CONTEXT
The work started here.
A private, file-based system prepared for a senior role using public research and structured mock interviews.
BUILD
What was built.
A file-based preparation system for a senior role at a security compliance company: a citation-enforced knowledge base compiled from 100+ sources, ten interviewer personas built from public data, and a mock-interview mode that rehearses in their voices. No app, no database, no server: a folder structure with rules, built in about seven evenings.
PROJECT DESIGN
How it was shaped.
Parallel research agents under a hard citation rule, a grading audit that triggers targeted follow-up research for anything below the bar, and a compile step forbidden from generating content: it only merges, deduplicates, and flags contradictions. Multiple agents worked the same folders overnight, coordinated by naming conventions and a handover note rather than any orchestration framework.
SECURITY & OPS
The operating decisions.
- Every claim in the knowledge base carries one of three states — confirmed, inferred, unknown — with inline citations, so facts travel intact when excerpted.
- The agent that writes research is never the agent that grades it: auditor skills are barred from generating content.
- Research files graded below B trigger targeted follow-up rounds aimed only at the specific gaps the audit found.
- Contradictions between credible sources are preserved side by side with an assessment — the system never silently elects a single truth.
BUILDER NOTES
What the build taught.
- No database, no app, no server: the whole system is markdown on disk, skill definitions, and directory contracts that concurrent agents follow without colliding.
- Coordination happens by filesystem convention — write-once research, append-only reference files, handover files between sessions — not an orchestration framework.
- Markdown compiles to self-contained HTML slide decks with zero external dependencies, and every deck is screenshot-verified before it counts as done.
- Skills were extracted from repetition, not pre-designed: the presentation skill appeared on day four, after the same conversion had been done by hand several times.
LESSONS
What changed afterward.
- Citations are a data format, not a virtue: claims must carry their sources inline so they survive being merged, split, and excerpted by later steps.
- Keep the writer and the grader separate: the auditing steps are non-generative by rule, so research is never grading itself.
- Realism in rehearsal comes from depth, not hostility. The first answer is the prepared one; the second question is where real experience shows.
- Contradictions between credible sources get preserved and labeled, not silently resolved, so downstream work inherits honest uncertainty instead of false confidence.
WHAT CHANGED
The compiler stopped writing.
The compile step was restricted to merging, deduplicating, and flagging contradictions so content generation could not hide uncertainty during assembly.
CONTROL
The constraint that held.
Claims carry inline citations, graders trigger targeted research, and credible contradictions remain labeled rather than silently resolved.
EVIDENCE
Where to inspect it next.
- Every claim in the knowledge base carries one of three states — confirmed, inferred, unknown — with inline citations, so facts travel intact when excerpted.
- The agent that writes research is never the agent that grades it: auditor skills are barred from generating content.
- Research files graded below B trigger targeted follow-up rounds aimed only at the specific gaps the audit found.
- Contradictions between credible sources are preserved side by side with an assessment — the system never silently elects a single truth.
CARRIED FORWARD
What survived the project.
A general blueprint for becoming rapidly, defensibly conversant in any organization and its people: reusable for due diligence, sales prep, or expert-witness work.
WRITING
Field notes from this work.
No public field notes for this project yet.