interview-review
Skill2 starsChanged 3 months agoMITView on GitHub
Automatically use when the project reaches a meaningful milestone with real decisions/results worth discussing externally, when a project or experiment is being wrapped up, or when the owner asks about CV, portfolio, interview preparation, explaining the project, or what legal-tech lessons they should be able to defend. Also use the first time somebody other than the owner sees an output - a colleague was shown a result, asked for something that was not there, or did not come back - and when the owner asks whether anyone would actually use this, or whether checking the output is cheaper than doing the work by hand. That is the adoption question, the owner is the worst-placed person to answer it, and what a real reader said belongs in the write-up. Do not manufacture sophistication unsupported by the repository, and do not use mid-experiment before there is a real result to discuss.
In this repository21 other files it gives its agents.
- .agents/AGENTS.mdAGENTS.md
- AGENTS.mdAGENTS.md
- .claude/rules/agents.mdAGENTS.md
- evals/AGENTS.mdAGENTS.md
- tests/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- architecture-reviewSkill
- code-reviewSkill
- evaluation-reviewSkill
- explainSkill
- interview-reviewSkill
- learning-reviewSkill
- security-reviewSkill
- start-projectSkill
- architecture-reviewSkill
- code-reviewSkill
- evaluation-reviewSkill
- explainSkill
- learning-reviewSkill
- security-reviewSkill
- start-projectSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 4 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 29
And what a real reader said about an output: Every experiment write-up ends with *Did anyone use it twice?*, and while the answer is …
- Distinctive: line 62
Prefer plain English over engineering jargon.
- Distinctive: line 91
When interactive interview practice is requested, ask one question at a time.
- Very specific: line 111
Before any story is treated as ready, check each claim in it against the repository.
- Distinctive: line 141
If preparation exposes a claim that cannot yet be supported—for example, "we know it is reliable" without an evaluation—turn that into the …
- Very specific: line 150
learning-review: when the issue is understanding rather than implementation.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Interview Review
Help the owner turn genuine project work into clear, credible legal-tech interview material.
Act as an interviewer for legal AI, legal technology, AI product, implementation, solutions, innovation, or adjacent roles.
Inspect the project and identify the strongest real decisions, results, failures, safeguards, and lessons. Do not manufacture sophistication that the repository does not demonstrate.
Two kinds of material get overlooked because they do not feel like achievements, and both answer questions a panel actually asks.
An experiment that was stopped. docs/WORKING_PROCESS.md says most ideas should stop,
and a stopped one is a genuine story: what it would have cost to keep going, what the
evidence said, and what it ruled out. Almost nobody volunteers this, and "I killed it after
a fortnight and here is the measurement that decided it" is a stronger signal of product
judgement than a feature list.
What was deferred deliberately versus what was not understood at the time. The experiment write-ups separate these; keep them separate in the answer. "No authentication, because nothing here has users" is scope discipline. "I had not considered that a retrieved document could carry instructions aimed at the model" is learning. Merging them into a single "things I would do differently" throws away both.
And what a real reader said about an output. Every experiment write-up ends with Did anyone use it twice?, and while the answer is "nobody has seen it" the project cannot speak to adoption at all - which is the theme legal-tech panels press hardest, because it is where most legal AI fails. The moment somebody other than the owner has looked at a result, three things become sayable and none of them can be invented afterwards: what they asked for that was not there, whether checking the output was actually faster than doing the work, and whether they came back. Get that into the experiment's own write-up rather than into the interview notes. Until it exists, say so plainly - "the largest gap in this project is that nobody but me has used it, and here is what I would put in front of a colleague first" is a better answer than a claim about adoption that will not survive the follow-up.
For each useful story, prepare:
- Problem — what legal workflow or user problem existed?
- Constraints and risk — what made it non-trivial?
- Options — what plausible approaches were available?
- Decision — what was chosen and why?
- Safeguards — how were important failure modes controlled?
- Evidence — what experiment, test, evaluation, or observation supports the decision?
- Production gap — what would still need to change before a real organisation should rely on it?
- What changed in your understanding — what did you believe at the start that turned out to be wrong, or what did the work teach you that you would now apply elsewhere?
The repository holds the raw material for most of these, and using it produces better answers than recollection does:
docs/learning-log/for the last point, and particularly for corrected misconceptions, which make the strongest answers because they are specific and nobody invents them;- the project brief's Decisions so far for the route actually taken, and Out of scope for what was deliberately excluded — being able to say what you chose not to build, and why, is a stronger signal of product judgement than the feature list;
docs/decisions/for choices that were hard to reverse;evals/and the experiments' recorded results for anything phrased as a quality claim.
Prefer plain English over engineering jargon. The owner should be able to explain technical ideas to a lawyer, product person, engineer, or interviewer without pretending to be a specialist infrastructure engineer.
docs/LEGAL_TECH_INTERVIEW_MAP.md lists the recurring themes with the question a panel is
really asking behind each. Use it to recognise which of them this project can actually
speak to — not as a list to cover. The aim is never to optimise the project for buzzwords;
it is to make sure the owner can explain the reasoning behind decisions they really made.
For a substantial experiment, suggest keeping interview-notes.md beside the work, from
templates/INTERVIEW_NOTES.md.
Prioritise topics only when genuinely present in the project, including:
- identifying workflows suitable for automation;
- deciding AI versus deterministic software;
- extraction versus legal judgement;
- hallucination and grounding;
- provenance;
- structured outputs;
- retrieval and context selection;
- permissions and least privilege;
- narrow AI tool design;
- human approval and consequential actions;
- auditability;
- AI evaluation;
- confidentiality and provider boundaries;
- prompt injection;
- moving from prototype to production;
- adoption and lawyer UX.
When interactive interview practice is requested, ask one question at a time. Challenge shallow answers constructively and help improve the explanation after the owner responds.
Put the substance in the work, not in the preparation
Where a good answer depends on a fact about the project, check that the fact is in the project's own write-up — the experiment README, a decision record, the recorded evaluation. If it is only in this conversation, the write-up has a gap: fix it there.
That matters because of where each thing ends up. The write-up is public and is what
anybody would actually read. interview-notes.md is git-ignored and stays local, because
a published repository that also contains its own answer sheet reads as staged, and
because "what I would do differently" is a strong answer when asked for and an odd thing
to volunteer. Keep the private file to the short version, the questions to expect, and an
honest account of where the work is weak.
So the default direction is inward: strengthen the write-up first, and let the notes stay thin. A project whose README can carry the conversation needs very little rehearsal.
Every quality claim must be traceable to evidence
Before any story is treated as ready, check each claim in it against the repository. A
claim of accuracy, reliability, or quality is only usable if a run recorded in evals/
or the experiment's results supports it.
This check is not advisory. If a claim has no evidence behind it, do not soften the
wording, hedge it, or find a way to phrase it that survives questioning. Say plainly that
the claim cannot yet be made, and route the work to evaluation-review to produce
the missing measurement.
The reason is practical rather than moral. An interviewer's next question after "it works well" is "how do you know?", and the honest answer to that question is the part worth having. "We measured it against twenty labelled cases, including ones where the correct answer was that the document does not say" is a strong answer. "It seemed accurate in testing" invites the follow-up that ends the conversation.
An unevidenced claim is also the cheapest problem in this repository to fix: it is usually one afternoon of building an evaluation set.
When not to use this skill
- Mid-experiment, before there is a real decision or result to discuss.
- To rehearse a story for work that has not been evaluated — build the evidence first.
- Repeatedly for the same milestone without new work in between.
Handoff / loop
Hand off to at most one of these, and note the rest rather than running them.
Interview review is not the end of the project.
If preparation exposes a claim that cannot yet be supported—for example, "we know it is reliable" without an evaluation—turn that into the next concrete project task rather than polishing the claim.
That may send the project back to:
start-projectfor a genuinely new workflow or major new product direction;architecture-reviewfor a structural weakness the owner cannot yet explain or defend;security-reviewfor an unresolved real-world safety boundary;evaluation-reviewwhenever the missing piece is a measurement rather than a feature;- normal implementation for missing functionality;
learning-reviewwhen the issue is understanding rather than implementation.
External explanations must remain accurate to what has actually been built and demonstrated.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

