steelmorgan/skill-learning

Knowledge Accumulation (Skill Learning) — procedure

The body was moved out of the skill-learning-policy rule. The trigger "after a cycle of ≥2 iterations" remains in the rule; here are the retrospective procedure, the recording format, and the criterion for choosing the level/file.

The agent learns from its iterations: what works (techniques) and what breaks (anti-techniques). Knowledge is recorded in references/learned-patterns.md of the owning skill.

Two levels of knowledge

| Level | File | What it stores | |---------|------|------------| | Universal | skill/references/learned-patterns.md | Techniques that work in any 1C configuration | | Project | {project}/.context/learned-patterns.md | Techniques tied to a specific configuration / project |

Separation criterion: if an entry mentions a specific metadata object, form name, attribute or configuration feature → project-level. If it describes a general platform/framework pattern → universal.

Using accumulated knowledge

The reading trigger ("before working with the skill, read its lessons") is in the skill-learning-policy rule (always-on). Here is the split by levels:

• references/learned-patterns.md in the skill directory — universal techniques;
• {project}/.context/learned-patterns.md — project techniques.

Apply confirmed entries as additional rules, candidate as hints.

When to run

After completing a work cycle in which there were ≥2 iterations (wrote → error → fixed → success). If the task is solved on the first attempt, a retrospective is not needed.

Procedure

1. Reconstruct the iteration chain — what was done → what failed → root cause → how it was fixed → what passed.

2. Abstract to a class. Name the CLASS of the error one level above the incident, not the incident itself:

   • bad (instance): "биг_Модуль.Метод:142 crashed because ИдентификаторСтроки was empty";
   • good (class): "accessing a record set field without checking whether it is filled before writing".

   The anti-technique is formulated as a generalized prohibition/check that can be carried over to not-yet-seen special cases. Concrete file:line references go in the source field, NOT in the body of the technique/anti-technique.

3. Test the abstraction. Ask: "To what other situation, besides the one that produced it, does this anti-technique apply?" No answer → this is an instance, not a class: reformulate it higher or reject the entry.

4. Formulate the entry (technique and anti-technique are one entry, two sides of one discovery):

status: candidate | confirmed
класс: <named class of error>
приём: <general rule — what to do, verified by success>
антиприём: <general prohibition — what NOT to do, verified by failure>
почему: <what happens when violated>
шаги: <concrete steps/code, if applicable>
источник: <task, iteration, file:line example>

5. Filter out — NOT a lesson:

   • Typos and accidental syntax errors;
   • One-off environment failures without a reproducible class;
   • An incident that did not pass the abstraction test (an instance without a class).

6. Determine the level and the owning skill:

   • mentions a specific metadata object / configuration → project-level → {project}/.context/learned-patterns.md (Russian, without mapping and synchronization);
   • general platform / framework pattern → universal → references/learned-patterns.md of the owning skill in the RU source of the framework;
   • the owning skill is the one to which the class belongs by content.

7. Find the RU directory of the universal lesson (using the skill-editing-from-project skill): .install-session.json → component_map → skill/{owning-skill-name} → ru_path → its directory → references/learned-patterns.md. Write to framework/ (RU), NOT to the installed ENG symlink framework_eng/ — it will be overwritten by synchronization.

8. Reconciliation cascade (read the target file before writing):

   • a similar class already exists → DO NOT duplicate; update the existing entry (expand the anti-technique boundaries, add the source), result = refined;
   • there is a candidate with the same class → promote to confirmed, result = refined;
   • there is no class → new candidate entry, result = new.

9. Append to the target file; if the file does not exist, create it; do not overwrite existing entries. For a universal lesson, after writing, synchronize the ENG mirror via sync_script from .install-session.json (see skill-editing-from-project).

MUST

| Requirement | Description | |------------|----------| | One level above the incident | The entry is about the CLASS of the error + anti-technique, not a specific incident; the abstraction test is mandatory | | Do not change the body of the skill | Entries go only into references/learned-patterns.md, not into SKILL.md | | Reconciliation / update | Read the target file before writing; do not create duplicates, update the existing class; result matched/refined/new | | RU source for universal | Write the universal lesson into the framework/ (RU) directory of the owning skill via skill-editing-from-project, NOT into framework_eng/; after writing, synchronize to ENG | | First case = candidate | A single case gets status: candidate | | Repeat = confirmed | When a similar situation repeats, promote it to confirmed |

depends_on:

• framework/skills/framework-meta/skill-editing-from-project/SKILL.md