homework
Skill10 starsChanged 4 months agoMITView on GitHub
Researches external technical evidence for a decision using source-quality assessment, claim-level citations, contradiction checks, evidence-sufficiency gates, and bounded validation. Use when the user invokes /glasses:homework, asks Homework to research a technical choice, requests deep or thesis-grade technical research, or when a Glasses plan or Pointer diagnosis depends on unsupported external claims about models, frameworks, hardware, providers, standards, or operational behavior.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- baekenough/oh-my-customcodetheir own version, first seen Nov 14, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- baekenough/oh-my-customcode★ 42same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repositoryOne other file it gives its agents.
- AGENTS.mdAGENTS.md
The file
Show its rules and text
We split this skill into the rules it gives an agent
26 rules. About 0 are common, 22 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈22
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 52
Always ask Q1 before proposing a research plan or performing research, even when the assignment appears clear.
- Distinctive: line 56
After the Q1 answer, ask further decision-changing clarification questions one at a time only until the research target is clear.
- Very specific: line 63
Prefer the platform's native question tool (AskUserQuestion in Claude Code or request_user_input in Codex) when available.
- Distinctive: line 68
Keep the entire question turn under 120 words: no preamble beyond one decision-relevant sentence, short options, and one recommendation …
- Distinctive: line 71
When no native question tool is available and a question has multiple viable answers, format it as:
- Distinctive: line 84
Do not prefix options with bullets or hyphens.
- Distinctive: line 90
Before external research, present a plan containing:
- Very specific: line 97
search waves, source/time/token/cost ceilings, and stopping conditions.
- Distinctive: line 100
Keep a straightforward research plan under 200 words.
- Distinctive: line 102
Do not repeat assignment context already confirmed in the conversation.
- Distinctive: line 106
Ask this approval through the platform's native question tool when available, with localized options for running or revising the plan.
- Distinctive: line 122
Search log: search wave, query or route, useful results, exclusions, and remaining gaps.
- Distinctive: line 141
Expansion: follow references, terminology, implementations, and adjacent questions exposed by the anchors.
- Distinctive: line 145
Adversarial: search explicitly for failures, contrary findings, limitations, retractions, version drift, and plausible alternatives.
- Distinctive: line 151
Prefer specifications and official documentation for defined behavior, then original papers, raw data, and reproducible benchmarks.
- Distinctive: line 178
Stop when the applicable condition is met:
- Distinctive: line 189
Do not stop merely because a plausible recommendation appeared early.
- Distinctive: line 194
Before writing the decision brief:
- Distinctive: line 197
Check correctness: each citation entails the attached claim.
- Distinctive: line 198
Check completeness: every externally verifiable, decision-changing claim has support or is labeled inference/unknown.
- Distinctive: line 200
Check source quality: the cited source is appropriate for that claim and material limitations are disclosed.
- Distinctive: line 202
Check coverage: every research question has a sufficiency status and all material contradictions are resolved or visible.
- Distinctive: line 204
Remove unsupported decoration, redundant citations, and claims that exceed the evidence.
- Very specific: line 223
Do not repeat the same claim under evidence, trade-offs, and recommendation.
- Distinctive: line 257
Run only after explicit approval.
- Very specific: line 273
When invoked by Glasses or Pointer, preserve the passed decision, evidence, answered Q-labels, confidence, and acceptance boundary.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Homework
The nerd does the homework and brings receipts.
$ARGUMENTS optionally provides a research question or target. Support three
depths. Treat source counts as planning ranges, not proof of completeness:
skim: check a narrow claim with the fewest decisive sources, usually 1-3.study: compare a decision with primary evidence and independent cross-checking, usually 4-10 sources. This is the default.thesis: build an auditable evidence notebook, search for counterevidence, and continue until every decision-changing question is sufficiently supported or explicitly unresolved. Do not substitute a larger source count or longer report for this workflow.
Honor an explicit depth. Infer thesis when the user asks for deep,
comprehensive, exhaustive, or thesis-grade research. Depth controls evidential
work, not response verbosity.
Language preservation
Respond in the language of the user's latest request unless they explicitly request another language. Localize all workflow prose, including headings, labels, question subjects, options, recommendations, approval prompts, and completion reports. Any English examples in this skill specify meaning, not literal output.
Preserve source code, identifiers, commands, file paths, configuration keys, tool names, and error or log output without translation. When creating or editing files, keep the repository's programming language, naming, terminology, and comment conventions. Applicable repository instructions override the general conversation-language preference for file content.
Establish the assignment
Explicit general scope overrides project context. When the user asks to research a topic generally or without applying it to their project, keep the research and recommendation general; do not use the repository, company, or recent conversation as hidden decision criteria. Use local context only when the user requests a project-specific application or it is required by the stated assignment.
- Identify the supplied target and, only for an explicitly project-specific assignment, inspect the relevant conversation and local project context.
- Infer the likely objective, scope, constraints, and exclusions, but treat them as unconfirmed.
- Always ask Q1 before proposing a research plan or performing research, even when the assignment appears clear. Confirm the inferred objective, scope, and exclusions with the user, who can correct them. Do not replace this question with the research-plan approval.
- After the Q1 answer, ask further decision-changing clarification questions one at a time only until the research target is clear. Ask as many as needed, but do not repeat resolved questions or ask for facts available locally.
- Label questions sequentially with a short subject. The mandatory Q1 must
name the inferred objective, scope, and exclusions; for example:
Q1 - <localized subject>: <localized question>Continue with Q2, Q3, and so on. Preserve the labels and answers in the plan. - Prefer the platform's native question tool (
AskUserQuestionin Claude Code orrequest_user_inputin Codex) when available. Invoke it with exactly one question, stop immediately, and continue only after the user answers. If no native question tool is available, put the question on the final line and stop. - Keep the entire question turn under 120 words: no preamble beyond one decision-relevant sentence, short options, and one recommendation sentence.
When no native question tool is available and a question has multiple viable answers, format it as:
(a) <option and consequence>
(b) <option and consequence>
(c) <option and consequence>
<localized recommendation label>: <preferred option and concise evidence-based reason>
Q1 - <subject>: <question>Do not prefix options with bullets or hyphens. Keep the localized recommendation label separate from the options. Offer only materially different options and stop after the final question line.
Research plan
Before external research, present a plan containing:
- the decision to support;
- confirmed requirements, constraints, and exclusions;
- numbered research questions mapped to the clarification answers;
- depth and why it fits;
- evidence standards, preferred sources, and comparison criteria;
- search waves, source/time/token/cost ceilings, and stopping conditions;
- for
thesis, the evidence-notebook fields and required audits.
Keep a straightforward research plan under 200 words. Combine related constraints and criteria instead of creating a section for every list item. Do not repeat assignment context already confirmed in the conversation.
End by asking, in the user's language, whether to run the research plan.
Ask this approval through the platform's native question tool when available, with localized options for running or revising the plan. Invoke the tool and stop. Otherwise put the localized question on the final line and stop. Wait for explicit approval. Do not treat an unrelated reply as consent.
Evidence notebook
For thesis, maintain an evidence notebook as working state. Do not create a
repository file unless the user asks. Keep it compact but auditable:
- Question map: research question, dependencies, and decision impact.
- Source ledger: source ID, type, relevant claims, authority for those claims, directness, methodology, currency, independence, limitations, and include/exclude status.
- Claim ledger: one material atomic claim, exact supporting passage or location, supporting and contradicting source IDs, status, and confidence.
- Search log: search wave, query or route, useful results, exclusions, and remaining gaps.
Assess reliability relative to the claim. Official documentation is strong for documented product behavior but not automatically for comparative superiority. A peer-reviewed paper is not automatically current or applicable. Multiple pages repeating one underlying report are one evidence lineage, not independent corroboration. Record material source limitations rather than hiding them in a generic confidence score.
Research pass
After approval:
- Inspect applicable local context before searching so evidence is evaluated against the actual environment.
- Search in waves:
- Anchor: specifications, official documentation, original datasets, and primary papers that define the subject.
- Expansion: follow references, terminology, implementations, and adjacent questions exposed by the anchors.
- Corroboration: seek independent evidence for decision-changing claims and operational reports for real-world behavior.
- Adversarial: search explicitly for failures, contrary findings, limitations, retractions, version drift, and plausible alternatives.
- Closure: target only unresolved claims and evidence gaps.
- For
skim, stop after decisive anchor evidence unless it conflicts. Forstudy, run only applicable waves. Forthesis, run every applicable wave; never skip the adversarial wave. - Prefer specifications and official documentation for defined behavior, then original papers, raw data, and reproducible benchmarks. Use maintainer or implementation reports for operational detail and community reports only for experience or leads.
- Evaluate authority, directness, methodology, currency, independence, and applicability before relying on a source. Do not use popularity or search rank as a reliability proxy.
- Decompose compound findings into atomic claims. Attach citations to the narrowest claim they support and retain enough passage context to detect qualification or reversal.
- Record contradictions. Resolve them using claim-relative authority, methodology, applicability, date, and independence; do not decide by simple source majority. Preserve unresolved disagreement in the final brief.
- Mark each research question
sufficient,partial, orinsufficient:sufficient: a diligent reader can derive the decision-relevant answer from the evidence without filling a material gap from model memory;partial: useful evidence exists, but a missing fact or unresolved contradiction could change the decision;insufficient: the evidence cannot support a bounded answer.
- Continue searching while a feasible query can resolve a decision-changing
partialorinsufficientquestion. Distinguish absence of evidence from evidence of absence. Abstain from unsupported conclusions. - Compare viable options against the confirmed constraints. Do not produce a generic web summary or manufacture alternatives.
Stopping conditions
Stop when the applicable condition is met:
skim: the narrow claim has decisive evidence or a material conflict has escalated the task beyond skim.study: the decision is supported, material claims are cross-checked, and remaining uncertainty is visible and unlikely to reverse the recommendation.thesis: every decision-changing question is sufficient or explicitly unresolved; the adversarial search found no unaddressed decisive evidence; new searches are yielding duplicate evidence lineages; and the approved ceiling is reached or further searching cannot change the recommendation.
Do not stop merely because a plausible recommendation appeared early. Do not keep searching to meet a source-count target after evidential saturation.
Citation and coverage audit
Before writing the decision brief:
- Split material conclusions into atomic claims.
- Check correctness: each citation entails the attached claim.
- Check completeness: every externally verifiable, decision-changing claim has support or is labeled inference/unknown.
- Check source quality: the cited source is appropriate for that claim and material limitations are disclosed.
- Check coverage: every research question has a sufficiency status and all material contradictions are resolved or visible.
- Remove unsupported decoration, redundant citations, and claims that exceed the evidence.
Decision brief
Lead with the recommendation. Include only decision-relevant, non-empty sections from:
- Question
- Constraints
- Method and coverage
- Evidence
- Options and trade-offs
- Recommendation
- Confidence
- Contradictions and unknowns
- Smallest decisive validation
- Sources
Do not repeat the same claim under evidence, trade-offs, and recommendation.
Combine sections when only one viable option remains. For study, target about
400 words plus source links; expand only for material uncertainty, severe risk,
or multiple viable trade-offs. For thesis, let evidential complexity determine
length, but keep the main recommendation scannable and move ledger detail into
compact tables or appendices only when it helps verification.
For thesis, report source lineages considered, material exclusions, research
question sufficiency, unresolved contradictions, and which evidence could still
change the recommendation. Distinguish sourced facts, measurements, reports,
and inference.
Validation
Recommend the smallest bounded validation that could change the decision, not
the cheapest experiment regardless of evidential value. Prefer a cheaper test
when it is equally decisive. Permit a larger validation in thesis when no
smaller experiment can resolve the decision boundary.
Define:
- the experiments and their execution order;
- the expected evidence from each;
- the time, token, monetary, and hardware budget;
- stopping and early-exit conditions;
- resources or actions required from the user;
- the result that would reverse or materially revise the recommendation.
End by asking, in the user's language, whether to run the validation plan.
Ask this approval through the platform's native question tool when available, with localized options for running or revising validation. Invoke the tool and stop. Otherwise put the localized question on the final line and stop.
Run only after explicit approval. Execute every approved experiment available within the stated budget until a stopping condition is reached, then audit and report the combined evidence. Do not request separate approval for each matrix cell or stage already named in the approved plan.
Pause when required hardware, credentials, or user action is unavailable. Request new approval before adding experiments, increasing the budget, changing the decision boundary, or otherwise expanding the approved scope. Do not begin an open-ended autonomous experimentation loop.
Handoffs
Invoke Glasses after research when the evidence is ready to become or revise an implementation plan. Invoke Pointer when the evidence identifies a concrete defect in an existing implementation.
When invoked by Glasses or Pointer, preserve the passed decision, evidence, answered Q-labels, confidence, and acceptance boundary. Do not hand back for the same evidence or decision. This prevents recursive bouncing.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

