# Week 6: Evidence of practice > Read the transcript and the git log closely, reconstruct one consequential decision, and write what you learned into the harness, which the next session loads # Welcome --- # Week 6 ## Evidence of practice --- ## what changed this week? Debian voted on [LLM usage in Debian](https://www.debian.org/vote/2026/vote_002) "responsible use of generative AI" won: disclosure of AI assistance is encouraged, not required also lots of model releases --- ## today's beats the transcript and the git log are the only account of the work that doesn't rely on your memory the "compound" step: write what you learned into the harness CI runs the same checks on a machine that isn't yours --- ## demo: can it reconstruct fourteen years of my tooling from the git log alone? then we leave it running and talk about what the work leaves behind --- # the practice leaves evidence --- ## three records of the work **the transcript**: every prompt, response and tool call **the git log**: every diff and every commit **the CI runs**: every pass and fail the transcript is what agentic work adds, between the commits, on disk (a running system that can examine itself is what programming languages call **introspection**) --- ## the logs are just files `~/.claude/projects//.jsonl`, one JSON object per line (thousands in a long session); one record, trimmed to five fields: ```text {"type":"assistant", "timestamp":"2026-09-22T03:25:30.649Z", "cwd":"/home/ben/projects/comp4020", "gitBranch":"HEAD", "message":{"content":[ {"type":"tool_use","name":"Bash", "input":{"command":"ls"}}]}} ``` --- ## querying the logs with `jq` ```sh jq -r '.message.content[]? | select(.type=="tool_use") | .name' \ .jsonl | sort | uniq -c | sort -rn | head -4 ``` ```text 170 Bash 30 Edit 9 Read 4 Write ``` (which tools did the agent call? one of my sessions, top four; your agent can write these for you) --- ## writing the learnings back **compound engineering** ([Kieran Klaassen, Every](https://every.to/c/compounding-engineering)): plan, work, review, **compound** — write the learnings back into persistent context, so the next cycle starts further along you've been doing this in `reflections/` and your `CLAUDE.md` since week 2 --- # **words**; what do they even mean? --- ## hoard things you know how to do work a trick out once, keep the documented example, and it becomes an input to every later project ref: [Agentic Engineering Patterns](https://simonwillison.net/guides/agentic-engineering-patterns/hoard-things-you-know-how-to-do/) what have I worked out once and kept, in my public dotfiles history? --- ## one line, thirteen months ago ```diff - use Australian English spelling - don't overuse lists (either bulleted/numbered or even just enumerating things in prose) + - do not ever use exclamation marks ``` a correction in the chat is gone next session; committed to `claude/CLAUDE.md` in August 2025, it is still in the file today softened once on the way through — "don't overuse lists or exclamation marks" --- ## the log says what changed — does it say why? the timeline falls out of the filenames and the dates alone the reasons, where they survived, are in the commit messages watch for the seam: where does the evidence stop and the story start? --- ## and what does it reckon I'll pick up next? a reading of the past is cheap; a prediction can be checked so we'll have it commit the call, and check it next semester --- ## the agent writes its own `CLAUDE.md` is what **you** wrote; **auto memory** is notes Claude writes without being asked — your corrections, your preferences, facts about the project `~/.claude/projects//memory/`, one markdown file per fact, with a `MEMORY.md` index read at the start of every session on by default, per repo, machine-local, and yours to read, edit or delete: `/memory` ref: [How Claude remembers your project](https://code.claude.com/docs/en/memory) --- ## a memory file, as Claude wrote it (trimmed) ```md --- name: uv-owns-the-venv description: "Fix broken venvs with uv commands, never a manual …" metadata: type: feedback --- Let uv manage virtualenvs end to end. When a venv is broken … reach for `uv sync --reinstall` …, not `rm -rf .venv`. ``` written after one correction; its one-line entry in `MEMORY.md` loads in every session in that project, and Claude reads the file when it looks relevant --- ## a file in every session's context Cisco Talos got a poisoned `MEMORY.md` onto a machine through an npm `postinstall` hook memory content was reaching the system prompt as though you had written it; Anthropic stopped that in v2.1.50 ref: [Cisco Talos](https://blogs.cisco.com/ai/identifying-and-remediating-a-persistent-memory-compromise-in-claude-code) --- ## the context cost of a rule every rule you add is context you pay for in **every** later session, and it goes stale > it had silted up into documentation: superseded decisions, dated changelog > entries, war stories my commit message, cutting a `CLAUDE.md` from 669 lines to 250 (the rule that fell out went straight back in) --- the chat lasts one session the **harness** is loaded into every one --- ## how much of this is true? the number that circulates is [300–700% faster](https://www.vincirufus.com/en/posts/compound-engineering/), sourced to a bracketed note citing nothing Klaassen's own claim is one hedged anecdote: [two weeks, "maybe", down to a few hours](https://every.to/podcast/transcript-how-two-engineers-ship-like-a-team-of-15-with-ai-agents) faster at _what_, measured _how_, against _what baseline_? --- ## Ben uses… a timer that ships the logs home every session from every machine, synced into one sqlite table ```text > select host, count(*) from sessions group by host; daysy |7625 weddle|8279 ``` because the logs are just **files** --- ## yes, and… how does it get pushed further? weird ideas welcome --- # Web --- ## continuous integration a series of checks that runs on every push, on a machine that isn't yours GitHub Actions in our case, configured by `.github/workflows/checks.yml` in your repo; the idea is [decades older](https://comp.anu.edu.au/courses/comp4020-agentic-coding-studio/topics/github-actions-basics/) --- ## one push, one run --- # which box does what? --- ## the same checks at two distances **local**: `pnpm check` in seconds, while the agent still holds the context that produced the mistake **the runner**: the same command on a clean machine, which tests whether the code works anywhere other than the machine it was written on if CI gives different results to local dev, then... --- ## runner-only failures ```text ✓ check in 29s ✗ deploy in 19s ✗ Run actions/configure-pages@v6 ##[error] Get Pages site failed. Please verify that the repository has Pages enabled and configured to build using GitHub Actions ``` the demo repo's first day, for real: every check green, and the deploy still failed --- ## reading a red run ```text $ gh run view 37084179269 X deploy in 2m32s (ID 111090928068) ✓ Verify the deployed site is online X Verify the LiveView socket upgrades $ gh run view 37084179269 --log-failed the LiveView socket did not upgrade (HTTP 400). If it is 403, the app thinks it is serving some other host: check PHX_HOST and FLY_APP_NAME … ##[error]Process completed with exit code 1. ``` the failing step's output gives the agent somewhere to look (lecture-demos, 3 October; trimmed) --- ## always-red checks week 3's fourth property was **trustworthy**: you learn to scroll past a check that's always red, and then the real failure goes unnoticed too e.g. the crit template's link check skips outbound links, because someone else's rate limiter or outage would turn every build red (see the comment in `checks.yml`) --- ## green when the sweep runs [the crit sweep](https://comp.anu.edu.au/courses/comp4020-agentic-coding-studio/topics/studio-crit-model/#in-crit-prototype-marks) starts fifteen minutes after your cutoff: pushed, live at its URL, reflection in the folder — and then it reads your checks **green**: the full shipping mark **red, still running, or no checks**: half those fifteen minutes are settling time — push early enough that the run finishes --- ## nothing is marked until you `/ship` it flips the repo public, turns Pages on, triggers the deploy, then checks the URL actually serves `/preflight` first — going public is the one **irreversible** thing you do in this course (C6 is a retro: no prototype and no checks, so that half is just `PROCESS.md` in the repo by the cutoff) --- ## over the break: fifteen minutes of fly from week 7 [CI deploys your repos](https://comp.anu.edu.au/courses/comp4020-agentic-coding-studio/topics/github-actions-basics/) on every push, and you never handle a deploy credential but `flyctl` is how you read your running app's logs and status: [onboarding step 7](https://comp.anu.edu.au/courses/comp4020-agentic-coding-studio/topics/onboard/#7-later-flyio-for-the-full-stack-half), no account, no billing of your own your tutor checks it works in c6 (no marks) --- ## yes, and… how does it get pushed further? weird ideas welcome --- ## your assignment 2 repo is **live** `/comp4020:start` clones it, like every other deliverable the [brief](https://comp.anu.edu.au/courses/comp4020-agentic-coding-studio/assessments/assignment-2/) is final now — and the one thing that moved since last week is this morning's lesson `PROCESS.md` has a spine: what you decided a good course looks like, which of those decisions you encoded in `CLAUDE.md` or `spec/`, and which you left out --- course ~~you'd like to take~~ you wish existed --- ## update the **course plugin** `claude plugin marketplace update comp4020`, then `claude plugin update comp4020@comp4020`, then restart Claude Code **do it before the break**: assignment 2 ships on whatever copy you're holding --- ## yes, and… how does it get pushed further? weird ideas welcome did the steer show up in the result? --- ## C6: the assignment 2 retro > no fresh provocation: the crit reflects on assignment 2 pick the specific change that made the course click and show it as a before/after — the breakthrough comes from the `PROCESS.md` you already submitted the [crit page](https://comp.anu.edu.au/courses/comp4020-agentic-coding-studio/crits/06-a2-retro/) has the spec --- ## deadline radar **noon, Monday 21 September**: assignment 2 — the first day back, so it's closer than it looks **Monday/Wednesday, week 7**: C6, with your `PROCESS.md` in the repo by the cutoff **the break**: two weeks, and fifteen minutes of it is installing `flyctl` --- ## tl;dr the transcript, the git log and the CI runs are the only account of the work that isn't recollection write the correction into the harness rather than the chat, then prune it: every rule is context you pay for in every session CI runs the same checks on a machine that isn't yours — be green when the sweep runs, and `/ship` before the cutoff --- ## ask your agent ask until you can explain these back: - **the compound step**: what should this session's corrections write back into your `CLAUDE.md`? - **auto memory**: open `/memory` and read what Claude has written about you - **CI**: which failures could only happen on the runner? - **fly**: install `flyctl` over the break, and confirm it runs --- # See you in the studio Source: \ Licence: [CC-BY-NC-SA-4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/)\ Cite as: Swift, B. (2026). _COMP4020/COMP8020 Agentic Coding Studio_ [Course materials]. ANU School of Computing. https://doi.org/10.5281/zenodo.21395760