Answer the bounded question
Choose a focused investigation, a justified experiment or a claim-linked manuscript.
Skill
Research
Make the answer traceable to the evidence.
A research skill for decisions, experiments and papers. Build a claim-to-source chain, keep contradictions visible and use a study design that can support the conclusion you want to make.
Fictional protocol: retrieval coverage Question: does the change surface required evidence? Baseline and treatment: frozen versions Unit: source-question pair Errors: retained under the declared denominator Judge: calibrated against literal source evidence Conclusion: pending measurements; no improvement claim yet
Fictional worked example. It illustrates the intended output; it is not a measured result or a real customer story.
Keep the evidence inspectable
Retain source locators, methods, populations, applicability and the reasoning between them.
Learn from an honest comparison
Freeze the protocol, preserve errors and failed trials, and separate exploratory findings from confirmation.
The film
Keep the failures in the count.
In 66 seconds, a search setting looks eighteen points better because nine crashed runs vanished into the hat and a friendly judge agreed. Research turns the house lights on: the crashes stay in the count, the judge is checked against the source text, and the plan is frozen before the next run. You get the protocol and an honest answer: no improvement claim yet. Captions are on the player; a transcript is below.
Read the transcript
Your team tests a new search setting on forty questions.
Nine runs crash. Poof. They vanish into the hat.
The rest look great. A friendly judge agrees. Ta-da! Eighteen points better.
Research turns the house lights on.
Behind the curtain, all nine crashes are still there. They stay in the count.
And the friendly judge gets checked against the actual source text.
Before the next run, you freeze the plan: the questions, the scoring, the stop rule.
You get the protocol, and an honest answer: no improvement claim yet.
On screen: Keep the failures in the count.
Research. Make the answer traceable to the evidence.
The experiment, its numbers and the judge are a fictional example. The film shows how the skill behaves; it does not report results.
See the work take shape.
Fictional worked example. It illustrates the intended output; it is not a measured result or a real customer story.
Define what the answer changes
Name the question, audience, competing explanations, needed evidence and budget.
Gather or measure
Read primary sources, or freeze the experiment and measurement rules before outcomes.
Write what the evidence permits
Link claims to records, include disagreement and uncertainty, and identify the smallest useful next check.
Fictional example
A high score built from broken trials
The brief A fictional experiment reports an apparent gain after several provider errors and a favorable model judge.
Before
Average only the successes, call the judge independent, and write an improvement claim.
After applying the method
Keep errors in the declared denominator, audit manifests and traces, inspect judgment semantics, retain null/disputed outcomes, and withhold the claim until a valid comparison exists.
An attractive number cannot repair a compromised measurement process.
A clear job.
A useful stopping point.
Use when a decision needs sources, an empirical claim needs a study, or a paper needs an auditable evidence chain.
Use it when
- A decision needs more than a confident summary.
- An empirical claim needs a test or a fair baseline comparison.
- A brief or paper needs real sources, methods and limits.
Choose another route when
- A citation generator that invents inaccessible papers or results.
- A reason to treat a model judge, popularity metric or passing self-test as research truth.
Investigate
A question needs a bounded multi-source answer.
A source and claim map, contradictions, applicability, and open gaps.
Experiment
A hypothesis has executable measurements and a justified budget.
A frozen protocol, retained trials, integrity checks, and cautious inference.
Write
Evidence is ready for a paper, brief, or technical report.
A claim-linked manuscript with real citations, figures, methods, and limitations.
Read the canonical method
Research
Build the claim from evidence, keep the limits, and make the work inspectable.
Start from the question and decision
State the question, intended audience/decision, competing explanations, required evidence level, time/cost limits, and what an answer would change. Choose Investigate, Experiment, or Write. Use the lightest valid route; do not launch an autonomous experiment without an executable measurement and a justified design.
Investigate
Decompose the question into claims. Search primary research and authoritative source material; retain exact locators, dates/versions, methods, populations, setting, outcomes, applicability, and limitations. Read sources before citing them. Separate supplied claims, source observations, your inference, contradictions, and unknowns. A secondary summary or model output does not become a primary source by repetition. Show how conflicting evidence changes the conclusion; do not count sources as votes.
Use investigation and writing roles to select the retained deeper agent/context guides. Do not invent citations or interpret an inaccessible paper as read. Write the answer the evidence permits, including a useful bounded negative result.
Experiment
Before outcomes, declare hypothesis, treatment/control, unit of analysis, sample size or minimum detectable effect, measures/guardrails, denominator, error/invalid-trial handling, seeds/repeats, budget, stop rule, and analysis. Freeze the measurement apparatus separately from the intervention. Use autonomous experiment discipline only when its prerequisites hold. Its hash-lock, append-only ledger, error retention, held-out promotion, and judge-calibration warnings are substantive controls, not a substitute for a valid study.
Audit source/data provenance and manifest completeness before trusting a number. Keep failed and ungraded trials visible. Resume only the declared missing/invalid subset, preserving clean records and exact input identities. Distinguish correlated repeats, model variants, and participants. Treat automated judgments as fallible: blind labels where possible, verify literal evidence and semantics, retain disagreements, and state the review independence actually achieved.
For hill-climbing, freeze baseline, intervention paths, fixtures, metric, guardrails, and stopping/cost rule; keep/revert with an append-only ledger. Use held-out confirmation before an effectiveness claim. If apparatus is broken, fix it under a new protocol and preserve the earlier run; do not rewrite outcomes. A cross-vendor judge is not independent proof of truth.
Write
Build a claim-to-evidence outline before prose. Keep methods sufficient to reproduce; distinguish preregistered and exploratory analysis; show negative/disputed results and limitations. A figure must reflect retained data, with units, uncertainty, and denominator. Use original conference templates and compilation utilities when available, but verify venue requirements from current official sources. Generated illustrations are not data figures. No invented authors, affiliations, citations, or results.
Output
Return the bounded answer/manuscript/study, claim-source map, methods and raw-record locators, actual checks, uncertainty, unresolved evidence, and the smallest next step. Optional engine checks establish software behavior only; a packaged skill, power calculation, judge score, or self-test does not establish research novelty or user benefit.
Limits and prerequisites
- A literature basis is not a measured result of this skill. Model judgments are fallible measurements.
- Optional statistical, audit, resume, provenance, and paper-generation engines retain their native dependencies.
- Do not begin a costly live experiment, generate paid media, or publish a manuscript without existing authorization.
Your first useful run.
Bring the question, the decision it affects and the evidence you already have.
The download contains the complete skill folder. Extract it into the skills directory supported by your agent, keeping its references and license together. Or use the source installer below. Python 3.11+ is required for that installer.
Start from the public source
git clone https://github.com/michaeljabbour/amplifier-skill-research.git
cd amplifier-skill-researchpython3 scripts/capability.py doctor
python3 scripts/capability.py install --dest ~/.agents/skillsNative tools and optional execution
amplifier bundle add ./bundle.mdThese commands are for the optional native surface. Follow its installation guide and prerequisites; portable method installation is separate.
Keep the complete skill folder together, including its references and license. The host applies the method in your context; the skill does not start a background service.
- Requires Git and Python 3.11+ for the repository installer. The portable method itself can be read by a compatible agent without a service or provider account.
- The destination is a skills-parent folder. Use the discovery path supported by your agent; ~/.agents/skills is an example, not a universal host setting.
- Keep the complete installed folder, references and license together. Identical reinstall is safe; conflicts are refused. A deliberate --replace keeps a dated backup.
Try this with your agent
Use Research to assess whether the fictional document-retrieval change improves source coverage. First propose a bounded question, treatment and baseline, coverage measure, error handling, budget and stop rule. Prepare a protocol; do not run paid calls.Use after installing in a host that loads skill folders. Supply your own task files and review the result.
Keep the guide nearby
Read the guide · License · Changes · Release metadata · Install guide · Migration · Validation · Native guide · consolidation
Know what you can rely on.
174 native integrity-tool tests and ten synthetic experiment-loop shell checks passed. They check software controls, not research novelty, an empirical treatment effect or an outside-user outcome.
Read the checks and their scope
- Six native integrity tool suites — passed
174 tests: audit 50; block hypothesis 22; power 40; provenance 19; resume 14; stage analyzer 29. Isolated Python 3.12.11 with recorded test dependencies. No new empirical study or live model comparison.
- Autonomous experiment-loop shell checks — passed
10 shell checks passed, including freeze gate, append-only ledger, keep/revert, promotion rules, attribution, templates, and an isolated three-row smoke loop. Synthetic plumbing, not empirical benefit.
- Portable packaging and install — passed
All operations passed. No runtime activation or effectiveness measured.
- Skill discovery metadata — passed
Skill is valid. This is a structural check, not a behavioral benchmark.
174 native integrity-tool tests and ten synthetic experiment-loop shell checks passed. They check software controls, not research novelty, an empirical treatment effect or an outside-user outcome.
Fictional examples explain the capability. Tests check the stated behavior. Outside-user outcomes need their own evidence.