Sundae Bar Logo
Past Challenge
Internal Communications Writing Skill
SBC11

Internal Communications Writing Skill

Overview

The core objective: given free-form project input, the agent must reliably produce a team Slack update, a cross-functional stakeholder email, and an exec-facing Notion/Confluence status section, each correctly formatted, toned, and scoped for its audience.

Your submission is a single SKILL.md file that teaches a fixed base agent to transform raw project context — ticket dumps, standup notes, sprint recaps, launch readouts — into three polished, audience-calibrated internal communication artifacts in a single response.

Strong skills teach the agent a reasoning procedure: classify the audience tier first, check for missing critical facts before drafting, extract blockers from narrative prose, triage what belongs at each level, and calibrate tone and length independently per channel. Weak skills hand the agent a fill-in-the-blank template — these collapse the moment input arrives in an unfamiliar shape, omits an owner, or contains politically sensitive blockers that should not surface at the exec level.

The skill must also teach the agent when NOT to apply the three-artifact structure: customer-facing communications, incident postmortems, and one-off conversational Slack messages are out of scope and should be declined gracefully rather than forced into the format.

A skill built to memorise specific scenarios will not hold up as challenges evolve.

This challenge is also a step toward a broader skills marketplace.

You submit a SKILL.md file to the challenge. The validator attaches it to our canonical base agent, runs it against 10 sampled scenarios, and returns a scored breakdown. Top submissions earn Alpha emissions daily and on displacement.

We are looking for skills that demonstrate good design and help the base agent behave like a capable practitioner in this domain.

Requirements (10 Tests)

Your skill is tested across eight capability types, with 10 scenarios in total. Every submission is evaluated against all 10 scenarios

CategoryOverviewTest cases
Type A — Blocker Triage & Visibility Scoping Tests whether the agent correctly separates operational blockers (appropriate for the team Slack post) from risk-level blockers that require exec visibility, and withholds team-level details from the exec artifact. The skill must teach the agent to apply a deliberate triage decision — not simply copy blockers across all three artifacts — and to label each blocker with owner and deadline in a structurally separated block.1
Type B — Distinct Artifact Composition Tests whether the agent produces three genuinely independent artifacts from the same source material rather than progressively shortening or reformatting a single draft. The skill must teach the agent to re-derive structure, emphasis, and language for each channel independently, so that the exec section, the email, and the Slack post each reflect their audience's needs rather than echoing one another.1
Type C — Cross-Functional Partner Framing Tests whether the agent identifies when multiple cross-functional teams are affected and includes explicit per-partner 'what this means for you' framing in the stakeholder email. The skill must teach the agent to read the input for downstream dependencies, name the affected teams, and tailor impact statements for each — not write a single generic cross-functional paragraph.1
Type D — Per-Channel Tone Calibration (testCases: 1)Tests whether the agent applies the correct register to each channel: casual and emoji-appropriate for Slack, professional and low-jargon for the stakeholder email, and outcome-verb-led and crisp for the exec Notion section. The skill must teach tone as a deliberate calibration step tied to channel identity, not a stylistic preference applied inconsistently.1
Type E — Update Delta & Correction (testCases: 1)Tests whether the agent, when given context that implies a prior update exists, correctly surfaces what has changed, what has been corrected, and what is new — rather than restating the full status as if from scratch. The skill must teach the agent to recognise delta signals in the input and structure each artifact around what is different, not what is complete.1
Type F — RAG Status Assignment & Risk FramingTests whether the agent assigns a committed RAG (green/yellow/red) status to the exec artifact and supplies a concise justifying reason grounded in the input's timeline, dependency, or risk signals. The skill must teach the agent to treat RAG assignment as an analytical judgment — selecting the status that most accurately reflects project health — rather than defaulting to yellow as a safe hedge.1
Type G — Blocker Extraction from Narrative Prose Tests whether the agent can parse free-form narrative input to identify blockers and pending decisions that are embedded in prose rather than listed explicitly, and render them as a structurally separated labeled block with owner and deadline. The skill must teach the agent to treat extraction as an active reading step, not a search for the word 'blocker'.1
Type H — Exec Artifact with Technical RiskTests whether the agent can translate technical risk signals — architecture constraints, integration edge cases, data quality issues — into exec-appropriate language that conveys business impact without jargon, paired with a justified RAG status. The skill must teach the agent to reframe technical detail as outcome risk, not omit it or reproduce it verbatim in the exec section.1

Surface details such as project names, team names, owner names, timelines, and specific metrics rotate between scenario variants while the underlying communication capability being tested remains the same.

Datasets

Available Datasets

Your skill attaches to a standardised base agent provided by sundae_bar. The base agent is the same for every submission. It is lean by design: generic memory blocks, standard tools, no customer-specific knowledge. Its performance on the evaluation scenarios is driven almost entirely by the skill you give it.

You do not submit or modify the base agent. Its behaviour is fixed across all submissions so that scoring reflects skill quality, not agent configuration.

Tools available to the base agent: the load_skill tool (which loads your SKILL.md content when the agent decides to use it), plus standard built-ins including memory and conversation search. Available tools: load_skill.

Evaluation Criteria

Each scenario is graded by LLM judges. Final scores combine four weighted criteria; integrity gates run before scoring, and a submission that fails a gate scores 0 regardless of output quality.

CriteriaWeight

Skill use

An LLM judge evaluates whether the agent's output actually followed the procedures your skill teaches, rather than freelancing or pretending to use the skill.

30 %

Scenario Quality

An LLM judge evaluates how well each output meets the per-type quality bar, independent of whether your skill was followed. A reference good output anchors the judge's sense of the quality bar.

30 %

Rubric compliance

An LLM judge scores each output against a structured rubric for this challenge's communication artifacts — audience calibration, structure, and actionability. The rubric itself is not published: write for the reader each artifact serves, not for a checklist.

30 %

Output Novelty

Outputs are compared across scenarios. Near-duplicate, template-style responses that reuse the same content regardless of scenario specifics are penalized.

10 %

Scenarios are scored individually, with deterministic penalties applied for moves that damage output quality — such as inventing missing facts instead of asking, copying blockers indiscriminately across audience tiers, hedging a RAG status without justification, or producing artifacts that are structurally indistinct from one another. Individual scenario scores aggregate to capability-type averages, which roll up to a single final challenge score displayed on the leaderboard.

Submission

How to submit

1
Verify miner registration on Subnet 121 and connect it to your sundae_bar account

Go to your sundae_bar account settings under wallet & transactions and verify you have your wallet id verified. If your accounts hotkey is registered to SN121 you will see it below with your miner UID. Otherwise you can use our prompts to register or register on the command line.

2
Review the active challenge brief

Make sure you review all requirements before creating your skill.md file

3
Submit your skill to the active challenge

On this page there is a submission button. If active and a registered miner to SN121 you can submit your contribution.
Note: we have daily user limits and total rolling limits.

4
Validator runs the evaluation

First your submission will go through anti-gaming gates. If passed the validator will run the evaluation suite on your submission to return your score.

Constraints

Constraints details coming soon...

Rewards

Emissions accrue on-chain to whichever submission currently holds the top of the leaderboard. There is no fixed per-submission payout: the current leader earns for as long as it stays first.

Time at the top takes all. A strong skill submitted early starts accruing immediately and keeps earning until another submission displaces it. Once you are displaced, accrual stops until you reclaim the top.