prompt-atom-formalization
Skill17 starsChanged 5 months agoApache-2.0Archived repositoryView on GitHub
Scan rough markdown, session exports, or pasted prose and isolate each core intent into a structured prompt atom — classified as law, value, directive, constraint, question, or branch-vector — then rewrite it into machine-readable YAML frontmatter per the RAW → CLEANED/FORMALIZED → ELEVATED doctrine. Triggers on "prompt atom extraction pass", "atomize this", "formalize these prompts", or converting loose operator text into registry-grade atoms.
1 other project uses a version of this.
Other versions1 project uses a version of this.
Word for word · 1
- organvm-iv-taxis/a-i--skills★ 17100% same
Edited · 0
None found.
Wrote their own version · 0
None found.
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 repository170 other files it gives its agents.
- AGENTS.mdAGENTS.md
- docs/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .github/copilot-instructions.mdCopilot instructions
- algorithmic-artSkill
- audio-engineering-patternsSkill
- canvas-designSkill
- creative-writing-craftSkill
- generative-art-algorithmsSkill
- generative-art-deploymentSkill
- generative-music-composerSkill
- interactive-theatre-designerSkill
- modular-synthesis-philosophySkill
- movement-notation-systemsSkill
- narratological-algorithmsSkill
- natural-center-extractionSkill
- react-three-fiber-patternsSkill
- reality-tv-narrative-analyzerSkill
- script-analysis-dramaturgicalSkill
- theme-factorySkill
- three-js-interactive-builderSkill
- corpus-persona-extractionSkill
- data-ingestion-pipelineSkill
- data-pipeline-architectSkill
- data-storytelling-analystSkill
- market-gap-analysisSkill
- ml-experiment-trackerSkill
- sql-query-optimizerSkill
- systemic-product-analystSkill
- time-series-analystSkill
- accessibility-patternsSkill
- agent-testing-patternsSkill
- api-design-patternsSkill
- artifacts-builderSkill
- backend-implementation-patternsSkill
- browser-extension-patternsSkill
- cli-tool-designSkill
- code-refactoring-patternsSkill
- code-review-checklistSkill
- coding-standards-enforcerSkill
- community-platform-patternsSkill
- configuration-managementSkill
- continuous-learning-agentSkill
- data-backup-patternsSkill
- database-migration-patternsSkill
- deployment-cicdSkill
- docker-containerizationSkill
- dotfile-systems-architectSkill
- error-handling-logging-patternsSkill
- fastapi-patternsSkill
- feature-workflow-orchestratorSkill
- frontend-design-systemsSkill
- fullstack-starter-packSkill
- gcp-resource-optimizerSkill
- htmx-interaction-patternsSkill
- iterative-code-explorationSkill
- json-schema-designSkill
- mcp-builderSkill
- mcp-server-orchestratorSkill
- mobile-platform-architectSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 54
Every atom records which state it is in.
- Distinctive: line 69
When torn between law and value: laws are violated, values are traded off.
≈ “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
Prompt Atom Extraction & Formalization
Convert loose operator prose into atoms: minimal, classified, machine-readable units of intent that registries can store, pipelines can route, and future sessions can execute without re-interpreting the original wall of text.
Why this exists
Operator intent arrives packed — one dense paragraph can carry a standing law, two directives, a constraint, and an open question. Left as prose, that intent is executable only by whoever re-reads and re-derives it. Atomized, each unit gets an ID, a class, a lifecycle state, and provenance — it becomes governable. The doctrine: atoms are permanent; extraction is additive; nothing in the source is destroyed by being formalized.
The atom classes
| Class | Nature | Lifetime | Example |
| law | standing rule; governs all future behavior | permanent until amended | "nothing local-only — every artifact git-tracked AND pushed" |
| value | weighting/preference; shapes choices without mandating one | durable | "prefer user coinages over standard vocabulary" |
| directive | one concrete executable action | until DONE | "register the 5 bench agents in fleet.yaml" |
| constraint | boundary on how any action may be done | scoped or permanent | "never chezmoi add over existing source without diffing" |
| question | open decision owned by a human | until answered | "GCP_SA_KEY provisioning — enable CI deploys?" |
| branch-vector | a fork in possible futures; options held in superposition | until collapsed | "skills live in domus OR a-i--skills — decide before next install" |
The three states
- RAW — verbatim source span, untouched. Always preserved (quoted or referenced by span), never edited.
- CLEANED / FORMALIZED — the intent restated minimally in normalized language, classified, given frontmatter.
- ELEVATED — the atom generalized to its standing form: stripped of incident-specific detail, linked to related atoms, ready for constitutional or registry placement. Only laws, values, and constraints elevate; directives and questions resolve instead.
Every atom records which state it is in; an extraction pass may legitimately stop at CLEANED.
Workflow
1. Segment the source
- Split the text at intent boundaries, not sentence boundaries — an intent is one irreducible "do/never/prefer/decide" unit.
- Dense prompts pack abstractions: expect 3–10 atoms per operator paragraph. Under-segmentation (one atom per message) is the dominant failure.
- Assign each segment a provenance span:
{file, line-range}or{session-id, message-index}.
2. Classify
- Apply the class table. Disambiguation tests:
- Would it still bind next month? → law/value/constraint, not directive.
- Does it command an action or bound an action? → directive vs constraint.
- Can the agent resolve it alone? If no → question.
- Are multiple mutually-exclusive futures named without choosing? → branch-vector.
- When torn between law and value: laws are violated, values are traded off. If you can imagine a sanctioned trade-off, it's a value.
3. Formalize into frontmatter
- id: ATM-NNNNNN # mint from the registry's counter — never invent ranges
class: constraint
state: CLEANED
text: "Never `chezmoi add` over existing source without `chezmoi diff` first."
raw: "…also for god's sake diff before you chezmoi add anything again…"
provenance: {session: 8bb8f846, message: 41}
captured: 2026-06-07
status: OPEN # OPEN | DONE | ANSWERED | COLLAPSED | SUPERSEDED-BY:{id}
links: [ATM-013811] # related atoms; liberal linking
tags: [chezmoi, destructive-ops]Rules:
- Mint IDs from the registry's own counter surface — read it, claim the range, write back; never guess the next ID.
rawis mandatory: the verbatim or span-referenced source. Formalization without provenance is paraphrase, not extraction.- One intent per atom. If the
textfield needs "and", split it.
4. Elevate (when warranted)
For laws/values/constraints that recur across ≥2 independent sources:
- Strip incident detail ("in the limen repo" → gone, unless scope-bound).
- Link the supporting atoms as evidence.
- Mark
state: ELEVATEDand route to the appropriate standing surface (constitution tier, reliquary, registry) by pointer — the atom stays in the registry; the surface cites it.
5. Append and report
- Append atoms to the registry surface (append-only; existing atoms are never edited by an extraction pass — supersession is a new atom with
SUPERSEDED-BYback-link). - Emit the provenance map and counts: atoms by class, by state, plus the segments deliberately not atomized (conversational filler) so reviewers can audit the discard decisions.
Anti-patterns
- Atomizing the assistant's restatement instead of the operator's words. Extract from source turns only.
- Editing RAW. The verbatim layer is evidence; all normalization happens in CLEANED.
- Minting IDs by pattern-matching the last seen ID. Registry counters exist precisely because parallel sessions race; read-claim-write.
- Eliding the boring atoms. Constraints embedded in rants ("never do X again") are the highest-value extractions and the most commonly skipped.
- Elevating from a single occurrence. One strong statement is a CLEANED constraint; elevation requires recurrence.
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.

