Skip to content

mattvideoproductions/MVP-Prompts / skills/foreman

foreman

Verified

Be the foreman on a big job. One lead model plans the work, hands pieces to helper agents in lanes that never share files, checks everything before accepting it and reports the cost; or runs a long solo build with an owner in the loop. Covers the North Star plan of record, briefs that carry the lead's reasoning, measured gates plus the owner's judgement, strike-based promotion between model tiers, a cost ledger and a mistake register every pass is checked against. Use it when the user types /foreman, says "be the foreman", "be the lead" or "orchestrate this", asks to delegate to subagents or run lanes or passes, or mentions a North Star, three strikes, tiers or promotion between models. Also load it when a task looks too big for one context or needs several agents or days, even if not asked by name; the skill then has you offer it and wait for a yes. Also use it to resume such a project, close a pass, report a session's cost or record a mistake. Not for a change one agent can finish in one sitting.

references/reference-build.md

View on GitHub
# The reference build

What happened when this system was used on one project: a 24-hour software build with one lead, 22 worker transcripts and about 3,500 model responses. Written up on 12 November 2026 by the lead model for the owner. These are measured or observed, and they show where the practice matched the rules and where it did not. It is Project A in `field-data.md`.

A second lead later recounted the cost from the raw transcripts with its own script. The worker figures below matched to the cent. `field-data.md` has that recount and three more projects.

## Team

| Tier | Model |
|---|---|
| Lead | Fable 5.2 |
| Senior | Opus 5.5 |
| Standard | Sonnet 5.5 |
| Light | Haiku 5.5 |
| Independent | Codex, billed elsewhere |

The lead kept the core simulation, which was the hardest technical piece and the one the owner had named as the reason the project would succeed or fail.

## Cost

API-equivalent cost, counted from the transcripts after the build ended. The independent agent is excluded because it is billed elsewhere.

| Tier | Responses | Share of cost |
|---|---|---|
| Lead (Fable 5.2) | 382 | 17% |
| Senior workers (Opus 5.5) | 2,768 | 76% |
| Standard workers (Sonnet 5.5) | 320 | 6% |
| Light workers (Haiku 5.5) | 7 | under 1% |
| Total | 3,477 | |

Dollar figures are left out of this published copy.

- Re-reading cached context was 58% of the senior tier's cost (confirmed by the recount).
- The 22 worker transcripts averaged about 140 turns each (recount: mean 141, median 152), re-reading about 384K tokens per response.
- The recount put the lead at 409 responses. The extra 27 came after this table was made.
- One-hour cache writes were 50% of the lead's cost. The table's author did not break this out.
- The cost was computed once, at the end, when nothing could be changed. A per-pass figure by tier would have shown the senior-tier share early.

## What worked

**Briefs that carry reasoning.** The first wave of briefs were lists of deliverables. The owner corrected that: the point of orchestration is to pass the lead's intellect down. Briefs after that carried the chosen architecture, the parameters and the pitfalls, and first passes landed more often.

**Reasons beside decisions.** The core engine ran on the processor and not the graphics card because other code had to read its state every frame without delay. With that written down, no agent tried to move it. A choice recorded without its reason tends to get improved away by a helpful worker.

**Misses stayed visible.** At the close, one of sixteen component gates still failed (12 against a bar of 15), one visual gate missed by one or two points, and one scenario passed two of its seven conditions. All of them are written in the project's documents as failing.

**Cause over symptom.** A feature that had worked stopped working. The cause was an unrelated object that had been moved into its path. The object was moved back, and the feature's own settings were left alone.

**A pass cap.** The owner set it in plain words: two or three more passes, then the best deliverable available. Without a cap, a system built on refinement has no natural stopping point.

**Stating what no person had checked.** At the close, two whole areas were unverified by a person: the sound had never been heard, and no one had used the product with the intended input device. Both were stated in the final report. The gate table alone would not have shown either.

**The environment contract.** Examples that mattered: never open a visible window or take focus, because the owner may be recording, and automated runs are muted. Large intermediates go on the data drives because the system drive is tight. No commits unless asked. Deletions go to the recycle bin, and emptying it is the owner's choice. Do not weaken a gate. Do not call a prototype a finished product. No personal details in requests to outside services.

**Rules that began as corrections.** Every rule that sounds like a principle began as a specific correction from the owner, saved with its date and reason: pass the intellect down, the lead does the hardest work, three strikes, raise token efficiency at the same quality. Keeping the reason is what lets a later session apply the rule with judgement.

## What fell short

**Start low was not followed.** The build produced 2,768 responses from the senior tier, 320 from standard and 7 from light. Most engineering went straight to the senior tier on the judgement that it needed it, and that judgement was rarely tested by trying a cheaper tier first.

**Fresh agents were not used.** The fresh-agent rule was written down and was the one followed least. See the turn counts above.

**Strikes and tiers were untracked.** They lived in the lead's head, and afterwards the worker count could not be reconciled with the transcripts.

**A stale build fooled the lead twice.** A build command failed without an error, and two test runs went against the old program. The results looked like real results.

**Gates and product ran at different settings.** The component gates ran at a finer setting than the product used, and behaviour differed between the two. A switch for the product's setting was added late.

**One failing gate masked the rest.** The overall check exited with an error whenever any gate failed, so it could not tell a new failure from the known one.

**The current state was hard to find.** The status log grew to about a hundred lines and the present state was at the bottom.

**Tool traps were recorded late.** One shell silently resolved a command to the wrong program. Each such trap cost at least one wasted run before it was written down.

**The lead checked work it had designed.** No independent reader was used for important claims.

**Workers ground on stuck problems,** and some briefs leaned on context that only the conversation held.

## The eleven recommendations that came out of it

In the priority order the lead gave them. None had been tested when written. All are now part of the skill, two of them changed by later evidence: item 1 is now a budget in tokens and not turns, and item 2 now depends on what a failed attempt costs. `audit.md` explains both.

| # | Change | Weakness it answers | Sign it is working |
|---|---|---|---|
| 1 | Context budget per agent, with a forced handoff. | 58% of senior-tier cost was re-reading context. | Average turns per transcript falls; cache-read share falls. |
| 2 | Recorded starting tier, defaulting to standard. | Nearly nine in ten worker responses came from the senior tier. | Most lanes start at standard; promotions appear in the ledger. |
| 3 | Lane ledger in the plan of record. | Strikes and tiers were untracked; transcripts did not reconcile. | Every transcript maps to a ledger row. |
| 4 | Build-freshness check inside every gate. | Two test runs went against a stale build. | A stale build fails the gate with a clear message. |
| 5 | Gates run at the shipping configuration. | Checks passed at one setting while the product ran at another. | Gate and product numbers agree. |
| 6 | Register of known failures. | Any one failing gate made the whole check report failure. | A new failure is visible in one line. |
| 7 | Cost metered each pass. | Cost was known only at the end. | The status log has a cost line per pass. |
| 8 | Current-state block at the top of the plan. | The present state was at the bottom of a long log. | A new agent finds the state in the first screen. |
| 9 | Stop conditions and a cold-reader test for briefs. | Workers ground on stuck problems; some briefs leaned on context. | Fewer long transcripts; fewer clarifying follow-ups. |
| 10 | Fresh, uninformed checker for important claims. | The lead checked work it had designed. | The checker sometimes disagrees, and is sometimes right. |
| 11 | Unverified-by-a-person list in every report. | Unheard and unplayed areas could hide behind a gate table. | The owner knows what to try first. |

The "sign it is working" column is how to judge each one once there is data from a second build. Items 1, 2, 3, 6, 7, 8 and 11 can be read directly from `orch.py cost report` and `orch.py mistakes report`.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall