Behind the scenes

One keynote, built three ways

Same brief, same fact sheet, same open-source Teaching Agent – only the AI engine changed. Here is what each route cost and what it gave.

V1 · Offline

Local open model

qwen3.8:27b on the university AI server (GX10)

46 min, no cloud tokens. Data stays on campus.

Open V1 in LiaScript ↗

V2 · Cloud only

Commercial cloud model

Claude Opus 5.5 builds the whole keynote

4 min, everything in the cloud.

Open V2 in LiaScript ↗

V3 · Offline + cloud

Draft offline, polish online

V1 draft, then one Claude editing session

46 min offline + 3 min cloud. Basis of the web keynote.

Open V3 in LiaScript ↗

How an agent builds a keynote

The language model never touches the computer. It writes requests as text – read_file, write_file – and the agent program checks them against the rules and runs them. The offline run took 23 such steps:

[1]    35s  read_file journal.md; read_file inputs/rahmen.md; read_file inputs/quelle_keynote.md
[2]    60s  read_file specs/tasks/build-session.md
 …
[7]   689s  write_file materials/01-giz-keynote/README.md      ← first full draft
[9]  1545s  write_file materials/01-giz-keynote/README.md      ← revision after self-review
 …
[20] 2611s  write_file agent_report.md
[23] 2812s  All self-checks pass …

Full log: run.log · task: T1.md · agent report: agent_report.md

Cloud effort

Claude input-token equivalents (fresh input and cache write × 1, cache read × 0.1, output × 5). Lower is cheaper.

Runtime

Minutes from start to finished deck (without human review).

Measured numbers

	V1 Offline (GX10)	V2 Cloud only	V3 Offline + improve
Model	qwen3.8:27b, local	Claude Opus 5.5	qwen3.8:27b → Claude Opus 5.5
Runtime	46 min (23 steps)	4.2 min (38 turns)	46 min + 2.9 min (12 turns)
Local model tokens (in / out)	947 k / 35 k	–	947 k / 35 k
Claude fresh input + cache write	0	162 k	41 k
Claude cache reads	0	1.19 M	0.32 M
Claude output (incl. thinking)	0	35.8 k (13.9 k)	18.5 k (6.9 k)
Claude input-token equivalents	0	≈ 459 k	≈ 165 k (–64 %)
Claude list price	0 €	≈ 2.25 USD	≈ 0.76 USD
Syntax errors (lint)	0	0	0
Slides	17	18	18

Not included and the same for all three: preparing brief and fact sheet (≈ 150 k equivalents, estimated) and the human review before the talk.

Pros and cons

	Pros	Cons
V1 Offline	no cloud cost data never leaves the university clean syntax follows brief and fact sheet faithfully	slow (46 min) copies the fact sheet almost word for word jargon kept, crowded slides its own self-check said “all fine” – and gave a wrong date
V2 Cloud only	best keynote straight away plain words, clear tables self-review caught jargon ready in 4 min	highest cloud cost all material goes to a cloud provider still needs the speaker's review
V3 Offline + improve	quality on the level of V2, with a thread of people (Amina, Grace, Joseph, Sara) ≈ 64 % fewer cloud tokens than V2 offline part can run overnight	two steps, ≈ 50 min in total only about 55 % of the offline draft survived the draft still goes to the cloud

What this means for GIZ projects

Local open models are good enough for first drafts
– and keep sensitive data in the country.

Quality comes from the improvement step
by a stronger model and by people.

Mix and match
offline for volume and privacy, cloud for polish – at about a third of the cloud cost.

Never trust the AI's own quality report
check it yourself.

The local model does the heavy, routine part. A stronger system – and finally a human – checks, shortens and makes it human.

Source files, logs and all three versions: github.com/OVGU-VET-TechEd/GIZ_impulse_lecture