Turn and Burn — every mission carries the skill-research organ
This is the BACKS AIOS mirror of the proven BACKS skill at
skills/turn_and_burn/SKILL.md. When a mission lands, do not start by
jumping into a build. Burn the roster first — the mission's plan and design
are blind until the platform's skills reflect the latest verified state of the
field.
When to run
Every mission, every burn. A trivial one-line edit may skip the absorb step but still consults the lessons-learned lane (see step 2) so the same mistake isn't repeated.
The burn ritual (six steps)
- Essence lane first. Load
core.steering_loader.load_model_essencefor every model that will touch the mission. The essence files are the know-thyself of each model — failure modes, mitigations, the good source/context/goal shape. Do not skip this; an agent that does not know its own failure modes is the failure mode. - Lessons-learned scan. Read
docs/BACKS_LESSONS_LEARNED.mdand the operator memory graph (memory/MEMORY.md) before planning. Every paid- for mistake is filed there — the same loop never burns twice when the agent reads the lessons first. Local prompts must be capability-aware: consultbackend/prompts/essence/models/*.essence.mdfor know-thyselfdocs/BACKS_LESSONS_LEARNED.md+memory/MEMORY.mdfor cross-session memory. This is the operator's "stop redundant failures across the board" floor.
- Play dispatch. Map the mission to the right playbook
(
config/playbooks/*.yaml): dev-mode-elite-build for builds/fixes, leap-bughunt for bughunts, grading-verification for tribunal lanes, parallel-work for fan-out, security-delivery for transport, design-taste for UI. The play'sinjected_blockIS the baseline — zero baseline pre- prompting, ever. - Quick tribunal on plan + design. Before any builder runs, blind-tribunal the plan and the design with at least 2 independent jurors in fresh, author-disjoint sessions. Builder ≠ grader. A plan that passes blind review is the only plan that gets built.
- Three-questions interrog. Ask: (a) what could fail silently here? (b) what is the cheapest sufficient route? (c) what is the operator's surface — does this prove on it? These three questions cut the high-cost-mistake rate to near zero when asked BEFORE the build.
- Absorb pass + skill update. Search verified-skill.com (and the BACKS memory lanes) for any skill the mission revealed as missing. Absorb it through the standard absorb flow; update existing skills with anything the mission surfaced that the skill did not already encode. Land the skill update alongside the mission's main artifact (whole-seam closure).
How the absorbs flow runs
skills/absorb/SKILL.md is the canonical absorb organ. The burn invokes it
on every missing skill the mission surfaced. After absorb, the mission
itself is rebuilt on the upgraded skill base, then re-tribunaled and landed.
No skill is ever deleted — everything BACKS learns is filed into its memory
lanes (inv 0: organize, never delete).
What never to do
- Do not skip essence + lessons. An agent that "knows" without loading its essence + lessons is hallucinating confidence. Both reads are free; both run before the mission starts.
- Do not start building without a quick tribunal. A blind review of the plan catches more defects than the build itself; skipping it means the defects land.
- Do not park decisions in a file. The existing decision bar (taste, vision, destructive risk) is the only reason to ask. Anything below the bar EXECUTES from invariants + memory + intent.
- Do not bench a model for a single 429. The dispatch doctrine says ollama sparse → local fleet. Swap to another configured, live-preflighted rung only when it is sufficient for the task; preserve the checkpoint and bounded backoff when cloud is hammered.
Cross-references
- The BACKS source of truth:
skills/turn_and_burn/SKILL.mdin the BACKS repo. - The absorb flow:
skills/absorb/SKILL.mdin the plugin (and BACKS repo). - The plays this skill bundles:
plays/elite-build.md,plays/agent-builds.md,plays/grading-verification.md,plays/parallel-work.md,plays/security-delivery.md. - Local prompt integration: see
skills/turn-and-burncompanions in this plugin.