Session record: 2026-10-05

This file is a chronological record of one working session between the owner and Claude Code. The session added three AI-written guides about one SNES video, reviewed them, and put their lessons on the site. Each round gives the request, the work that Claude Code did, and the result. The decisions are in ../DECISIONS.md. The plan is in ../PLAN.md. The first session is in 2026-09-27.md.

Round 1: three new documents about one video

• What the owner asked: add three new documents to the review. Two AI tools wrote them from the video “It Took Every SNES Hardware Trick To Make My Game” by Inkbox (51:14). In the video, Inkbox finishes a game for the Super Nintendo in pure assembly.
• The documents: a cautious research report that says that it could not get the captions (document 13), a long report with nine “stages” (document 14), and an interactive page made from document 14 (document 15). The YouTube oEmbed service confirmed the title and the channel. The file review/findings/videos.json records the video for the three files.
• The split: 41 new document parts (11, 25 and 5). The splitter needed no change.
• The captions: yt-dlp got the English captions. They stay on this computer, in a scratch folder, and are not in the repository (decision D25).

Round 2: adversarial review of the three documents

• Done: three helper agents reviewed one document each against the captions. The reviews are adversarial-review-pass1/13-…, 14-… and 15-…. The script extract_findings.py put the 80 findings into findings/pass1-snes-video-guides.json (26, 27 and 27). The three files in findings/video-moments/ give the moment of each part in the video. 34 of the 41 parts have a moment.
• Results: document 13 Partly (3 critical, 11 major, 12 minor). Document 14 No (6, 10, 11). Document 15 No (5, 13, 9).
• Largest results:
  • Documents 14 and 15 say the opposite of the video three times. Transparent sprites on tile zero were the problem, not the fix (28:34). Sound effects have three voices of their own and steal nothing from the music (18:18). The developer did not rotate sprite priority (30:21).
  • The tile address formula of documents 14 and 15 joins two shifts with OR. The shifts must be added. The page also uses one byte for each tile, but the video gives two (06:53).
  • Documents 14 and 15 say that the CPU has no hardware division. The 5A22 has a multiplier and a divider.
  • Document 13 puts the audio section at 36:15. It starts at 08:40.
  • Every chart number of the page is invented, and its BCD simulator adds decimal strings, not BCD.

Round 3: consolidation and the concept review

• What the owner asked: consolidate the findings of the three documents, divide them into learning units, and challenge each unit.
• Done: a consolidation brief listed what the video covers, minute by minute, with eleven candidate units and six Fortran ideas. A concept review attacked each unit with the rule: an idea stays only if it serves the 8086 lander or the Fortran laboratory.
• Result: eight units stay, three merge, three drop. The dropped ideas are sound from samples (the PC speaker plays one square wave), a maze with a flood-fill check (the lander has no maze), and the SNES dialect words. The review also found that the dictionaries lacked words that the laboratory code already used: SHIFTA, IAND, MERGE, MODULO, ERROR STOP, NEWUNIT, ACCESS='STREAM', INT16, REAL64, GET_COMMAND_ARGUMENT, XLAT, INCBIN and ADC.

Round 4: where the lessons live, and the review section of the site

• What the owner asked: as the system architect, decide where the learning points belong on the site. Update the review section of the site. Use the review to strengthen the Fortran side. Plan, decompose, implement, regenerate.
• The positioning (D26): the SNES video is a source of ideas, not of words. Each unit became a strength, a limit or a “Watch out” note on an existing segment page. No new page and no SNES page. The table of units and pages is in ../PLAN.md.
  • Keeping values: segment:offset is not bank:offset (U6).
  • Doing arithmetic: a table replaces a slow verb (U2), DAA after ADD (U3), wrap for free on the A side and MODULO on the B side (U8), and the OR formula of the documents as a trap.
  • Working with many values: the terrain lookup (U2), REP MOVSW copies every byte (U5), and the check runs all inputs when they are few.
  • Organizing a program: the table is the contract, and each table gets a wrong twin (U4).
  • Showing text and reading keys: the packed decimal counter (A) and the formats I4.4 and I0 (B) (U3).
  • Controlling the hardware: the frame budget (U1), no HDMA and no sprites on the VGA, Mode X pages and the line compare, and the PIT note protocol (U7).
• The Fortran side: three planned tests and a rule in PLAN.md, Phase 4d. Test 4 checks a decimal counter for all 10,000 values. Test 5 checks the row-start table. Test 6 makes and checks a terrain with one flat pad. The note table gives one PIT divisor for each semitone, with its error in cents. Each test has a wrong twin that must fail. No game code changed in this session.
• The dictionaries: 8 new A-side entries (XLAT, ADC, SBB, DAA, DAS, AAM, CBW / CWD, INCBIN) and 13 new B-side entries (INT8 / INT16, REAL64, MODULO, IAND, ISHFT / SHIFTA, RANDOM_NUMBER, ERROR STOP, BTEST, MERGE, NEWUNIT=, ACCESS='STREAM', the formats I4.4 and I0, GET_COMMAND_ARGUMENT), in the same format as the other words.
• The review section of the site: findings/documents.json has the three documents with their verdicts, so the dashboard, the scorecard, the heatmap and the sidebar show them. A new summary, adversarial-review-pass1/00-SUMMARY-snes-video-guides.md, consolidates the three reviews. The sync script gives the two video summaries different labels in the sidebar. The status list and the log of review/README.md have the new rows.
• A new open question (O9): which PC speed does the lander target? The frame budget of a 4.77 MHz 8086 is about 68,000 cycles, and a full-screen copy costs about 544,000.
• Checks: the splitter ran again, the site built, and the STE lint stayed under 2.5 for each file that changed.

Where things stand at the end of the session

• Live: https://kairin.github.io/000-111-learn/. GitHub Actions builds it from site/.
• Decided: D25 (the SNES review) and D26 (the lessons live on the segment pages).
• Open: O9, the PC speed of the lander.
• Next: the game design (phase 5), then tests 4 to 6 (Phase 4d) inside the 12-week course.