Agent-environment retrospective

SkillAI & models

Use when a completed session needs an agent-environment retrospective: a severity-ranked list of environment improvement candidates, each backed by session evidence. Not for an engineering retrospective from telemetry — use engineering-retrospective.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Agent-environment retrospective skill

What this skill tells your AI

The instructions your AI receives, as published by outlinedriven/outline-driven-development in .devin/skills/agent-environment-retrospective/SKILL.md and read by ahel’s review.

Contract

FieldBound contract
TriggerA completed session needs an agent-environment retrospective.
AuthorityRead-only. No file, VCS, credential, paid, published, deployed, or remote mutation.
Side effectChat output: severity-ranked environment improvement candidates.
DoneEvery candidate names evidence and the friction it removes.

Inputs

  • Session artifact (required): the completed session transcript or state record. Must contain observable agent-environment interaction.
  • Environment context (optional): the agent's working environment at session time. Use only if supplied; do not infer it.

Procedure

  1. Gather inputs. Receive the session artifact and any supplied environment context. Done when: the session artifact is received and any supplied environment context is noted.
  2. Identify friction. Scan the session artifact for patterns where the agent's environment created friction: tool failures, slow retries, missing context, state loss, repeated navigation, or unclear feedback. Done when: every friction pattern in the artifact is identified or the artifact is confirmed friction-free.
  3. Classify candidates. Assign each friction point a type: tool-failure, slow-retry, missing-context, state-loss, navigation-overhead, or unclear-feedback. Done when: every identified friction point has an assigned type.
  4. Rank by severity. Order candidates: high (blocks progress) → medium (degrades efficiency) → low (minor friction). When severity ties, prefer candidates with stronger evidence. Done when: candidates are ordered by severity with ties broken by evidence strength.
  5. Validate evidence. For each candidate, confirm the named evidence appears in the session artifact. Candidates without traceable evidence are omitted. Done when: every candidate is either confirmed against traceable evidence, omitted, or retained as unconfirmed with its severity downgraded.
  6. Return report. Output the severity-ranked candidate report. Done when: the report is emitted with every surviving candidate carrying type, evidence, severity, and friction_removed.

Failure and recovery

  • No session artifact: return an empty report stating "No session artifact supplied."
  • No friction observed: return a report stating "No environment friction detected." with zero candidates. Do not fabricate candidates.
  • Ambiguous evidence: downgrade the candidate to unconfirmed severity rather than guess. Include the ambiguity in the evidence field.

Output

A severity-ranked markdown report. Each candidate entry contains:

  • type: friction type
  • evidence: verbatim session evidence
  • severity: high, medium, or low
  • friction_removed: what eliminating this friction would achieve

Signals

GitHub stars
52
Forks
9
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
agent-environment-retrospective-outlinedriven
Source
github.com/outlinedriven/outline-driven-development