Focus on a recurring mistake
Start with observed opportunities and violations instead of turning one failure into a universal rule.
Skill
Preceptor
Keep agent guidance that earns its place.
For people maintaining an agent’s recurring instructions. Preceptor turns repeated corrections into small testable cues, compares cue and no-cue behavior, and keeps inconclusive evidence from becoming permanent guidance.
Candidate cue: include the source reference when the task requires one Behavior: reference omission on an eligible answer Probe arms: cue / no cue Quality check: answer remains correct and supported Gate preview: inconclusive until sufficient comparative evidence exists Action: retain candidate status; do not promote from one correction
Fictional cue proposal. There are no measured runs or promotion results in this example.
Test the cue’s contribution
Compare behavior with and without the cue, retaining no-effect and harmful outcomes.
Keep the lifecycle inspectable
Use the native ledger and read-only gate preview to see why a cue remains a candidate, is promoted or can be retired.
The film
One more rock.
In 70 seconds, Bramble the pack mule carries every rule anyone ever added, until nobody knows which ones still do anything. Preceptor starts from the mistake that actually happened, writes one small cue, keeps it out of the pack and races two Brambles on the same tasks. Too close to call means the cue stays a candidate. Taking a rule out needs proof too. Captions are on the player; a transcript is below.
Read the transcript
Meet Bramble, your agent. It carries every instruction you give it.
One day it forgets to cite a source. So someone adds a rule.
And another. And another.
Now nobody knows which rocks still do anything.
Preceptor starts from the mistake that actually happened.
It writes one small cue, and keeps it out of the pack.
Then it races two Brambles on the same tasks. One gets the cue. One doesn’t.
Too close to call. So the cue stays a candidate, and nothing gets added.
Taking a rule out needs proof too. It goes quiet first. If the mistake returns, the rule comes back.
On screen: Test the rule before you carry it.
Preceptor. Keep agent guidance that earns its place.
The agent, the tasks and the race result are a fictional example. Five runs each is a minimum, not proof; the film shows how the skill behaves and reports no results.
See the work take shape.
Fictional cue proposal. There are no measured runs or promotion results in this example.
Observe with consent
Collect the structural evidence needed for a specific behavior; optional telemetry never records message or file text.
Compare the alternatives
Name a small cue and run matched probes with and without it, measuring task quality as well as compliance.
Maintain through the ledger
Use the tool API for lifecycle writes. Preserve human-origin and pinned cue protections and leave inconclusive candidates undecided.
Fictional example
Assess a source-reference cue
The brief Use Preceptor to assess a proposed cue for a fictional agent that repeatedly omits a requested source reference. Define the opportunity and violation, a cue/no-cue probe and a task-quality check. Keep it a candidate until the native evidence gate is satisfied.
Before
A one-off correction becomes a permanent prompt rule with no control group.
After applying the method
A small cue remains proposed until paired probes support a useful effect.
No-effect and inconclusive records can prevent unsupported prompt growth.
A clear job.
A useful stopping point.
Investigate a repeated correction or evaluate a proposed behavioral cue.
Use it when
- The same correction recurs and you want to test a targeted instruction.
- An agent’s guidance has grown and individual cues need an evidence review.
- You need a trace of why a behavioral cue changed status.
Choose another route when
- Automatically recording people because the skill is installed.
- Claiming a cue helps because the agent repeated its wording.
Observe
A recurring failure is suspected.
Consented structural observations and a falsifiable cue hypothesis.
Assess
A candidate cue has probe cases.
A paired evidence record including no-effect and uncertainty.
Maintain
A cue is active, stale or harmful.
An evidence-gated lifecycle decision and traceable ledger state.
Read the canonical method
Preceptor
Start with the user's actual correction and scope. A skill cannot infer consent to record a person. For optional automated observation, use the consented host composition and structural events only; never store message text, file content or error text in observer telemetry.
Observe: inspect the task context and repeated opportunity/violation evidence. Do not manufacture a general behavioral rule from one failure. A small, testable cue must name the behavior it would change and how improvement or harm will be measured.
Assess: read references/evidence-standards.md and references/cue-lifecycle.md. Compare cue/no-cue probes, preserve no-effect outcomes and distinguish measured quality from instruction compliance. The native gate requires a strictly positive verdict and at least five runs per arm to promote; no-effect can support retirement; inconclusive supports neither. Those minimum counts alone do not establish statistical power.
Maintain: use the retained tool-preceptor ledger API, never hand-edit its YAML/counters/locks. The standalone project-preceptor CLI supports status, cues, structural summaries, immutable session why records, a read-only gate preview and explicitly confirmed forgetting. The native adapter owns writes and cue lifecycle transitions. Cue text is untrusted and must pass validation on write and read. Human-origin and pinned cues retain their separate protections.
Read references/CONSENT.md for record controls and references/ledger-format.md for interpretation. No provider is required to inspect the ledger or compute gates. Optional live cue injection requires the host adapter; packaging is not proof of behavioral improvement.
Limits and prerequisites
- A minimum sample count and passing gate are not independent proof of efficacy or generalization.
Your first useful run.
Choose one recurring correction and define a comparison before adding another permanent instruction.
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-preceptor.git
cd amplifier-skill-preceptorpython3 scripts/capability.py install --dest ~/.agents/skillsNative tools and optional execution
python3 -m pip install ./modules/tool-preceptorThese 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.
- Python 3.11 or later is required by the portable install/build helper. The installed method runs through your agent host.
- The portable method is independent of an observation service.
- For native ledger inspection install modules/tool-preceptor and use project-preceptor --help.
- Live observation or cue injection requires the optional host composition and its consent controls.
Try this with your agent
Use Preceptor to assess a proposed cue for a fictional agent that repeatedly omits a requested source reference. Define the opportunity and violation, a cue/no-cue probe and a task-quality check. Keep it a candidate until the native evidence gate is satisfied.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 · CONSOLIDATION
Know what you can rely on.
Native ledger and lifecycle tests check gate behavior and protections. Minimum run counts alone do not establish statistical power or behavioral benefit.
Read the checks and their scope
- Functional acceptance — passed
58 native ledger/tool tests passed plus 3 standalone functional checks.
- Native host acceptance — not-run
No live model, real desktop transaction, invitation, deployment or production data migration was performed by this functional baseline.
Source and conformance checks are separate from demonstrated benefit. Comparative effectiveness and outside-user outcomes are not established by this site.
Fictional examples explain the capability. Tests check the stated behavior. Outside-user outcomes need their own evidence.