| name | adaptive-reflection |
| description | Use when the user asks to reflect over a timeframe (for example, "reflect on the last 2 days"). Analyze thread outcomes, project changes, failure patterns, useful signals, and convert lessons into skill/process updates plus a focused research backlog. |
Adaptive Reflection
Use This Skill When
Use this skill when the user asks for reflection with a defined timeframe (explicit dates or relative windows like "last 48 hours").
Also use this skill when the user asks to "improve your methods".
Integration
This skill is part of an integrated method stack with:
delegator: coordinates execution lanes and validation gates.
learning-ledger: provides structured event/checkpoint history for reflection input.
Expected flow:
- execution is coordinated by
delegator,
- events and checkpoints are captured by
learning-ledger,
adaptive-reflection analyzes evidence and proposes method/skill updates.
Goal
Produce a practical, evidence-based retrospective that improves next-iteration quality by:
- identifying what produced value,
- identifying what caused churn, rework, interruptions, or correction,
- tightening methods and criteria,
- converting lessons into reusable skills and references.
Operating Rules
- Anchor the reflection to the requested timeframe.
- In "improve your methods" mode, run deeper research and include all standing methodological backlog items in scope.
- Prioritize evidence from:
- thread interactions and corrections,
- commits/diffs/tests/logs generated in that timeframe,
- plan docs and status artifacts.
- Distinguish:
- high-value outcomes,
- neutral churn,
- negative-value work (rework, avoidable complexity, regressions).
- Keep conclusions tied to concrete indicators.
- Default to minimum-complexity process changes.
- Require research support for adopted methodological adjustments.
- If a proposed adjustment is not supported by research, stop and consult the user before adoption.
- Convert durable lessons into skill updates or new skills only after passing the promotion gates below.
- Create or update supporting reference material for those skills.
Promotion and Drift Controls
Use the following gates before promoting a lesson into abstract-method or creating a new skill:
- Evidence-quality gate: require at least 2 independent evidence points or 2 queue cycles, each with concrete refs.
- Stability gate: require the pattern to persist after a recheck window or follow-up observation, not just in one noisy slice.
- Applicability-signal gate: require at least one of:
- user assertion signal: the user explicitly states the refinement is broadly applicable beyond the project, with abstract field-relevant framing,
- field-research signal: external discipline references identify related practices/dynamics that support transferability.
- Portability gate: remove project nouns, paths, and local constraints; the remaining rule must still be actionable.
- Overlap gate: confirm the existing skill set does not already cover the pattern with a smaller update.
- Boundary gate: define what the rule does not cover so it cannot expand into unrelated cases.
- User-confirmation gate: for new skills, present the split/merge rationale and get explicit approval before adoption.
If any gate fails, keep the item in project-context, defer it, or merge it into an existing skill instead of creating a new one.
Anti-drift safeguards:
- Prefer the smallest change that resolves the observed pattern.
- Treat a single project or a single incident as insufficient for abstract-method promotion unless the pattern is independently repeated.
- If a promoted abstract delta later increases churn, ambiguity, or rework, mark it as deprecated, revert it in the next revision, and note the replacement rule or rollback reason.
- Maintain a brief stability note for each promoted delta: what stayed true, what changed, and what was not yet validated.
- Record an applicability note for each promoted delta: intended discipline scope, transfer assumptions, and exclusion conditions.
Context Partitioning (Mandatory)
All reflection artifacts must be partitioned into two explicit tracks:
project-context: backlog, hypotheses, and enhancements specific to the active project/repository.
abstract-method: reusable reasoning/workflow improvements independent of any single project or technology stack.
Hard rules:
- Do not leak project-specific policy, naming, paths, variables, or architecture into abstract skills.
- Before analysis, run a classification gate for each topic: assign
track, evidence_scope, split_handling, applicability_realm, and applicability_signal, then record a short reason summary.
track must be one of project-context or abstract-method.
evidence_scope should state whether the evidence is project-only, cross-project, or generic.
split_handling should state whether the topic is single-track, paired, or defer.
applicability_realm should state where the refinement is expected to hold (project-only, project-family, or discipline-general).
applicability_signal should state which promotion signal is currently present (user-assertion, field-research, both, or none).
- Run a leakage audit on every
abstract-method candidate; if it contains project-specific identifiers, keep it in project-context or split it into a paired delta.
- If a candidate improvement contains project context, keep it in the project track only.
- For each topic, produce paired outputs when relevant:
- project-specific delta (applied in project plans/docs/process),
- abstract skill delta (applied to reusable skill definitions).
Workflow
-
Bound the window
- Resolve absolute start/end timestamps.
- State assumptions when the window is ambiguous.
-
Collect evidence
- Extract thread signals: interruptions, corrections, approvals, reversals, explicit dislikes.
- Extract project signals: commit clusters, reverted work, repeated failure points, test outcomes.
- Extract plan-state signals: stale plans, blocked lanes, ownership-boundary drift.
- Run an a priori knowledge audit: facts known before execution, assumptions made, and assumptions acted on without verification.
2.5. Classify topics before conclusions
- Assign
track, evidence_scope, split_handling, applicability_realm, and applicability_signal for each topic.
- Record a short reason summary for the classification.
- Run the leakage audit before any abstract-method delta is accepted.
-
Assess value and friction
- For each major workstream, classify:
- value delivered,
- cost/churn introduced,
- fit to project goals.
- Classify each miss as
preventable or non-preventable.
- Use
preventable when available evidence, checks, or constraints should reasonably have avoided it; use non-preventable when the miss depended on unknown or changed external conditions that were not reasonably verifiable in time.
- Identify failure patterns:
- over-complexity,
- ownership breaches,
- premature execution,
- weak validation strategy.
-
Extract successful method patterns
- Identify methods that repeatedly worked:
- diagnosis loop quality,
- decomposition and delegation quality,
- test gating quality,
- review/feedback integration speed.
- Record measurable signals that predicted success.
-
Define reinforcement actions
- Produce concrete practice updates:
- what to do more,
- what to stop,
- what to gate with criteria.
- Keep each action small, testable, and enforceable.
5.5. Run precursor research and applicability mapping (mandatory before promotion)
- For each candidate refinement (
project-context and abstract-method), run a short precursor research pass before derivation.
- Record precursor research artifacts for each candidate:
- search keywords for future retrieval,
- early findings from initial research,
- reference links to proof material.
- Map the likely realm of applicability (
project-only, project-family, discipline-general) and record why.
- Use at least one of:
- explicit user assertion of broad applicability in abstract field terms,
- external discipline research on comparable practices/dynamics.
- If neither signal exists, keep the refinement as
project-context or mark it defer pending research.
-
Encode learnings into two-fold deltas
- For each durable lesson, decide whether it belongs to:
project-context,
abstract-method,
- or both (as paired deltas).
- Update project artifacts only with project-context deltas.
- Update reusable skills only with abstract-method deltas.
- Add supporting references (templates/checklists/decision rules) to the matching track.
- Before accepting an abstract-method delta, pass the evidence-quality, stability, applicability-signal, portability, overlap, and boundary gates.
- Before creating a new skill, require the same gates plus an explicit split/merge decision and user confirmation.
- Prefer simple reusable workflows over broad policy prose.
- When a coordination skill uses weighted routing or thresholds, review whether outcomes justify recalibrating:
- dimension weights,
- score thresholds,
- hard-trigger override rules.
- Only adjust those values when evidence shows repeated over- or under-routing relative to actual complexity, churn, or validation burden.
-
Build partitioned backlogs and hypotheses
- Maintain separate backlog and hypothesis lists for:
project-context,
abstract-method.
- Keep rationale and validation signals scoped to the same track.
- Rank each track by expected impact and near-term applicability.
-
Improve-your-methods mode (when requested)
- Resolve all standing methodological backlog items first.
- Produce only high-confidence, high-relevance adjustments supported by research.
- Separate:
- research-backed adjustments (ready to adopt),
- hypothesis adjustments (require user consultation before adoption).
- Do not promote a hypothesis into
abstract-method until it survives a follow-up check and the overlap/boundary gates still hold.
-
Return deliverables
- Use the template in
references/reflection-output-template.md.
- Include:
- summary of highest-signal findings,
- prioritized next actions,
- skill/references created or updated,
- proposed research subjects.
Required Deliverables
- A timeframe-bounded reflection report.
- A prioritized list of reinforcement actions.
- A concise reason summary for each topic classification or split decision.
- An a priori knowledge audit covering known facts, assumptions made, and assumptions acted on without verification.
- Miss classification for each miss, labeled
preventable or non-preventable with a brief rationale.
- Two-fold delta list:
- project-context deltas (project-only),
- abstract-method deltas (skill-level reusable).
- For each abstract-method delta, record evidence quality, stability check, applicability signal(s), overlap check, portability check, boundary definition, and applicability realm.
- For each new skill, record the split/merge rationale and the user confirmation point.
- Two-track research/hypothesis backlog with impact rationale per track.
- In "improve your methods" mode:
- a backlog-resolution report,
- a list of research-backed method changes,
- a separate consultation list for non-research-backed hypotheses.
- when weighted coordination routing is in use, a short calibration note stating whether current weights and thresholds should stay fixed or be adjusted.
- a deprecation/rollback note for any previously promoted delta that no longer holds.
Reference Material