression shows its value most clearly. Opposing Counsel Review is explicitly forbidden from identifying strengths. A drafting solicitor needs both perspectives. The judicial skill has correctly identified that the chronology of payment and complaint is genuinely helpful to the Claimant, and that the quantum point at paragraph 18 is the skeleton's strongest single paragraph. A reviewing partner would want both of those observations preserved in the final version.

Comment. This is a judicial voice, not an advocate's. Note how the skill engages with the statutory framework at a level the adversarial skill did not: the reference to section 49 of the Consumer Rights Act 2015 and the Supply of Goods and Services Act 1982 is the sort of detail one would expect from a judge reading the skeleton, and the point about indemnity costs is a specific, correctable drafting defect. The observation that the skeleton takes the Defence's case at its lowest rather than its reasonable height is a genuinely judicial instinct.

Comment. This section is the one this author finds most useful. The framing ("the argument appears to assume X") is the right way to think about unspoken premises in any legal submission, and the skill reliably produces it. The three suspected assumptions named here are the three load-bearing pieces of the skeleton that the Claimant has not actually supported on paper.

Comment. "Low to medium" is marginally a hedge, and the skill's own guidance warns against hedging. But the reasoning that follows is honest and calibrated. A reader focused on the application as drafted might say "low", since the skeleton would not win it. A reader focused on the underlying claim might say "medium", because the contractual architecture is strong. The skill has split the difference for the right reason: it is assessing the skeleton, not the underlying claim, and the skeleton has failed to bridge between the two.

Comment. This is a clean punch-list. A reviewing partner could take this seven-item list, walk the drafter through it, and the next version of the skeleton would be meaningfully better. That is what the skill is for. It is worth noting that the list does not tell the drafter how to fix any of these things, in keeping with the skill's express rule against supplying authorities or content the submission omits. That restraint is harder than it sounds for an AI, and it is preserved here.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

Both skills are published under the Apache 2.0 licence. That matters, more than many legal readers may realise.

In short: UK law firms can download, use, modify and combine these skills with their own work, including for paying client matters. Attribution must be preserved. The detail follows for those who want it.

Apache 2.0 is a permissive open-source licence. It allows anyone, including commercial law firms, to:

In exchange for that freedom, the licence requires three things. The first is attribution: any copy or derivative must retain the original copyright notice and a copy of the licence. The second is that any modifications to the original files must be marked as modified. The third is an express patent grant: Meredith-Flister has granted users a licence to any patents she holds that are necessary to use the skill, and that grant terminates automatically if the user sues anyone else alleging that the skill infringes a patent. That patent clause is one of the reasons Apache 2.0 is preferred over some other permissive licences (such as MIT) for professional use: it removes a specific and otherwise awkward risk for firms.

For UK law firms, the practical implication is this: you can download, use, and modify these skills freely, including for commercial client work. You must preserve the attribution. You do not have to publish any bespoke additions you build on top. Some firms' internal open-source policies require sign-off before introducing any Apache 2.0 dependency, particularly into client deliverables. If yours does, this is the licence your internal team will need to be told is in play. But as open-source licences go, Apache 2.0 is about as friendly to commercial and professional use as it gets, and it is a deliberate choice on Meredith-Flister's part that the skills are published this way. It is another piece of generosity alongside the skills themselves.

The most useful single observation from running both skills on the same document is that they catch overlapping but not identical issues. Both noticed the UCTA silence. Both noticed the "material interruption" qualifier at paragraph 5. Both flagged the lack of engagement with the Defendant's actual pleaded case. But the adversarial skill went further on strategy (the judicial question that would sink the application; the five oral-submission points), while the judicial skill went further on drafting defects (the statutory framework on implied terms; the unsupported indemnity costs request; the awkward sequencing of the October 2025 letter and the April 2026 Defence). Neither is a substitute for the other. Used in sequence, as Meredith-Flister suggests, they give a drafting solicitor two different kinds of information: one tells you how the other side will attack; the other tells you what a judge will actually notice on first reading.

This is also where the structural discipline of both skills earns its value. A prompt you write yourself in the moment will catch some issues. A skill with fixed output headings will catch the same kinds of issues on Tuesday that it catches on Friday, and on this hypothetical that it catches on your live drafts. For a firm introducing AI-assisted review into litigation practice, that reproducibility is not a nice-to-have. It is the difference between a tool you can build a supervisory process around and a tool you cannot.

This author promised a balanced review, not a marketing piece. Two honest concerns are worth naming about Judicial First Impression specifically, plus a final observation that applies to both skills.

A third concern, around citation risk, was discussed in Part 1 and is not repeated here. However, in writing this Part 2, it is noted that Meredith-Flister has created a new skill called Source Locked Verification, which will be examined in a future article.

The first is the pull towards the middle on confidence calibration. Judicial First Impression explicitly instructs the model not to default to "medium" out of politeness, and the output here came back "low to medium" with reasoning that genuinely earned the split. But "low to medium" is still a hedge relative to the three levels the skill asks the model to choose between. This writer suspects that a genuinely hopeless submission would get "low" and a genuinely persuasive one would get "high", but that the grey zone is where calibration wobbles. Users should read the confidence level as directional and weight the reasoning more than the label.

The second is the gap between the "UK" tag and substantive jurisdictional encoding. Neither SKILL.md cites the Civil Procedure Rules, Practice Direction 57AC, the Commercial Court Guide, the Chancery Guide, or any other source specific to English procedure. The UK-ness is a register instruction (British English, the tone of English advocacy) and a set of usage conventions. The substantive English law in the outputs comes from the underlying model's training knowledge, which is competent but not reliable in the way a firm's internal precedent or a procedural rulebook is reliable. This does not undermine the skill. Designing for cross-model use (Claude, ChatGPT, Gemini) makes deep jurisdictional encoding impractical, and treating "UK" as a register and tone tag is a sensible design choice. Users should simply read the label as such, and not as a jurisdictional guarantee.

A final observation, less critical. Lawvable's marketing copy describes skills as producing "consistent, structured output every time it runs". The structure is indeed consistent (headings, order, overall shape). The substance is not. Two runs on the same document, even with the same model, produce meaningfully different primary lines of attack and different surgical strikes. This is a property of the underlying model, not a failing of the skill, but the word "consistent" in the marketing is doing work it cannot fully sustain. Readers should understand what "consistent" means in this context: the frame is the same, the content of the frame is not.

This writer has used both skills on real work over the past month (on drafts, not on live submissions, and only on anonymised or synthetic material). They do what they say. They do it with a register and a discipline that is unusually good for prompt-based AI work. Opposing Counsel Review is particularly useful during the drafting phase of a contested interim application, where the real question is "what will the other side say?" and where self-editing tends to be weakest. Judicial First Impression is useful at the point of finalisation, as a proxy for the hour of judicial reading the court will never give an advocate a proper preview of. Both, used in sequence, have saved this author meaningful time and have surfaced real issues this writer would otherwise have missed.

Reader response to Part 1 has been warm, with open rates well above the newsletter's typical run and replies that suggest these skills are landing with UK practitioners. The wider community response to Meredith-Flister's work has been similarly positive. That is the right reception, in this writer's view.

Larissa Meredith-Flister has done the profession a service by building these tools, thinking about them carefully enough to write proper instruction files, releasing them openly under a permissive licence, and inviting the rest of us to test and improve them. That is not the default direction of travel for legal tech, and it deserves to be recognised. This author's encouragement to other UK lawyers is straightforward: download the skills, try them on your own drafts (anonymised where appropriate), and, if you find them useful, say so publicly. Work like this gets better when the profession engages with it, and the profession gets better when work like this exists.

If readers would like to look at the skills themselves, both are hosted on Lawvable at the links below, and both can be downloaded and used on Claude, ChatGPT, or Gemini without any vendor lock-in.

If you have tried these skills, hit reply and tell me what you think. I would love to hear what other types of arguments people have tested it with. I read every email!

If there are other skills or AI tools that you want to see tested, let me know!

Thanks for reading,

Serhan, UK Legal AI Brief

Guidance and news only. Not legal advice. Always use AI tools safely and in line with best practice.