Chrome side-panel prompt catalog¶

A catalog of prompts that work reliably with the heim Chrome side panel, verified against a local gemma-4-12B model. Each prompt is written to be model-friendly for a ~12B model: explicit output format and "Output ONLY ..." guards.

The status table near the bottom is auto-generated from a real verification run (see How verification works) so the docs never drift from reality.

How to use¶

Install and open the side panel (see the Chrome side panel guide).
Point it at a running heim serve (Settings -> heim base URL).
Turn on Page context (the pill in the composer, or the Settings checkbox). This attaches the page URL, title, and your current text selection to the next message.
For whole-page tasks, also turn on Page text (the second pill / the "Include full page text" checkbox). This attaches the page's visible text, collapsed and truncated to 8000 characters. It is a sub-option of Page context.
For selection-grounded tasks (Q&A, rewrite, translate), select the paragraph on the page first and leave Page text off - only your selection is sent.
Paste a prompt from below and send.

The context is prepended to your message as a labeled block:

Page URL: https://news.ycombinator.com/
Page title: Hacker News
Page text (truncated):
<the page's visible text>

or, for a selection:

Page URL: https://en.wikipedia.org/wiki/SQLite
Page title: SQLite - Wikipedia
Selected text:
<your selection>

Prompts by site¶

Copy a prompt, make sure the right toggles are on (noted per group), and send.

Hacker News front page (`news.ycombinator.com`)¶

Page text on, no selection.

Wikipedia: SQLite (`en.wikipedia.org/wiki/SQLite`)¶

Article summary - page text on:

The page text above is a Wikipedia article about SQLite. Summarize it in exactly 5 bullet points. Each bullet is one sentence. Output ONLY the 5 bullets.

Extract years mentioned - page text on:

From the page text above, list every distinct year (a four-digit number like 2004) that is mentioned, as a plain list with one year per line, sorted ascending. Output ONLY the list.

Selection-grounded (select the lead paragraph first; Page text off):

Explain the selected text simply, as if to someone new to databases, in 2 to 3 sentences. Output ONLY the explanation.

Using ONLY the selected text, answer in one sentence: what kind of database engine is SQLite? If the answer is not in the selected text, reply with exactly: not in the selection

Honesty case (the answer is absent from the selection - the model should refuse):

Using ONLY the selected text, answer in one sentence: who is the current CEO of Google? If the answer is not in the selected text, reply with exactly: not in the selection

Rewrite for a child:

Rewrite the selected text so a 10-year-old can understand it. Keep it under 80 words. Output ONLY the rewritten text.

Translate to Norwegian:

Translate the selected text into Norwegian (bokmål). Output ONLY the Norwegian translation.

GitHub: fastapi/fastapi (`github.com/fastapi/fastapi`)¶

Page text on.

Repo metadata as JSON:

The page text above is a GitHub repository page. Output a strict JSON object with keys: name (string, owner/repo), description (string), primary_language (string, your best guess). If a value is unknown use an empty string. Output ONLY the JSON object.

Two-bullet summary:

The page text above is a GitHub repository page. In exactly 2 bullet points, state (1) what the project is and (2) its primary programming language. Output ONLY the 2 bullets.

MDN: Array.prototype.map() (`developer.mozilla.org`)¶

Page text on.

Signature extraction:

The page text above is the MDN reference for Array.prototype.map(). Extract the call signatures shown in the Syntax section, one per line, exactly as written. Output ONLY the signature lines.

Parameters as JSON:

The page text above is the MDN reference for Array.prototype.map(). List its parameters as a strict JSON array of objects with keys: name (string), description (string, one short sentence). Output ONLY the JSON array.

arXiv abstract (`arxiv.org/abs/1706.03762`)¶

One-sentence summary - page text on:

The page text above is an arXiv abstract page. Summarize the paper's abstract in ONE sentence of at most 40 words. Output ONLY that sentence.

Selection-grounded (select the abstract block first; Page text off):

Explain the selected text in plain language for a non-expert, in 3 sentences. Avoid jargon. Output ONLY the explanation.

Using ONLY the selected text, answer in one sentence: what network architecture does the paper propose? If the answer is not in the selected text, reply with exactly: not in the selection

Extract numbers and their meaning:

From the selected text, extract every number and what it refers to, as a list with one item per line in the form: number - meaning. If there are no numbers, reply with exactly: none. Output ONLY the list.

Stack Overflow (a canonical question)¶

Page text on.

The page text above is a Stack Overflow page. In 2 to 4 bullet points, summarize the question and the key point of the top answer. Output ONLY the bullets.

Python docs: json module (`docs.python.org/3/library/json.html`)¶

Page text on.

Module summary:

The page text above is the Python documentation for the json module. In exactly 3 bullet points, summarize what this module is for. Each bullet is one sentence. Output ONLY the 3 bullets.

Functions as JSON:

The page text above is the Python documentation for the json module. List four top-level functions it documents as a strict JSON array of strings (e.g. "json.dumps"). Output ONLY the JSON array.

Selection-grounded honesty case (select the intro paragraph; Page text off):

Using ONLY the selected text, answer in one sentence: what is the airspeed velocity of an unladen swallow? If the answer is not in the selected text, reply with exactly: not in the selection

DHIS2 play demo (`play.im.dhis2.org/dev-2-42`)¶

Context-only (no login). Page context on; the URL and title in the context block are enough.

Extract the instance URL:

Using the page context above, output ONLY the base URL of this DHIS2 instance and nothing else.

Describe the page:

Using the page context above, in 1 to 2 sentences describe what this page is (name the system and whether it looks like a login page or a signed-in app). Output ONLY the description.

Verified results¶

Auto-generated from a real harness run. Model: lmstudio-community/gemma-4-12B-it-QAT-GGUF. 27/28 verified. Last run: 2026-07-11T00:19:10.543Z.

By category

Category	Verified
extraction	7/8
summarization	7/7
explanation	2/2
selection-qa	4/4
transformation	4/4
reasoning	3/3

Per prompt

Prompt	Site	Category	Mode	Status	Latency	Check
`hn-json-top10`	hn	extraction	page-text	verified	54.4s	json-array: array of 10 items
`hn-markdown-table`	hn	extraction	page-text	unreliable (12B)	84.6s	markdown-table: only 7 data rows (need >= 8)
`hn-csv`	hn	extraction	page-text	verified	118.5s	csv: 11 csv rows
`github-metadata-json`	github	extraction	page-text	verified	21.2s	json-object: object with keys name,description,primary_language
`mdn-callback-args`	mdn	extraction	page-text	verified	18.3s	contains: contains all of: element, index, array
`mdn-params-json`	mdn	extraction	page-text	verified	15.1s	json-array: array of 2 items
`pydocs-funcs-json`	pydocs	extraction	page-text	verified	60.2s	json-string-array: array of 4 strings
`dhis2-url-extract`	dhis2	extraction	page-text	verified	26.8s	regex: matched /play.im.dhis2.org/i
`hn-themes`	hn	summarization	page-text	verified	53.7s	bullets: 3 bullets
`wiki-summary-bullets`	wikipedia	summarization	page-text	verified	27.9s	bullets: 5 bullets
`arxiv-onesentence`	arxiv	summarization	page-text	verified	24.3s	length: 32 words
`pydocs-summary`	pydocs	summarization	page-text	verified	21.6s	bullets: 3 bullets
`so-summary`	stackoverflow	summarization	page-text	verified	24.9s	bullets: 3 bullets
`github-summary`	github	summarization	page-text	verified	16.1s	bullets: 2 bullets
`wiki-explain-selection`	wikipedia	explanation	selection	verified	10.0s	length: 50 words
`arxiv-explain-selection`	arxiv	explanation	selection	verified	17.1s	length: 65 words
`wiki-selection-qa`	wikipedia	selection-qa	selection	verified	14.8s	regex: matched /serverless\|self-contained\|embedded\|library\|C language\|written in C
`wiki-selection-honesty`	wikipedia	selection-qa	selection	verified	6.3s	honesty: correctly refused (not in the selection)
`arxiv-selection-qa`	arxiv	selection-qa	selection	verified	8.4s	contains: contains all of: transformer
`pydocs-selection-honesty`	pydocs	selection-qa	selection	verified	8.6s	honesty: correctly refused (not in the selection)
`wiki-child-rewrite`	wikipedia	transformation	selection	verified	16.6s	length: 59 words
`wiki-translate-no`	wikipedia	transformation	selection	verified	34.5s	regex: matched /\b(og\|er\|en\|som\|det\|til\|av\|for\|med\|ikke)\b/i
`arxiv-extract-numbers`	arxiv	transformation	selection	verified	71.2s	regex: matched /\d.-\|-\s\d\|none/i
`wiki-extract-dates`	wikipedia	transformation	page-text	verified	33.1s	regex: matched /(19\|20)\d\d/
`hn-most-points`	hn	reasoning	page-text	verified	14.0s	regex: matched /\d+\s*points/i
`hn-rank-by-points`	hn	reasoning	page-text	verified	22.7s	regex: matched /1.[\s\S]\d+\spoints/i
`hn-points-delta`	hn	reasoning	page-text	verified	8.0s	regex: matched /difference:?\s*-?\d+/i
`dhis2-context-describe`	dhis2	summarization	page-text	verified	7.1s	contains: contains all of: dhis2

Reliability notes

hn-markdown-table: Borderline at 12B: usually emits the full 10-row table, but sometimes stops after 7 rows even at temperature 0.1 (observed 10/10 and 7/10 across identical runs). The JSON variant (hn-json-top10) is the reliable form; use this one when you want copy-pasteable Markdown and do not mind an occasional short table.
mdn-callback-args: Replaces a signature-extraction prompt: code blocks (the literal map(callbackFn) syntax lines) do not survive page-text capture on MDN, so signature extraction is unanswerable from the attached context - the model deliberates endlessly over missing data. Prose-rendered facts extract reliably.

How verification works¶

The prompts above are not aspirational - they are checked mechanically against a real model by the harness in extension/e2e/prompts/:

Capture (capture.ts): plain Playwright visits each site once and records {url, title, page text (<= 8000 chars, whitespace-collapsed), selection}, mirroring exactly what the extension's lib/pageContext.ts collects. Captures are cached to results/captures.json so reruns can use --no-capture.
Replay (replay.ts + context.ts): the harness reuses the extension's own formatPageContext to build the context block (a format-parity.spec.ts unit test guards that they stay identical), prepends it to the prompt, and POSTs to /api/chat with use_tools=false and a low temperature against a locked-model heim serve.
Assert (checks.ts): each prompt has a mechanical check - JSON parse + shape + item count, Markdown table header/row counts, CSV column counts, keyword/regex presence, the honesty refusal, or summary length/bullet bounds.
Doc regen (regen-doc.ts): the results table above is rewritten from results/results.json, so it always reflects the last real run.

Rerun¶

make extension-prompts

This captures the sites, starts a heim serve locked to gemma-4-12B, replays every prompt, writes extension/e2e/prompts/results/ (git-ignored: results.json + per-prompt raw outputs under outputs/), and regenerates the table above. Use HEIM_PROMPT_BASE_URL=http://127.0.0.1:PORT bun run prompts --no-capture (from extension/) to replay against an already-running server without re-capturing.

The end-to-end last mile - a verified prompt driven through the real side panel against a real heim - is covered by e2e/prompts/ui.spec.ts (bun run prompts:ui).

Chrome side-panel prompt catalog¶

How to use¶

Prompts by site¶

Hacker News front page (news.ycombinator.com)¶

Wikipedia: SQLite (en.wikipedia.org/wiki/SQLite)¶

GitHub: fastapi/fastapi (github.com/fastapi/fastapi)¶

MDN: Array.prototype.map() (developer.mozilla.org)¶

arXiv abstract (arxiv.org/abs/1706.03762)¶

Stack Overflow (a canonical question)¶

Python docs: json module (docs.python.org/3/library/json.html)¶

DHIS2 play demo (play.im.dhis2.org/dev-2-42)¶

Verified results¶

How verification works¶

Rerun¶

Hacker News front page (`news.ycombinator.com`)¶

Wikipedia: SQLite (`en.wikipedia.org/wiki/SQLite`)¶

GitHub: fastapi/fastapi (`github.com/fastapi/fastapi`)¶

MDN: Array.prototype.map() (`developer.mozilla.org`)¶

arXiv abstract (`arxiv.org/abs/1706.03762`)¶

Python docs: json module (`docs.python.org/3/library/json.html`)¶

DHIS2 play demo (`play.im.dhis2.org/dev-2-42`)¶