Adam MassarschSession Replay
  1. plan.md loaded
  2. 0 Foundation
  3. 1 Content before design
  4. 2 Design direction
  5. 3 Home page
  6. 4 Case pages
  7. 5 Polish
  8. 6 Launch
  9. 7 Review
  10. 8 Fixes
  11. STOP — waiting for Adam
  12. /kolla ✓ typecheck ✓ lint ✓ build
  13. review: fresh session, different model

I build working AI software with Claude Code: a real-time study app, a prospecting engine with parallel agents, and an agent pipeline that turns process pages into film.

Seminar for groups2/5 · Nov 2026

A shared AI room where a study group uploads its material, asks one AI together and rehearses for the seminar.

$0.47 → $0.03 cost per question in a room with seven papers: the first after a pause, then follow-ups read from the prompt cache.

$0.47cost per question

cold114,000 tokens

114,000 tokens. Each dot is one token. Only the total is drawn, not how much of it is cached.

  1. system prompt
  2. room's documents
  3. one-hour cache breakpoint
  4. what changes per question

Blocks are not drawn to scale.

the first question after a pause114,000 tokens$0.47
follow-ups within the hourread from the prompt cache$0.03

Prompt caching by design.

The prompt is ordered stable-first: system prompt, then the room's documents with a one-hour cache breakpoint, then what changes per question. Measured cost per question: about $0.47 for the first question after a pause in a room with seven papers (114,000 tokens), about $0.03 for follow-ups within the hour.

59 %

PDFs
Text

Text instead of PDFs.

Uploaded PDFs are converted to page-marked text before they reach the model. On real files this cut input tokens by 59 %. Scanned PDFs without a text layer fall back to the original file.

13/13 chat markers

17/17 analysis markers

In testing, 13 of 13 chat markers and 17 of 17 analysis markers passed, each also checked by hand against the PDF.

Read the case: Seminar for groups

Blue Mesa3/5 · Summer 2026

A prospecting service that finds companies with a reason to buy now, built on parallel AI agents and one rule: every contact needs a source.

Two paying customers in summer 2026.

  • verifiedwith link
  • patternwith the example address that proves the format
  • generic
  • not foundNot found is a valid answer.

No rows lost.

Simulated run. The rows are made up and their outcomes are random.

Try to lose a row: drag one out and let go.

For contact research I wrote a Claude Code skill that runs the work in parallel: the company list is split into groups of 15–25, one subagent per group, all started at once.

Every row is classified.

Verified (with link), pattern (with the example address that proves the format), generic, or not found. Not found is a valid answer.

Code writes the result, not the model.

Agents save JSON; a script writes it into the spreadsheet, never overwrites old data silently, and a check script compares the new file with the original so no rows or values go missing.

Read the case: Blue Mesa

How I build4/5

I build with Claude Code. The point is not that an AI writes code; it is that I set up the work so the result can be trusted.

  1. 0Foundation

    STOP — waiting for Adam

    Stop: I ran /kolla and saw CI pass.

  2. 1Content before design

    STOP — waiting for Adam

    Stop: I approved or corrected all text. No design until it was settled.

  3. 2Design direction

    STOP — waiting for Adam

    Stop: I picked one.

  4. 3Home page

    STOP — waiting for Adam

    Stop: I went through it in the browser.

  5. 4Case pages

    STOP — waiting for Adam

    Stop: I approved it before the others were built.

  6. 5Polish

  7. 6Launch

  8. 7Review

    STOP — waiting for Adam

    Stop: I went through the list and decided what to change.

  9. 8Fixes

  1. Plan first, one piece at a time.

    My projects start with a written plan in phases, with a stop after each phase where I test before anything continues. This site was built that way too. Read how this site was built

  2. CLAUDE.md as the rulebook.

    The project file gives every session the same starting point: what the project is, how to run and check it, and the hard rules, such as only stating confirmed facts and stopping at every review point. Read this site’s CLAUDE.md

  3. Budgets in code.

    Every route that calls a model checks a budget before the call and logs the cost after it. Prompts are ordered so the expensive part is cached.

  4. Measure, then decide.

    Built-in citations were measured at 31–46 % more tokens, so I built my own and verify them in code. Text extraction was measured at 59 % fewer tokens before it became the default.

  5. Never invent.

    The agents I write have explicit rules: a source for every claim, “not found” as a valid answer, and code, not the model, writing the final result.

Let’s talk5/5

I’m looking for part-time work or an internship building AI products.

email adam