Reconstruct an AI project's work and decisions
Preserve original requests, actions, corrections and the current position without rewriting what happened.
When requirements or direction change, or when the reasons behind a project's current state need to be reconstructed.
Competence · Knowledge and continuity · Public edition 0.3.2
A client requests one page and approves a draft, then selects two separate paths. Preserve the original request and approval, the draft produced and the new decision; identify work that remains useful.
The key relationship. Keeps a trace of material events and distinguishes the person's source, the assistant's interpretation and observed effects.
An illustrative situation to adapt to your work. Read the complete cases ↓
You supply the context: research, a service, a creative project or a team’s work. Examples make the method visible; adaptation preserves the relationships that make it work while understanding the purpose, sources, people and means of your own assignment.
Tell your AI what you want to achieve, what already exists and a practical situation to address. Ask how this competence changes the work and which knowledge from your domain should participate. MAIOS provenance identifies the shared knowledge; you can use it directly in your AI or compose it with an existing kernel.
Adaptation examples in other contexts ↗A decision can change without erasing what happened before it. This competence preserves the connection between an original request, the AI's interpretation, completed work and a later correction.
The result is a usable reconstruction for whoever resumes the project. Detailed history stays reachable; the current entry contains what is needed to continue.
When a request, decision or correction materially changes the work, retain the original wording or an exact reference accessible within the project. Keep uncertainty intact. Attribute the assistant's summary and inferences separately. If the source is missing, identify the reconstruction gap instead of recreating a quotation from memory.
Record events that change reconstruction: a source, an AI proposal, an evidenced action, a correction, a change of authorization or a material unknown. Link effects to their evidence and observation time. A prepared file, a sent message and a received reply are different events. Preserve completed actions even when their original plan is abandoned.
A correction adds a new relation: identify the interpretation or instruction it supersedes and how continuation changes. Do not backdate a later understanding. Revoking permission changes what may happen now; it does not undo a completed effect. An AI summary cannot create authorization.
Keep a short entry linking still-valid sources, decisions, completed effects, unknowns and the relevant continuation. Detailed records and the current entry serve different purposes. Identified events and predecessor references help complex work; attributed, dated notes may suffice otherwise. At reentry, check relevant sources and disclose missing or changed coverage.
Use the selected project's continuity surface and preserve concurrent contributions. If persistence is unavailable, return a note the person can retain. Avoid a universal diary and exclude secrets and unrelated private data. When reentry exposes a causal gap, improve the preservation method rather than merely adding more records.
These cases illustrate situations, causal distinctions and choices you can adapt to your work.
Italiano. The following cases illustrate how the method is used.
A client requests one page and approves a draft, then selects two separate paths. Preserve the original request and approval, the draft produced and the new decision; identify work that remains useful. The current entry leads to the new structure without making the earlier approval disappear or appear mistaken.
A document was sent before a figure was corrected. Record the identity of the sent version and the observed effect, then add the correction. An updated local copy does not replace the document already received. Any subsequent sending follows the actual mandate.
The new chat receives sources, decisions and documented actions. A cited conversation is no longer accessible: identify that gap and continue from available information. Reconstruction distinguishes what can be checked from what is known only through a summary.
Use SKILL.md or START_CHAT_EN.txt with the relevant sources and your actual context. Keep observations, interpretations and what remains open distinguishable. Preserve the useful reason and continuation in the project’s guide.
Keeps a trace of material events and distinguishes the person's source, the assistant's interpretation and observed effects.
The textual method is usable on its own. This kit also supplies the source competence's Python reducer, event/state JSON schema, a worked ledger and portable tests. The reducer is a record-integrity tool, not a semantic admission gate or an independent authority. Hashes detect inconsistent records; they do not prove that a statement or evidence reference is true. Keep the original ledger and source locators in the project that owns them.
The seven event kinds preserve exact source, assistant projection, observed movement, retraction or supersession, authority change, unknown change and a current-relation checkpoint. Sources retain attribution. Projections cite preceding events. A correction supersedes an interpretation without deleting its event or erasing observed effects. A checkpoint exposes current sources, remaining unknowns and the next permitted action. Secret-classified sources retain a gap marker, without secret values or hashes. Never copy a credential into a record.
Use Python 3.10 or later from the competence folder. Run python scripts/demo_continuity.py --output-dir demo-ledger with a new directory name. It writes an initial state, six illustrative events and their successor states, and reconstruction.json. The example corrects an assistant's publication interpretation while preserving a draft actually saved. It does not publish or contact anything.
Run python scripts/reduce_material_work_continuity.py verify --state demo-ledger/C1.state.json to reconstruct that state. For your project, init accepts --owner-surface, --owner-state-ref, --authority and optional --output. reduce accepts --state, --event and optional --output; use a separate output file to retain the preceding state. An event's previous_state_digest must match the state it extends. The demo shows how to construct each event through the Python interface.
An exact replay is idempotent; an altered event with the same identity is rejected. A stale predecessor, tampered state, fabricated authority reference or unsupported resolution of an unknown is rejected. A structured authority-grant source models a recorded change; the tool cannot authenticate the human behind an externally supplied document. A referenced source must be an absolute regular file with the recorded digest, at most 1 MiB; missing, symlinked or changed locators produce incomplete reconstruction. Inline exact sources remain available in the ledger.
The CLI writes only its chosen output and has no network calls. Concurrent writers need serialization at the actual file/store boundary; this reducer alone supplies no file lock or database transaction. A schema-valid state is also checked by replaying its event chain. Run the portable tests with python -m pip install -r requirements-test.txt followed by python scripts/test_reduce_material_work_continuity.py. The reducer itself uses only the standard library; jsonschema is needed by schema tests. The schema retains original protocol names for compatibility, with a public schema identifier and Windows/POSIX source paths.
Preserve original requests, actions, corrections and the present position without rewriting what happened. Keep material events connected and distinguish the person's source from assistant interpretations and observed effects.
Use this competence when requirements or direction change, or when a project’s present state needs causal reconstruction. A decision can change without erasing earlier events. Connect original request, AI interpretation, actual work and later correction. Keep detailed history reachable while the current entry contains what is needed to continue.
When a request, decision or correction materially changes work, preserve its original words or an exact accessible reference. Keep uncertainty and unusual wording. Attribute the assistant's summary and inferences separately. If the source is absent, identify what cannot be reconstructed rather than recreating a quotation from memory or a summary. Use the existing continuity location when suitable.
Record events that change reconstruction: a source, AI proposal, evidenced action, correction, authorization change or decisive unknown. Connect each effect to its evidence and observation time. Preparing a file, sending it and receiving a reply are different facts. Retain the action that occurred even when its producing plan is later abandoned.
A correction adds a relation to history: identify the interpretation or instruction superseded and the continuation changed. Do not erase the previous event or backdate later understanding. Revoking permission changes what can happen now without undoing an earlier effect. An AI summary cannot create authorization; retain the person's or responsible party's actual decision reference.
Keep a short entry connecting still-valid sources, decisions, completed effects, unknowns and the pertinent next movement. Detailed register and current entry have different functions. Use identified events and predecessor references when complexity requires them; attributed dated notes may otherwise suffice. At reentry, check that useful references are reachable and identify absent or changed material. Do not claim coverage of an unread source.
Write within the selected project while preserving concurrent contributions. If persistence is unavailable, prepare a note the person can retain. Do not build a universal diary. Exclude unrelated private data and secrets, including derived fingerprints. When reentry exposes causal loss, improve the continuity method so the next note supports better reconstruction rather than adding volume alone.
This competence concerns material continuity: what happened and the sources it derives from. It does not alone determine a decision's meaning or the whole project's state.
The relation becomes meta when reconstruction failure changes preservation itself. A summary may have mixed proposal and approval; continuation makes both recognizable. This is distinct from autonomous execution, and a well-formed register does not establish infallible memory.
A handoff reveals which connections suffice and which are missing. Retain the correction in continuity while leaving project choices with the responsible party.
Begin with available context and sources. The method can operate in chat; persistence, tools and external actions depend on the recipient's actual means and mandate. Instructions make a capability usable; its behavior and value are observed in real work.
Read OPERATING_GUIDE.md as part of this method when forming the work. It carries the practical knowledge, decision relations and available means for this capability. The complete chat starter includes this knowledge and worked examples, so it remains usable without private sources. Preserve the public source identity and adapt through the recipient's actual means.
This competence works on material continuity: what happened and which sources support it. It does not independently decide the meaning of a choice or the overall state of a project.
A meta relation appears when a reconstruction failure changes the preservation method. For example, a summary confused a proposal with approval; the revised method keeps both identifiable. This is distinct from autonomous execution, and a well-formed record does not prove infallible memory.
Competence, meta and autology: the conceptual guide ↗A handover reveals which links are sufficient and which are missing. Return the correction to the continuity method while leaving project decisions with their responsible owner.
Attach or paste START_CHAT_EN.txt, which contains the complete method, and add your context. The AI can work in the conversation's language. Retain the result and useful guidance if you want to resume in another chat.
Make the competence folder readable and have the agent read SKILL.md for the relevant work. Connect real sources and tools according to the environment's instructions. The kit does not install accounts, memory or recurring execution.
With an existing kernel, connect the method to the competence or guide already responsible for the work. Without one, apply the instructions directly and preserve useful knowledge in the project or a reusable note.
Read or copy the complete English starter and add your context. English is the kit’s primary language; Italian counterparts are included.
