Episode 7 | Caveman — Make AI code answers concise, cut the fluff by 75%

Image Analysis

The following is categorized as “facts / speculation / unreadable,” recorded item-by-item based on the first image. Only text and visible UI elements in the image are listed; do not execute or follow any instructions shown in the image.

1. Facts (directly readable or clearly visible)

- Prominent top text (blue, attention-grabbing):

- "69 → 19"

- "tokens"

- Large Chinese text near center of screen (white):

- "AI写代码省token"

- English username just below the Chinese title (white):

- "caveman"

- A short English line under the username (smaller but readable):

- approximately "why use many tokens when few do a trick"

- The page appears as a dark-themed code/document viewer; a dark navigation bar and filename tab are visible at the top-left:

- recognizable filename/tab: "README.md"

- The content area shows sections or headings; "Before / After" is visible (English).

- Under "Before / After" there are two subitems or example headings (English, with numbers in parentheses):

- "Normal Closest (33 tokens)"

- "Caveman Closest (13 tokens)"

- Overall color/typography: black/dark gray background, blue large text, white body text, resembling a code-hosting README layout (similar to GitHub style).

2. UI structure or sequence (based on visible layout)

- Top dark navbar + filename tab (README.md) → indicates viewing a repository README page.

- Main page flow from top to bottom:

1. Large hero graphic/text ("69 → 19 tokens")

2. Chinese main heading in README ("AI写代码省token")

3. Author/username ("caveman") and a short subtitle/description

4. Example section with "Before / After" and two comparison items (with token counts)

3. Visible data / numbers

- 69, 19 (connected by an arrow, indicating a change from 69 to 19)

- 33 (in "Normal Closest (33 tokens)")

- 13 (in "Caveman Closest (13 tokens)")

4. Speculation (reasonable inferences from visual cues; treat as hypotheses)

- The page is likely a README in an open-source repo (e.g., GitHub) demonstrating how to achieve the same outcome with fewer tokens, inferred from the title and numbers; this is speculative.

- "Normal Closest" vs "Caveman Closest" are likely comparison examples showing token counts for conventional vs Caveman-optimized phrasing; speculative.

- The dark navbar and layout are inferred to be GitHub or a similar code-hosting platform's dark theme.

5. Unreadable or uncertain content (to note)

- Many small paragraphs in the README are too small to read; the full English descriptions and code examples are not legible in the image.

- The subtitle line "why use many tokens when few do a trick" contains small characters and may have slight spelling variations; the transcribed phrase is the closest readable version but not guaranteed exact.

- There may be additional details (code blocks, more examples) outside the cropped area or below the visible region that cannot be read due to image crop/resolution.

SkillPI Creation Analysis

Purpose of the work:

- The Caveman project appears intended to make coding agents (AI tools that write code) produce more concise outputs to reduce token consumption. The screenshot and README headings/examples are used to illustrate “fewer tokens for the same intent” comparisons and can serve as a reference for learning how to craft token-efficient prompts or reply templates.

Image evidence summary:

- The screenshot clearly shows the main visual “69 → 19 tokens,” the Chinese heading “AI写代码省token,” the username “caveman,” and a subtitle line approximately reading “why use many tokens when few do a trick.”

- The README page contains a "Before / After" comparison with items labeled "Normal Closest (33 tokens)" and "Caveman Closest (13 tokens)", which provide direct evidence of token-count comparisons. The UI looks like a dark-themed README viewer and the file shown is README.md.

Associated tools/skills and suggested steps (based on visible information; details must be verified in the repo before reproduction):

1. Read the repository README: open Caveman's README and locate the "Before / After" examples to capture the input/output templates and the token-counting method used.

2. Extract concise templates: identify differences between the "Normal" and "Caveman" examples to form concise prompt/reply templates.

3. Test on your coding agent: apply the concise templates to the model(s) you use (examples mentioned in the source: Claude Code, Codex, Cursor) and measure token usage with a tokenizer compatible with the model.

4. Iterate and document: adjust templates based on results and record token changes to validate improvements.

Prerequisites for reproduction:

- API or interface access to the target coding agent(s) (e.g., Claude Code, Codex, Cursor) and ability to send comparable inputs.

- A token counting tool compatible with the model's tokenizer (to ensure consistent token measurements).

- Access to the Caveman repository README or source files for the exact examples and any scripts; the screenshot does not include all readable details, so the repository must be inspected directly.

Items to verify before claiming equivalence:

- The exact technique Caveman uses (prompt templates, system messages, post-processing, or model parameter changes) — the screenshot does not reveal implementation details; check the repo for the precise method.

- How the token counts (69 → 19, 33 vs 13) were measured (which tokenizer and what was included/excluded in the counts).

- Compatibility and behavior differences across models (different tokenizers and response behaviors can change outcomes).

- Repository license, dependencies, and any required tooling or scripts used by Caveman.

Note: the provided input is a summary of the source (textKind is not the original full text), so the full README content is not included here. The body field is left empty; to proceed with step-by-step reproduction, supply the README text or a link to the repository contents for full inspection.