Skip to content

Latest commit

 

History

8 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Overpowered

The skills that make capable agents work like excellent professionals.

Overpowered is a composable Agent Skills suite for knowledge work, enterprise automation, evidence-driven decisions, and reliable agentic execution. It encodes the habits that strong knowledge workers apply almost automatically: get the right context before acting, distinguish authority from precedent, surface contradictions, hunt for exceptions, choose the right automation mechanism, preserve human authority where it matters, and verify outcomes with evidence.

It is intentionally not a collection of role prompts and it does not try to replace harness features such as goals, subagents, memory, RAG, or tool calling. Overpowered sits above those capabilities and teaches agents when and how to use them well.

The design target is the kind of leverage that disciplined software-agent workflows provide, applied to documents, data, business processes, organizational knowledge, policies, contracts, proposals, operations, and mixed human/agent workflows.

One-sentence philosophy

Know enough before you act, make uncertainty explicit, use the right evidence, model exceptions, preserve human authority where it matters, and prove the outcome.

Why “Overpowered”?

A capable model already knows how to summarize a document, query a tool, or draft an answer. The largest quality gains often come from a small number of disciplined behaviors that the model does not apply reliably by default. Overpowered packages those behaviors as reusable skills.

capable agent
    + know-enough
    + what-changed
    + find-precedent
    + find-the-exceptions
    + automate-this
    + evidence / verification primitives
    = dramatically stronger real-work behavior

The goal is not to invoke every skill on every task. The goal is to make the right professional move at the right time.


Why this suite exists

Modern harnesses increasingly provide excellent runtime capabilities: tool calling, delegation, goals, subagents, memory, retrieval, and self-improvement. Those capabilities do not by themselves tell an agent how to work well.

This suite lives one layer above the runtime:

runtime / harness capabilities
        │
        │ tools, RAG, subagents, goals, filesystem, databases
        ▼
Overpowered
        │
        │ decision policies and reusable procedures
        ▼
contracts, proposals, datasets, policies, processes, operations...

A retrieval extension can answer a query. know-enough decides whether a query is needed, what information could change the outcome, which source is appropriate, whether the result is authoritative or merely precedent, and when to stop retrieving.

A database tool can execute SQL. ask-the-data turns a folder of structured data and dictionaries into an evidence-backed analytical workflow.

A harness can keep running until a goal is reached. completion-audit independently reconstructs what “done” means and checks the evidence behind the claim.


Architecture

The suite has five layers. The first three encode professional working methods, Level 3 routes them, and Level 4 lets an agent acquire a missing capability without turning the harness into a junk drawer.

Level 0 — Discipline primitives

Small, composable rules that improve other skills and workflows.

Skill Core rule Typical use
evidence-first Claims require proportionate evidence. Research, analysis, decisions, reporting
assumption-audit Separate known facts, inferences, assumptions, and unknowns. Planning, analysis, architecture
completion-audit Activity is not completion; reconstruct success criteria and verify them. Before declaring work done
human-gates Keep humans only where authority, ambiguity, risk, or accountability requires them. Automation design and execution
dry-run Simulate side effects before committing risky or broad changes. Data updates, automation, migrations
checkpoint Make work resumable by a fresh agent without relying on chat history. Long tasks, cross-agent handoff

These primitives may be invoked directly, but their main value is composition.

Level 1 — Knowledge and evidence skills

Skill Promise
know-enough Acquire exactly the knowledge that can materially improve the next decision; no less, no more.
find-precedent Find comparable historical cases and extract reusable lessons without blindly copying them.
ask-the-data Query structured files through a durable analytical layer and return traceable answers.
reconcile Resolve or expose disagreement across sources without silently choosing one.
what-changed Explain semantic changes between versions and why they matter.

Level 2 — Process and automation skills

Skill Promise
find-the-exceptions Turn happy-path rules into an exception-aware decision model.
automate-this Compile a human process into the right mix of deterministic automation, agents, systems, and human gates.

Level 3 — Orchestration

Skill Purpose
using-overpowered Select and sequence the smallest useful set of skills, reuse existing capabilities first, and route to adaptive capability only for a proven execution gap.

Level 4 — Adaptive capability

Skill / protocol Purpose
gear-up Create and activate the smallest temporary capability needed to close a real execution gap, then measure whether it helped.
Skill Academy Evidence lifecycle for generated capabilities: ephemeral → candidate → qualified → graduated. See ACADEMY.md.
skillify Generalize and package a qualified/proven workflow into a durable portable Agent Skill.

The separation is deliberate: gear-up optimizes for the task in front of you; the Academy decides what deserves to survive; skillify optimizes a proven method for durable reuse.


The core workflow

The suite does not prescribe one giant pipeline. using-overpowered chooses only what the task needs. A typical knowledge-work flow looks like this:

request
  │
  ├─ material unknowns? ───────────────► know-enough
  │                                      ├─ historical analogy? → find-precedent
  │                                      └─ structured data?    → ask-the-data
  │
  ├─ sources disagree? ────────────────► reconcile
  │
  ├─ comparing versions? ──────────────► what-changed
  │
  ├─ hidden rules / edge cases? ───────► find-the-exceptions
  │
  ├─ process to automate? ─────────────► automate-this
  │                                      ├─ human-gates
  │                                      └─ dry-run when risky
  │
  ├─ material execution gap remains?
  │      ├─ missing knowledge? ────────► know-enough, not creation
  │      ├─ installed/Academy match? ─► reuse / evaluate
  │      └─ truly uncovered? ─────────► gear-up
  │                                      └─ validate → activate → measure
  │
  └─ claiming completion? ─────────────► completion-audit

The important property is conditional composition. Do not load every skill “just in case,” and do not generate a capability “just in case.” gear-up is the last-mile response to a proven execution gap, not a preparation ritual.


Example 1 — Contract clause review

User request:

Review this liability clause and tell me whether we can accept it.

A disciplined agent should not immediately answer from generic legal knowledge.

using-overpowered
  ↓
know-enough
  ├─ Need current company legal position
  │    → authoritative legal-policy knowledge base
  └─ Need negotiation precedent
       → find-precedent over signed historical contracts
  ↓
reconcile, only if current policy and precedent appear inconsistent
  ↓
human-gates, because final legal approval may require delegated authority
  ↓
evidence-first response

Possible result:

Current policy
  Standard liability cap: 100% of contract value.

Relevant precedent
  Three comparable signed agreements were found.
  One accepted 150%; two remained at 100%.

Interpretation
  150% is a historical precedent, not the current normative position.

Decision
  Do not auto-approve. Route to the authorized legal approver.

Evidence
  Policy: Legal Playbook §4.2
  Precedents: Contract A §12, Contract B §9, Contract C §14

The key distinction is:

policy      = normative authority
precedent   = historical evidence

The suite must never collapse those concepts.


Example 2 — Reusing previous proposal material

User request:

Draft the maintenance section of this proposal. Reuse what we have done before where appropriate.

know-enough
  ├─ What capabilities are currently approved?
  │    → product reference KB
  └─ Have we answered similar requirements before?
       → find-precedent over previous proposals

find-precedent
  → extract reusable structure, rationale and wording patterns
  → flag pricing, dates and customer-specific commitments as non-transferable

evidence-first
  → current product claims must come from authoritative current sources

This avoids the dangerous pattern “retrieve a similar proposal and copy it.”


Example 3 — Automating a weekly supplier process

Human description:

Every Friday, Sophie downloads the supplier export, checks anomalies, asks people to correct some of them, updates the master sheet and emails Finance.

A good agent should not jump directly to an “AI agent” architecture.

find-the-exceptions
  ↓
clarify anomaly categories, missing files, duplicate suppliers,
late responses, urgent cases, approval limits, reprocessing rules
  ↓
automate-this

Possible compiled design:

Step Best implementation
Detect new export deterministic event / schedule
Validate schema deterministic code
Match supplier IDs SQL / rules
Classify ambiguous anomaly agentic reasoning
Approve financial exception human gate
Update master dataset deterministic transaction
Notify Finance automation

Before deployment:

dry-run
  → run on a representative historical export with side effects disabled
completion-audit
  → prove all valid source records are accounted for and no duplicate IDs were introduced

Example 4 — Policy update impact

User request:

The purchasing policy was revised. Tell me what actually changed and what we need to update.

what-changed
  → semantic diff, not line diff
  → identify threshold, role, SLA and obligation changes
  ↓
know-enough
  → retrieve the workflows, training material and automations governed by those rules
  ↓
reconcile
  → where dependent artifacts still encode the old rule

Example output:

MATERIAL CHANGE
Approval threshold: €20,000 → €10,000

Likely impact
HIGH    purchasing workflow
HIGH    invoice approval automation
MEDIUM  buyer training guide
LOW     supplier onboarding

Evidence
Old policy §4.3
New policy §4.1

Example 5 — Ask structured organizational data

Project layout:

project/
├── data/
│   ├── test-facilities.xlsx
│   ├── organizations.csv
│   └── data-dictionary.md
└── TASK.md

User request:

Which facilities can perform environmental tests above 80°C and are available in Q4?

ask-the-data
  ↓
inspect files and dictionaries
  ↓
load useful relations into DuckDB through structured-data-duckdb when available
  ↓
inspect schema and samples
  ↓
query only the required relations
  ↓
return answer + query/evidence provenance

data/ and TASK.md are conventions, not requirements.


Example 6 — Cross-agent continuation

When a long task must move to another harness or a clean context:

checkpoint

produces a state package containing:

Goal
Current state
Authoritative inputs
Decisions made + rationale
Rejected paths + rationale
Artifacts produced
Evidence gathered
Known issues
Open questions
Next recommended action
Completion criteria

The receiver should be able to continue without the original chat transcript.


Example 7 — Gear up during a live task

User request:

Build a cross-document requirement matrix. Existing extraction works, but every available normalization path loses exact requirement IDs. Keep every source traceable.

A disciplined agent should not immediately write helpers. First prove the gap:

using-overpowered
  ↓
existing skills/tools checked
  ↓
knowledge gap? no — documents and rules are already known
  ↓
Academy match? none
  ↓
material capability gap
  ↓
gear-up

gear-up states a value test before creation:

Before
  sampled normalization loses or collides on source identity.

Smallest missing capability
  one deterministic provenance-preserving normalization tool.

Success
  100% of the validation sample preserves document + original ID,
  with zero collisions.

The generated tool is written to an ephemeral task workspace, validated, activated only when the runtime confirms it, and used on the sample first. If it closes the gap, the task proceeds.

After the task, the default is cleanup. If the capability created material value and plausibly generalizes, only its artifact/evidence package becomes a Skill Academy candidate. It is not installed globally.

See the complete walkthrough in examples/07-adaptive-capability.md.


Example 8 — Skill Academy graduation

A temporary capability that worked once is candidate evidence, not a permanent skill. On a later distinct task, gear-up searches Academy metadata before generating again. If the candidate fits, it is staged and evaluated rather than reinvented.

first real task
  → gear-up → useful ephemeral capability
  → Academy candidate

second distinct task
  → reuse candidate → succeeds again
  → clean-context eval vs baseline
  → qualified

qualified method
  → skillify
  → generalized, eval-backed portable skill
  → human/project graduation gate
  → durable installation/publication

See ACADEMY.md for the lifecycle and examples/08-academy-graduation.md for the full example.


gear-up, runtime adapters, and Pi

gear-up deliberately separates adaptive capability policy from hot-loading mechanics.

gear-up
   = prove gap / choose minimum artifact / budget / validate / measure / retain?

runtime adapter
   = create isolated workspace / activate / deactivate / reload / report runtime state

This keeps the skill portable. A harness that cannot hot-load a generated artifact must report it as staged, not active.

For Pi, this repository now includes the executable @raguets/pi-overpowered runtime extension. It uses Pi's resource discovery and reload lifecycle for temporary skills, injects explicitly lower-authority temporary context, and registers constrained generated tools behind a trusted subprocess wrapper. See adapters/pi.md for its lifecycle and security boundary.

Overpowered Runtime for Pi

Install the skills and optional runtime companion directly from Git:

pi install npm:@raguets/overpowered

Or

pi install git:github.com/raguets/overpowered

The extension exposes two model-facing tools, overpowered_capabilities and overpowered_runtime, plus /overpowered:status, /overpowered:cleanup, and /overpowered:academy. gear-up remains responsible for proving the gap and choosing the smallest artifact; the extension only stages, validates, activates, records, and cleans it.

Generated executable tools require interactive confirmation. They run in a subprocess with a sanitized environment, bounded time/output, controlled working directory, and no Pi ExtensionAPI, but v0.3 does not claim OS-level sandboxing. Review generated code and its declared effects before approval.

Install

To install only overpowered kills on any harness:

npx skills add raguets/overpowered

Skill Academy

The Academy is a selection mechanism, not an ever-growing memory dump.

EPHEMERAL  →  CANDIDATE  →  QUALIFIED  →  GRADUATED
     │             │              │
     └─ discard    └─ reject      └─ retire/supersede later

Key rules:

  • temporary by default;
  • one successful task can justify candidate status, not generality;
  • qualification needs evidence beyond the originating task;
  • search lightweight Academy metadata before loading candidate bodies;
  • a candidate stays visibly experimental until graduated;
  • skillify packages/generalizes only after adequate evidence;
  • durable deployment remains an explicit authority decision.

The protocol is documented in ACADEMY.md; templates live in academy/.


know-enough and RAG / pi-rag

know-enough deliberately separates retrieval policy from retrieval capability.

know-enough
   = when / what / where / enough?

pi-rag, MCP retriever, RAG API, vector DB tool...
   = execute retrieval

This makes the skill portable while allowing pi-rag to be an excellent implementation backend in Pi.

A knowledge-source registry is recommended. See:

  • skills/know-enough/references/knowledge-source-registry.md
  • skills/know-enough/references/knowledge-sources.example.yaml
  • skills/know-enough/references/pi-rag-integration.md

The registry should describe meaning, not tool plumbing: scope, authority, freshness, intended uses, and exclusions.


ask-the-data and DuckDB

ask-the-data is the user-facing behavior. A lower-level skill such as structured-data-duckdb can remain the technical implementation layer.

ask-the-data
      ↓
structured-data-duckdb
      ↓
DuckDB / Excel / CSV / JSON / Parquet

If structured-data-duckdb is present, ask-the-data should delegate workbook inspection, ingestion, schema inspection and database operations to it rather than duplicating those mechanics.


Installation model

Each folder under skills/ is a valid Agent Skill unit with a required SKILL.md plus optional references/ and evals/.

Install only the skills you want, or make the entire skills/ directory visible to a skills-compatible harness.

Typical layouts include:

<project>/.agents/skills/<skill-name>/SKILL.md

or harness-specific skill directories. Consult your harness documentation for discovery paths.

using-overpowered provides the suite-level routing policy. It is useful when the harness can automatically invoke matching skills. Explicit invocation remains useful for interactive commands such as what-changed, find-precedent, automate-this, or gear-up. Hot activation of artifacts generated by gear-up additionally requires a compatible runtime adapter; see adapters/.


Skill-writing principles used here

This suite follows the current Agent Skills format and intentionally applies these practices:

  1. Descriptions are routing interfaces. They say both what the skill does and when it applies.
  2. Progressive disclosure. Core behavior stays in SKILL.md; deep material lives in references/.
  3. Moderate detail. Instructions focus on mistakes an otherwise capable agent is likely to make.
  4. Procedures over generic advice. Every skill contains an execution loop or decision protocol.
  5. Defaults over menus. Each skill chooses a preferred method and gives escape hatches only when needed.
  6. Concrete output contracts. Short templates make expected results easy to follow and verify.
  7. Gotchas are first-class. Each skill calls out tempting but incorrect shortcuts.
  8. Composable scope. Skills are designed like functions: coherent enough to be useful alone, narrow enough to sequence.
  9. Eval-ready. Every skill includes evals/evals.json with realistic prompts and assertions.
  10. Evidence before confidence. Uncertainty and source authority are explicit.
  11. Reuse before creation. Adaptive capability is a fallback after installed/runtime/Academy options.
  12. Ephemeral before durable. Generated capability must earn persistence through evidence.

Primary design references:

The suite is also influenced by the broader pattern of strong process skills: a memorable rule, explicit trigger conditions, a short procedure, failure-mode guardrails, and a measurable stopping condition.


Testing and validation

Static validation

Run:

python scripts/validate_suite.py

It checks:

  • every skill directory contains SKILL.md;
  • frontmatter parses as YAML;
  • directory and name match;
  • names satisfy the Agent Skills naming constraints;
  • descriptions are non-empty and within the 1024-character limit;
  • every SKILL.md is under the recommended 500-line limit;
  • referenced local Markdown/YAML/JSON files exist;
  • every skill has evals/evals.json;
  • every eval file names the correct skill and contains prompts, expected outputs, and assertions;
  • the expected Overpowered skill set is present;
  • the Skill Academy protocol/templates and runtime adapter docs required by the adaptive layer exist.

If skills-ref is installed, also run the official validator against each skill:

for d in skills/*; do skills-ref validate "$d"; done

Behavioral evaluation

The included eval files are test specifications, not proof that every model/harness already passes them. For meaningful evaluation:

  1. Run each eval once with the skill and once without it (or against the previous skill version).
  2. Use clean contexts.
  3. Grade the assertions with concrete evidence.
  4. Inspect execution traces for wasted steps and instruction failures.
  5. Tighten the skill only where the evals show a real gap.

This is important: static correctness proves the package is structurally valid; only runtime evals show whether a particular model/harness follows the skill reliably.


Repository layout

overpowered/
├── .github/workflows/validate.yml
├── .gitignore
├── LICENSE
├── README.md
├── ARCHITECTURE.md
├── ACADEMY.md
├── CATALOG.md
├── REFERENCES.md
├── CONTRIBUTING.md
├── CHANGELOG.md
├── RELEASE_CHECKLIST.md
├── VERSION
├── adapters/
│   ├── README.md
│   └── pi.md
├── academy/
│   ├── README.md
│   ├── candidate.template.yaml
│   └── index.template.yaml
├── examples/
│   ├── 01-contract-review.md
│   ├── 02-proposal-reuse.md
│   ├── 03-process-automation.md
│   ├── 04-policy-change-impact.md
│   ├── 05-data-question.md
│   ├── 06-cross-agent-checkpoint.md
│   ├── 07-adaptive-capability.md
│   └── 08-academy-graduation.md
├── scripts/
│   └── validate_suite.py
└── skills/
    ├── evidence-first/
    ├── assumption-audit/
    ├── completion-audit/
    ├── human-gates/
    ├── dry-run/
    ├── checkpoint/
    ├── know-enough/
    ├── find-precedent/
    ├── ask-the-data/
    ├── reconcile/
    ├── what-changed/
    ├── find-the-exceptions/
    ├── automate-this/
    ├── using-overpowered/
    ├── gear-up/
    └── skillify/

GitHub readiness

The repository includes a GitHub Actions workflow that runs the static suite validator on every push and pull request. CONTRIBUTING.md documents the acceptance bar for new skills and CHANGELOG.md starts version history.

The repository includes the MIT License, matching the public GitHub repository.


Suggested publication strategy

For public positioning, lead with a small memorable set rather than marketing all sixteen skills equally:

know-enough
what-changed
find-precedent
find-the-exceptions
automate-this
ask-the-data
gear-up

Keep the primitives installed underneath as quality infrastructure. using-overpowered can later become the suite-level entry point once the individual skills have accumulated enough real-world eval evidence.

Before release, follow RELEASE_CHECKLIST.md: run the behavioral evals on the harnesses/models you want to support, verify any claimed runtime-adapter behavior, and keep release/version notes current.

About

The skills that make AI agents dramatically better at real work.

Resources

Contributing

Stars

5 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages