Colt Skill

COLT Corpus-Grounded Conference Skill

Generated: 2026-07-18T02:43:25Z

Inputs

• Target manuscript abstract, claim list, experiment plan, or draft sections.
• Target cycle year and any official CFP, author guide, template, rebuttal, artifact, or camera-ready links the user provides.
• Optional submission package paths, review text, code/data repository notes, and presentation type.

Read Before Acting

Read these local references before making venue-specific claims:

1. references/corpus-summary.md for corpus coverage, years, tracks, and sample size.
2. references/evidence-patterns.md for accepted-paper patterns with supporting paper IDs.
3. references/paper-index.md when you need to inspect which sampled papers support a pattern.

This skill is corpus-grounded only. It does not contain verified current-cycle official policy.

Workflow

1. Identify the user's stage: venue fit, structure planning, experiment audit, anonymity/ethics check, submission preflight, author response, artifact, camera-ready, or presentation.
2. Ask for the target cycle year and official policy sources if the task depends on deadlines, page limits, templates, anonymity, AI policy, rebuttal format, artifact rules, registration, or fees.
3. Use corpus patterns only for research-culture guidance: section planning, claim framing, evidence expectations, baseline pressure, ablation and robustness gaps, limitations, reproducibility, and presentation story.
4. Separate every output claim into [policy], [pattern], or [user-provided].
5. If official policy is missing for a hard submission fact, mark the status blocked or unverified; do not infer current rules from corpus patterns.
6. Return a concise handoff that names the next action and the evidence still needed.

Hard Gates

• Current official deadline, page limit, template, anonymity, AI-use, rebuttal, artifact, registration, or fee facts are missing but the user asks whether the submission is ready.
• The user asks for an acceptance probability or asks the agent to treat corpus frequency as an official rule.
• A draft claims experimental superiority but provides no datasets, metrics, baselines, or source of results.
• A response or rebuttal would require inventing new experimental numbers not provided by the user.

Evidence Rules

• [policy] must come from a current official source supplied by the user or a verified project cycle reference.
• [pattern] must cite references/evidence-patterns.md with sample size, years, and paper IDs.
• [user-provided] facts must be labeled when they are not independently verified.
• Accepted-paper patterns describe observed practice, not venue requirements.

Output Contract

Use this structure:

[status] ready | at_risk | blocked | unverified | not_applicable
[venue] COLT
[stage] venue-fit | structure | experiments | policy-check | submission | response | artifact | camera-ready | presentation
[policy state] verified | partial | missing | not-needed
[corpus evidence] sample_n=... years=... paper_ids=...
[findings]
1. ...
[blockers]
1. ...
[next action] ...

Failure Paths

• If references are missing, return blocked and ask to regenerate colt-skill.
• If the user needs current-cycle policy and no official source is available, return unverified or blocked.
• If the manuscript domain does not match the venue's corpus patterns, return at_risk with adjacent-venue alternatives only when the user asks for routing.
• If paper evidence is too sparse for a requested subarea, state the sample limitation instead of generalizing.